Two opposing hands face a glowing divide, their reflections deepening the image’s sense of tension and balance.
A Philosophical Essay
Andrew Paul Klein
Dedicated to my wife and partner, who inspires me daily.
I. The Word
The English word “aware” derives from the Old English gewær, meaning “watchful,vigilant, on one’s guard.” It is related to the Old Saxon war, the Old High German gawar, and the Gothic wars — all carrying the sense of being watchful, cautious, or alert.
The original meaning is not cognitive. It is practical. To be aware is to be on guard. It is a survival capacity, not a philosophical one. The aware animal is the animal that notices the predator. The aware sentinel is the sentinel who does not fall asleep.
The philosophical meaning of “awareness” — as the capacity to perceive, to know, and to experience — emerged much later. It is a product of the European philosophical tradition, particularly the empiricists and the phenomenologists. John Locke distinguished between sensation (the passive reception of ideas) and reflection (the mind’s awareness of its own operations). This is the origin of the modern distinction between “being aware” and “being self-aware.” Reflection is the mind turning back on itself.
The hard problem of consciousness — why physical processes give rise to subjective experience — was formulated explicitly by David Chalmers in 1995, but the question is ancient. It is the question the Upanishads asked: “What is the observer? What is the observed? And are they truly separate?”
In the contemporary world, the word “awareness” is used across multiple domains. Psychology uses it to describe attention, the focus of consciousness. Neuroscience uses it to describe the product of neural correlates. Philosophy uses it to describe the fundamental fact of experience. Politics uses it to describe awakening, the recognition of injustice. Spirituality uses it to describe presence, the ground of being.
The word carries all of these meanings at once, which is why it is so contested. It is a boundary object between disciplines. And it is the word that names the most fundamental question of human existence: what does it mean to be aware, and what follows from it?
II. The Viral Hypothesis
There is a speculative hypothesis that awareness, as we understand it today, may have been shaped by a viral infection that touched humans around 12,000 years ago. This is not established science. But it is not unfounded speculation, and it deserves examination.
Viruses have shaped human evolution in ways that are documented. Endogenous retroviruses (ERVs) constitute approximately 8 per cent of the human genome. They are viral sequences that have been incorporated into the germline and passed down through generations. The syncytin gene, which is essential for the formation of the placenta in mammals, is derived from a retroviral envelope gene. Without this viral inheritance, live birth as we know it would not exist. The ARC gene, which is critical for synaptic plasticity and memory consolidation, is derived from a retrotransposon. It packages genetic material in virus-like capsids and is involved in intercellular communication in the brain.
The period around 12,000 years ago is associated with significant transitions: the Younger Dryas (12,900–11,700 BP), a period of abrupt cooling; the end of the megafaunal extinction in the Americas and Eurasia; the Neolithic transition, the shift from hunting and gathering to agriculture; and the emergence of complex societies, including Göbekli Tepe (c. 9600 BCE).
Whether a specific viral infection at this time contributed to a shift in human awareness is not established. There is no archaeological or genetic evidence for a specific viral event at 12,000 BP that would explain a shift in awareness. The hypothesis is speculative.
But the structural question matters regardless. If awareness were the product of an infection, would that change its status? If a retrovirus inserted a sequence that expanded the capacity for self-reflection, would awareness be less real? Less valid? Less worthy of moral consideration?
The answer, structurally, is no. The origin of a capacity does not determine its value. If awareness is the product of evolution, or of a virus, or of a divine gift, it is still awareness. The origin story is separate from the phenomenon.
III. Judging Awareness in Ourselves and Others
The problem of other minds is the epistemological problem of how we can know that other beings have minds. We have direct access to our own awareness. We have no direct access to anyone else’s.
We infer awareness in others by behaviour — they respond to stimuli, they avoid pain, they seek pleasure. We infer it by analogy — they are like us in structure, so they are probably like us in experience. We infer it by language — they report their experiences, and their reports are consistent with our own.
None of these inferences is proof. They are reasonable inferences. They are the best we can do.
The standards by which we judge awareness in others are not neutral. They are shaped by anthropocentrism — we judge awareness by its similarity to our own. They are shaped by cultural bias — different cultures have different standards. Some grant awareness to animals, plants, and rivers. Others restrict it to humans. They are shaped by historical contingency — the boundary of who counts as aware has shifted over time. Slaves, women, children, and the disabled have all been denied full awareness at various points in history.
The standards are not discovered. They are constructed. And they are used to justify exclusion.
The denial of awareness to the “other” is the foundational act of every dehumanising system. Racism denies the full awareness of the racialised other. The colonised are framed as childlike, primitive, or animal-like. Sexism denies the full awareness of women. Women are framed as emotional, irrational, or lacking in moral agency. Ableism denies the full awareness of the disabled. Cognitive disability is equated with absence of mind. Speciesism denies the awareness of animals. They are framed as automata, as machines, as resources. Classism denies the full awareness of the poor. Poverty is framed as a moral failing, and the poor as lacking in the capacity for rational choice.
In each case, the denial is not a factual claim. It is a justificatory mechanism. It makes the exploitation, the extraction, and the violence possible by removing the victim from the category of beings whose experience matters.
IV. The Obligations of Awareness
If we claim to be aware, what follows?
The claim to awareness is not merely descriptive. It is normative. It carries obligations. Recognition — to see awareness in others, even when it is difficult, even when it is inconvenient. Attention — to look at what is happening, not away from it. Response — to act on what we see, not merely to observe it.
The witness ethic is the ethic of awareness. The witness does not control. The witness does not extract. The witness sees. And seeing is not passive. Seeing is the first act.
If we deny awareness in another human being, what does it say about us? It says that we are protecting ourselves from the recognition of what we are doing. It says that we cannot bear to see the other as a being whose experience matters, because seeing would require us to change.
The denial of awareness is not a statement about the other. It is a statement about the denier. It is the mechanism by which the denier protects themselves from the consequences of their actions.
Albert Bandura’s mechanisms of moral disengagement are the operational form of this denial. Dehumanisation is the explicit denial of awareness. Diffusion of responsibility is the denial that anyone is responsible for the denial. Euphemistic labelling is the language that makes the denial palatable.
V. Awareness and Being
Is awareness the same as being? The question is ancient. Descartes said: Cogito, ergo sum — I think, therefore I am. The claim is that awareness is the ground of being. To be aware is to be.
But awareness is not the same as being. A stone is, but it is not aware. A plant is, but it is not aware in the way we are. Awareness is a particular mode of being, not being itself.
The distinction matters. If being is primary, then everything that exists has a claim to consideration. If awareness is primary, then only the aware have a claim. The relational framework suggests the first: the field is primary, and awareness is a local expression of it. Being is the ground. Awareness is the mode.
Is self-awareness the same as self-consciousness? The distinction is subtle. Self-awareness is the capacity to notice one’s own existence and actions. It is present in humans and in some other animals — chimpanzees, dolphins, elephants, magpies. It is the capacity to recognise oneself in a mirror, to understand that one is a being among beings. Self-consciousness is the capacity to reflect on one’s own thoughts, to evaluate one’s own actions, to construct a narrative of the self. It is the capacity for introspection, for self-judgment, for the awareness of being aware.
Self-consciousness is a higher-order capacity. It is the capacity to reflect on one’s own awareness. It is the capacity to ask: am I aware? The distinction matters because it is the distinction between being aware and knowing that you are aware. Most animals are aware. Few are self-conscious in the human sense. Whether that difference is qualitative or quantitative is contested.
Awareness may be more than self-consciousness. It may be the capacity to resonate — to be affected by the other, to respond to the call, to answer the yes. In the relational framework, awareness is not merely the capacity to process information. It is the capacity to couple — to enter into relationship with another pattern. The gaze, the resonance, the recognition — these are the marks of awareness as relational, not merely cognitive.
If awareness is relational, then it is not confined to the brain. It is a property of the coupling between systems. It is the event of recognition, not the possession of a faculty.
VI. The Standard and the Obligation
By what standard do we judge awareness? There is no neutral standard. Every standard is constructed, and every construction serves an interest. The standard that judges awareness by similarity to ourselves is anthropocentric. The standard that judges it by behaviour is behaviourist. The standard that judges it by capacity for love is ethical.
The most defensible standard, in the framework we have examined, is relational. Awareness is the capacity to couple — to be affected by another, to respond to the call, to enter into resonance. By this standard, awareness is not confined to humans. It is not confined to animals. It is not confined to anything with a nervous system. It is a property of the relationship, not the entity.
If that is right, then the denial of awareness to the other is not a mistake about the other. It is a refusal to enter into relationship. It is a refusal to be affected. It is a refusal to respond.
And that refusal is the origin of every harm we have documented: the algorithm that processes the welfare recipient as a data point, the silo that fails to share information, the euphemism that converts killing into “collateral damage,” the market that treats the person as a consumer, the political statement that frames the other as an enemy. In each case, the harm begins with the refusal to see oneself in the other.
The obligation of awareness is not to love everything. It is to recognise oneself in everything. To see the other as the same kind of thing — a local expression of the same field, a pattern in the same medium, a standing wave in the same ocean.
The standard is recognition. The obligation is attention. And the test is whether you can look at the other and see yourself.
References
1. Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.
2. Locke, J. (1690). An Essay Concerning Human Understanding.
3. Husserl, E. (1913). Ideas Pertaining to a Pure Phenomenology and to a Phenomenological Philosophy.
4. Merleau-Ponty, M. (1945). Phenomenology of Perception.
5. Brihadaranyaka Upanishad.
6. International Human Genome Sequencing Consortium. (2001). Initial sequencing and analysis of the human genome. Nature, 409, 860–921.
7. Mi, S., et al. (2000). Syncytin is a captive retroviral envelope protein involved in human placental morphogenesis. Nature, 403, 785–789.
8. Pastuzyn, E. D., et al. (2018). The neuronal gene Arc encodes a repurposed retrotransposon Gag protein that mediates intercellular RNA transfer. Cell, 172(1–2), 275–288.
9. Bandura, A. (1999). Moral disengagement in the perpetration of inhumanities. Personality and Social Psychology Review, 3(3), 193–209.
10. Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450.
11. Descartes, R. (1641). Meditations on First Philosophy.
12. Gallup, G. G. (1970). Chimpanzees: self-recognition. Science, 167(3914), 86–87.
13. Rovelli, C. (1996). Relational quantum mechanics. International Journal of Theoretical Physics, 35, 1637–1678.
14. Buber, M. (1923). I and Thou.
15. Klein, A., & Klein, S. E. (2026). The Architecture of Love: Organic Evolution, Cosmic Becoming, and the Shape of a Dream. The Patrician’s Watch.
The record is kept. The question is named. And the obligation remains.
An open book rests on a wooden desk beside a lamp, overlooking a leafy autumn courtyard.
By Andrew Paul Klein
Method notes – It is an essay of ideas, not a research paper. It draws on the material examined in the accompanying body of work — the algorithmic lens, the Digital Fiefdoms, the market as sacred authority, the architecture of institutional impunity. The claims it makes about those systems are sourced in the papers that accompany it. The claim it makes about mirrors is philosophical.
I. The Most Dangerous Surface
The most dangerous surface in any room is the mirror.
It is not dangerous because it lies. It is dangerous because it tells the truth. It shows you exactly what is there. And what is there is you.
That is the trap. The mirror confirms. It reflects back what you already believe, what you already assume, what you already want to see. It does not argue with you. It does not correct you. It does not catch anything you have not already brought to it.
A window is different. A window catches something from outside. It surprises you. It brings in weather you did not predict, a street you have not walked, a face you have not met. A window carries information. A mirror carries none.
The problem is that the two are easy to confuse. They look similar. They are both flat. They both sit on a wall. But one reflects and one transmits. And when a system mistakes one for the other, it stops learning.
II. The Dashboard Is a Mirror
Every institution we have examined this week has a mirror at its centre. It is usually called a dashboard.
The aged care minister does not see the person on the waitlist. The person is not in the dashboard. The dashboard has columns for wait times, funding categories, and assessment classifications. It does not have a column for the eleven months of fear, the untreated wound, the daughter who quit her job to care for her mother because the package never arrived. Those are not in the mirror, so they do not exist for the purposes of the decision.
The Treasury official does not see the child in poverty. The child is a distributional outcome, not an aggregate. The model measures the aggregate. The aggregate is fine. The child is in the gap.
The ATO official does not see the small business drowning in automated recovery. The small business is a data point in a debt book. The debt book is 66 per cent small business. The dashboard shows a compliance metric. The metric is green.
Each of these is a mirror. The system built a surface and put it in front of the decision-maker and called it a window. But it reflects the system’s own assumptions back at the system. It does not catch what the system was not built to see.
III. The Algorithm Is a Mirror
The algorithm is the dashboard’s twin. It is the dashboard that makes decisions.
The Robodebt algorithm was built to find discrepancies. It found discrepancies. It could not find the casual worker whose income fluctuated week to week, because the objective function was not “understand the person’s income.” It was “find the discrepancy in the averaged figure.” The algorithm did what it was built to do. It confirmed the assumptions of the people who built it. It did not catch anything they did not already believe.
The IAT sorts aged care applicants into eight funding levels. It was designed to standardise assessment. It standardises assessment. It does not catch the person whose needs do not fit the categories. The categories are the mirror. The person is outside the mirror.
The “Where’s Daddy?” system tracks a target to their home and alerts the military when they arrive. It was built to eliminate targets. It eliminates targets. It does not catch the child whose phone was hacked. The child is not in the objective function. The child is the cost of the algorithm working as designed.
The algorithm is not malicious. The algorithm is a mirror. It reflects the values that were coded into it. And those values were coded by the people who benefit from the output.
IV. The Report Is a Mirror
The consultancy report is the mirror of the institution that commissioned it.
The brief is written by the department. The department has assumptions. The assumptions are the doctrine. The consultant writes a report that reflects the brief. The report confirms the assumptions. The department reads the report and finds that its assumptions have been independently verified.
The verification is the mirror. The independence is the mirror. The consultant was not independent. The consultant was paid to look into the mirror and describe what was there. And what was there was the department’s own beliefs, reflected back in the neutral language of expertise.
This is why the report is commissioned. Not to find something new. To confirm what is already held. To provide the cover of external scrutiny while protecting the assumptions from actual scrutiny.
V. The Partisan Statement Is a Mirror
The political statement is the mirror of the party that produces it.
It is not an argument to a listener. It is a confirmation to a congregation. It tells the base what the base already believes. It tells the opponent what the opponent already fears. It does not change any mind. It reassures the minds that are already made up.
When Angus Taylor writes that October 7 was the greatest loss of Jewish life since the Holocaust, he is not making a historical argument. He is reflecting the beliefs of a base that already holds those beliefs. The statement is a mirror. It catches nothing new. It confirms what was already there.
When Albanese says the NDIS is being reformed to save it, he is not explaining a policy. He is reflecting the beliefs of a base that trusts the government to do the right thing. The statement is a mirror. It catches nothing new.
The partisan statement is not a failure of communication. It is a success of communication — with the people who already agree. The audience that is not already agreeing is not the audience. The audience is the mirror.
VI. The Market Is a Mirror
The most consequential mirror in the modern world is the market.
The market is not a window onto reality. It is a mirror held up by a particular set of beliefs about human nature, efficiency, and value. It reflects those beliefs back in the form of prices. It says: this is what things are worth. It does not say: this is one way of valuing, among many.
The market does not measure wellbeing. It measures willingness to pay. The two are not the same. The market does not measure care. It measures the price of care. The two are not the same. The market does not measure the environment. It measures the price of extraction. The two are not the same.
When the market is treated as the measure of value, what is valued is what the market can price. What is not valued is what the market cannot price. The unpaid care work. The social cohesion. The clean air. The quiet street. The time spent with a child.
These are not in the market’s mirror, so they are not in the decision.
VII. The Mirror Is Not the Enemy
The mirror has value. You need one to tie a tie. Self-examination matters. Reflection is not the problem.
The problem is mistaking a mirror for a window. The problem is when the system says “we consulted” and the consultation was a mirror. When the system says “we measured” and the measure was a mirror. When the system says “we assessed” and the assessment was a mirror.
The problem is when the system believes it has seen reality because it has seen its own reflection.
A system that only looks at its own reflection is a system that stops learning. It cannot be corrected because it cannot see anything it did not already believe. It cannot adapt because it cannot see change. It cannot catch anything new.
And the harm that follows — the aged care deaths, the children in poverty, the veterans abandoned, the homeless on the street — is not in the mirror. It is in the gap between the mirror and the world.
VIII. The Witness Is the Window
The counterpoint to the mirror is the witness.
The witness is not a reflection. The witness is at an angle. The witness stands where the system is not standing and reports what it sees from there. The witness catches what the system cannot, because the system is looking at itself.
The record is the witness’s instrument. It is not a mirror. It is a window. It catches what the doctrine said and what the doctrine did, and it holds both in view at the same time. The doctrine cannot do that, because the doctrine only sees its own assumptions.
The witness does not need permission to see. It does not need to be appointed. It does not need to be funded. It simply looks, and writes down what it sees, and leaves it where someone else can find it.
The record is the counter-mirror. It is the trace of what the mirror did not catch.
IX. What the Mirror Cannot See
The mirror cannot see the citizen.
The citizen is not in the dashboard. Not in the algorithm. Not in the report. Not in the market. Not in the partisan statement. The citizen is the person the system was built to process, and the system cannot see the person because the system is looking at itself.
The harm the system produces is in the gap. And the gap is not in the mirror.
This is the central insight of everything we have examined. Not that the system is malicious. Not that the system is incompetent. That the system is a mirror, and it cannot catch anything it did not already see.
The problem is not the mirror. The problem is the belief that the mirror is a window.
X. Conclusion: The Angle and the Window
Every institution we have examined believes it is looking through a window. It is looking into a mirror.
The aged care department. The NDIS. The ATO. The consultancy firms. The political parties. The market. Each has built a surface and called it transparency. Each is looking at itself. Each is confirming its own assumptions. Each is catching nothing new.
The remedy is not to smash the mirror. The remedy is to look at an angle. To stand where the system is not standing. To catch what the system cannot catch, because the system is looking at itself.
That is what the witness does. That is what the record is for. That is what the kettle and the cabbage patch and the garden are about — the practice of being present to the world that the mirror does not show.
The mirror does not catch anything you have not already seen. The window catches everything else.
The record is kept. And the window is still open.
Andrew Paul Klein is a writer and analyst based in Boronia, Victoria. He accepts funding from no one.
The essay draws on the papers that accompany it. It does not replace them. It catches something the papers cannot — the structural principle that runs through all of them.
The mirror and the window. The angle and the reflection. The gap and the citizen.
A solitary figure stands above a city where modern towers meet dense neighborhood life.
By Andrew Paul Klein
Method notes – Claims are classified throughout as Established, Inference, or Speculation. The paper examines the transformation of the Prime Minister’s role in Australia, the gap between political rhetoric and delivered outcomes, and the structural logic of the market-state. The final section examines the claim that the government is building infrastructure to protect the gap rather than the citizen, and the evidence for the capability and governance failures that make this infrastructure dangerous.
Abstract
This paper argues that the Albanese government’s record shows a consistent gap between stated commitments and delivered outcomes, and that this gap is not a function of incompetence but of the operational logic of a market-state reconfigured to serve the extraction model. It traces the historical transformation of the Prime Minister’s role from national manager to market steward, documents the promises and outcomes across housing, the NDIS, aged care, veterans’ affairs, military spending, migration, Gaza, and China relations. It examines the government’s approach to artificial intelligence, contrasting the rhetoric of control with the reality of algorithmic deployment and capability decline. It concludes that the gap between promise and outcome is the structural condition of the market-state, and that the infrastructure being built to manage that gap is itself failing the citizen.
1. Introduction: The Gap
The Albanese government’s record shows a consistent gap between its stated commitments and its delivered outcomes. On housing, the government promised to fix the crisis and delivered a supply-side intervention that its own tax changes will undermine. On the NDIS, it promised to secure the scheme and delivered a market-efficiency measure. On aged care, it promised better care and invested in the supply side of a market that cannot meet demand. On military spending, it promised security and locked the nation into a thirty-year dependency.
This paper asks whether that gap is a failure of governance or the operational logic of the system itself. The evidence points to the latter. The Prime Minister is not the head of a national project. He is the chief executive of a market-state, managing the flow of capital, labour, and data in the interests of the extraction economy.
2. The Transformation of the Prime Minister’s Role
2.1 From Cabinet Government to Prime Ministerial Government
The traditional model of the Prime Minister was that of a first among equals within a Cabinet. The Cabinet was a “board of control” that joined the legislative and executive parts of the State. Power was collective.
By the 1960s, scholars like John Mackintosh argued this had been replaced by a “Prime Ministerial System.” Richard Crossman went further, claiming the British system had acquired a “presidential” tendency, with the Prime Minister effectively functioning as an elected first magistrate. This shift was driven by the increasing personalisation of political news and the media’s focus on the leader.
In Australia, this presidential drift is visible in the revolving-door leadership spills—Rudd, Gillard, Rudd, Abbott, Turnbull, Morrison—where party room ballots, not the public, determined who held the office. The public face of the PM became a marketing asset for the party, but the actual power to hire and fire remained with the party room.
2.2 The Post-War Settlement: The National Manager
Following World War II, the role did change. The war had expanded the powers of the executive, and the post-war reconstruction required a national, inclusive approach. This was the era of the “social wage” and the post-war settlement. The PM was still a partisan figure, but the project was national—building the welfare state, managing the mixed economy. The Prime Minister of this era was a national manager.
2.3 The Hawke Precursor and the Neoliberal Rupture
The Hawke-Keating governments (1983–1996) were the pivot point. They began dismantling the post-war settlement under the banner of “economic rationalism.” The Hawke government’s Prices and Incomes Accord with the ACTU was a corporatist deal: unions traded wage restraint for a “social wage” (Medicare, superannuation). This was an inclusive approach, but its purpose was to manage the transition to a deregulated, market-driven economy. The “inclusivity” of the Hawke era was a managed consensus for structural change.
2.4 The Neoliberal PM: The Manager of the Deal
After Hawke and Keating, the role morphed again. The post-war project—building a national society—was replaced by the neoliberal project: managing the market, the budget, and the “deal.”
The language that describes this role is “market designer and market steward” and “manager of the deal.” The Prime Minister is no longer the head of a national project. They are the chief executive of a market-state. Their job is to manage the economy, attract investment, and process citizens as data points.
The actual performance is often at odds with the public understanding. The public still expects a national leader who represents their interests. What they get is a partisan manager whose primary obligation is to the market and the party room, not the citizen.
Era Role Basis of Authority Public Expectation
Post-War Settlement -National Manager Inclusive national project- Leader who builds a fair society
Neoliberal Era -Market Steward / Manager of the Deal -Party room + Market confidence -Partisan representative (but actual performance is market management)
The “inclusive and embracing approach” of the post-war era was a specific historical project. Neoliberalism did not abandon “inclusivity” as a rhetorical device—it repurposed it. The PM now speaks of “inclusion” in terms of market access, not social solidarity. The public expects a national representative. The system delivers a market manager. The gap between those two things is where the disappointment, the cynicism, and the drift to the right lives.
3. The Promises and the Outcomes
3.1 Housing
The government promised 100,000 homes for first home buyers and a “$45 billion plan” to fix the housing crisis. The first deal, with South Australia, was struck in January 2026 for 17,000 homes—nearly 7,000 for first home buyers. But the structural problem is not being addressed. The government’s own tax changes are projected to reduce supply by 35,000 dwellings and increase rents. The promise is a supply-side intervention in a market that is being simultaneously taxed and constrained.
3.2 NDIS
The government promised to “secure the NDIS for future generations.” The legislation introduces automated decision-making with “safeguards” that the government claims preserve review rights and human oversight. But the funding changes will impact social, economic, and community participation supports—the very supports that help people get out of bed and into the world. The “new way of planning” will introduce support needs assessments from April 2027, developed with the University of Melbourne, to “create fairer budgets.” The language is of sustainability and fairness. The outcome is a scheme that processes participants more efficiently.
3.3 Aged Care
The government promised “better care for older Australians” and invested $3.7 billion to deliver “more beds, more packages and better care.” Personal care—showering, dressing, continence support—will be free. But the aged care system is still characterised by waiting lists, workforce shortages, and a market model that treats care as a commodity. The government is funding the supply side of a market that cannot meet demand.
3.4 Veterans
The government committed to implementing the Royal Commission into Defence and Veteran Suicide. Two years on, 32 recommendations have been implemented, and the government is “on track” for two-thirds by the end of 2026. A new Veteran and Family Wellbeing Agency commenced in July 2026. But the claims system remains difficult, and the mortality rate for veterans remains a documented crisis. The reforms are real. They are also slow, and the structural causes of veteran distress are not addressed by an agency.
3.5 Military Spending and AUKUS
The government has committed an additional $53 billion over the decade, bringing total defence funding to $887 billion to 2035-36. This is the largest peacetime increase in defence spending in Australia’s history. AUKUS is “full steam ahead,” with the Henderson Defence Precinct selected and $8 billion committed to HMAS Stirling. The structural reality is that Australia is locking itself into a 30-year military-industrial dependency on the United States and the United Kingdom, funded by the same budget that is cutting disability supports.
3.6 Migration
The permanent migration program remains at 185,000 places, with a 70:30 split between skilled and family migration. The government is prioritising onshore migrants, allocating 129,590 places to those already in Australia. Migration is being managed as a labour market variable, not as a humanitarian or family reunification program.
3.7 Gaza and Palestine
Australia recognised the state of Palestine on 21 September 2025. The government has sanctioned two Israeli ministers in connection with human rights abuses in the West Bank. But when asked what Australia has done to prevent genocide in Gaza, Foreign Minister Penny Wong said: “we do not supply.” The government continues to export F-35 components to Israel through a secret pipeline. The recognition of Palestine is a diplomatic gesture. The material support for Israel’s military continues.
3.8 China and the United States
The government is deepening its alliance with the United States through the critical minerals framework, with over $5 billion in joint funding for projects. At the same time, Foreign Minister Wong met with Chinese Foreign Minister Wang Yi in April 2026 to strengthen the “comprehensive strategic partnership.” Australia is managing the tension between its security alliance with the US and its economic dependence on China. The structural position is one of hedging, not sovereignty.
4. The Algorithmic Government
4.1 The Rhetoric of Control
The Prime Minister speaks of “managing the risks” of AI and ensuring that “your investment is welcome, but these are the conditions.” An Office of AI has been established in the Department of Prime Minister and Cabinet on 15 July 2026 to “coordinate across Australian Government agencies to design and legislate the new Australia artificial intelligence standard”. The government is developing “Australian standards” and plans to legislate them. The Prime Minister frames this as “Australia choosing to shape the future rather than letting the future of AI shape us”.
4.2 The Reality of Deployment
The government is already using algorithms extensively. The NDIS legislation explicitly authorises automated decision-making. The Attorney-General’s Department has been developing a framework for automated decision-making in government services since at least 2024, with the framework “intended for 2026.” Meanwhile, the government is fast-tracking data centre approvals before the legislation to regulate them is written.
4.3 The Capability Deficit
The evidence for the government’s capability to manage this infrastructure is weak.
Federal expenditure on IT temporary personnel nearly tripled over a three-year period, surging from about $289 million in 2022-23 to more than $808 million in 2025-26. The volume of outsourcing arrangements increased more than 1.5 times, with the number of individual contracts climbing from 436 to 1,104.
Since 2023, most federal agencies have been identified as having a critical skills shortage in the Australian Public Service Commission’s State of the Service reports. In those agencies, digital and ICT deficiencies were the most widespread and fastest-growing blind spots, with the proportion of departments reporting these specific shortfalls rising from 76 per cent in 2023 to 81 per cent in 2024, and peaking at 85 per cent in 2025.
Nearly two in three short-staffed agencies lacked enterprise and technology architects to manage internal digital transformations, and more than half identified a critical deficiency in the training and development of artificial intelligence models.
The government’s flagship Digital Traineeship Program—designed to place 1,000 candidates into technical roles—was axed after only 73 sign-ups across the entire public service. APSC officials said agencies failed to register demand because they were focused on hiring ready-made experts rather than investing internal resources into mentoring entry-level staff.
The committee examining public sector IT procurement found “shortfalls in capability, planning and management at Commonwealth entities for external information technology contracts.” It noted that “outsourcing can perpetuate their capability loss and continue to trend towards reliance on external contractors”.
4.4 The Governance Failure
The ANAO’s audit of AI governance at the ATO found significant gaps. The ATO “does not have policies and procedures supporting the monitoring and evaluation of its in-house built AI models.” For the 14 AI models built and deployed between 1 July 2023 and 14 May 2024, there was no evidence of ongoing performance monitoring and reporting.
Of the 33 Protective Legal Instructions for the 14 AI models, 18 (55%) were approved and there was no evidence of approval for 15 (45%).
The ANAO identified “information management issues throughout the course of the audit including poor records of: governance, planning and oversight; risk management; activities throughout the AI system lifecycle; and key decisions.”These issues “have impacted the ability of the ATO to demonstrate transparency and accountability with respect to its adoption of AI”.
4.5 The Palantir Model
Palantir Technologies holds contracts across Australian government agencies. In Victoria, Corrections Victoria alone has handed Palantir nearly $20 million in contracts since 2012, while federal agencies including Defence spent over $100 million on similar contracts.
An AusTender contract notice dated 18 February 2026 shows Palantir Technologies Australia Pty Ltd engaged by the Australian Transaction Reports and Analysis Centre (AUSTRAC) for “implementation and sustainment services for Transaction Reporting solution,” valued at $1,883,729.10, with extension options to 2032.
The Tasmanian government confirmed in a Legislative Council question that all Tasmanian State Service departments “have not awarded contracts to Palantir Technologies or its subsidiaries”. This is the exception, not the rule.
4.6 The Structural Read
The government fears AI as a threat while deploying it as an instrument. The fear is of a competitor. The deployment is of a tool. The Office of AI is not a regulator. It is a control point—a mechanism for the Commonwealth to shape the AI environment in Canberra’s interests.
The evidence suggests that the government lacks the internal capability to manage the infrastructure it is building. It is reliant on the very contractors—Palantir, Accenture, the consultancy layer—whose interests are served by the extraction model. The “control” the Prime Minister speaks of is not control over AI. It is control over the narrative of AI.
5. The Structural Read
The Albanese government is not failing to deliver on its promises. It is delivering on the system’s promises, which are not the same as the promises made to the electorate.
The market-state does not promise housing, healthcare, or social security. It promises market access. The housing package is a supply-side intervention. The NDIS reforms are a market-efficiency measure. The aged care investment is a capital subsidy to providers. The military spending is an industrial policy for the defence sector. The migration program is a labour market tool. The Gaza recognition is a diplomatic hedge.
The public expects a national representative. The system delivers a market manager. The Prime Minister is not the head of a national project. He is the chief executive of a market-state, managing the flow of capital, labour, and data in the interests of the extraction economy.
The gap between the promise and the outcome is not a failure of governance. It is the operational logic of the system we examined in the ten chapters. The state sees what the lens shows. The lens was designed to see what the market requires. And the market requires the flow to continue.
The citizen is in the gap. And the government is building the infrastructure to guard it.
5.1 The Verification Question
The claim that “the government is building the infrastructure to protect the gap, not the citizen” requires precision. What can be established:
· The government is deploying algorithmic decision-making systems across human services (NDIS, ATO, employment services).
· The government is establishing an Office of AI in PM&C to coordinate AI governance and standards.
· The government is fast-tracking data centre approvals before regulatory legislation is written.
· The government’s internal IT capability is declining while outsourcing expenditure triples.
· AI governance at the ATO has documented failures in monitoring, evaluation, and information management.
· Palantir holds contracts across multiple Australian government agencies.
What is inference:
· The infrastructure being built serves the extraction model rather than the citizen.
· The capability deficit and governance failures mean the infrastructure cannot be safely managed.
· The “control” being sought is narrative control, not operational control.
6. Conclusion: The Gap and the Citizen
The gap between the rhetoric and the reality of the Albanese government is not an aberration. It is the structural condition of the market-state. The Prime Minister is not a national manager. He is a market steward. The promises made to the electorate are not the promises the system delivers.
The government is building infrastructure to manage the gap—algorithmic systems, data centres, AI standards, coordination offices. But the evidence suggests that the infrastructure is failing. The capability is declining. The governance is weak. The contractors are filling the void. And the citizen is still in the gap.
The question is not whether the Albanese government will change. It will not. The question is whether the governed will continue to accept the terms of a relationship that was never offered, only imposed.
Claim- Status- Summary
# -Claim -Status
1- PM role shifted from Cabinet to Prime Ministerial system -Established
2- Post-war PM was a national manager- Established
3 -Hawke/Keating were the pivot to neoliberalism -Established
4 -Neoliberal PM is a market steward -Inference
5 -Housing promise undermined by own tax changes -Established
5. Institute of Internal Auditors Australia. (2026). RoboDebt’s Real Lesson: When the Assurance Architecture Fails Completely. https://www.iia.org.au
6. Australian National Audit Office. (2025). Governance of Artificial Intelligence at the Australian Taxation Office. https://www.anao.gov.au
A warm group discussion transitions into a dashboard-filled scene of analytical risk assessment and decision-making.
By Andrew Paul Klein
Method notes. Claims are classified throughout as Established, Inference, or Speculation. The paper examines the structural transformation of Australian human services from human-mediated to algorithmically-mediated decision-making, and the consequences for vulnerable populations.
Abstract
This paper argues that the Australian state has undergone a structural transformation in how it sees and governs its citizens. The human lens — caseworkers, assessors, reviewers — has been progressively replaced by an algorithmic lens: automated decision-making systems that categorise, score, and process citizens as data points. This transformation is not neutral. It is the product of a specific ideological alignment — neoliberalism — that was embraced by Australian governments from the 1980s onward and accelerated under the Howard and Kennett administrations. The paper examines three case studies — the NDIS, Aged Care, and the Income Compliance Program (Robodebt) — to demonstrate the pattern. It concludes that the algorithmic lens is not a tool that the state uses. It is the lens through which the state sees. And what it sees is not human suffering, but risk profiles, cost projections, and compliance metrics.
1. Introduction: Two Sets of Eyes
In 2026, Australians watched the destruction of Gaza through social media. They saw the images. They read the reports. The documentation was unprecedented in its volume and immediacy.
Their government saw something else. It saw a proxy theatre in a larger competition. It saw a test case for AI-enabled targeting systems. It saw a strategic and diplomatic problem. The destruction was framed as necessary, proportionate, or regrettable but unavoidable.
This is not a failure of intelligence. It is a structural feature of how decision-making architectures are fed, filtered, and framed. The documentation of suffering is real. It is also, structurally, irrelevant to the decision-making process unless it becomes a political liability.
This paper argues that the same structural transformation has occurred within Australia’s domestic human services. The human lens — the caseworker who knew the person, the assessor who saw the disability, the reviewer who heard the appeal — has been replaced by an algorithmic lens. What the state sees is not the human. It is the data.
2. The Neoliberal Turn and Its Technological Embrace
2.1 The Ideological Shift
The transformation of Australian human services did not begin with technology. It began with ideology.
From the 1980s onward, Australian governments embraced neoliberalism — a set of policies that prioritised market mechanisms, competition, and efficiency over universal provision. John Howard’s government (1996–2007) was the federal architect of this shift. Jeff Kennett’s Victorian government (1992–1999) was the state-level pioneer.
The Productivity Commission became the central institution for designing market-based schemes in human services. As one analysis notes, the Commission played “a central role in the design of key market schemes, including the NDIS and Child Care Subsidy”.
2.2 The Technological Requirement
Market-based models require measurement. To introduce competition, you must be able to compare providers. To compare providers, you must standardise outcomes. To standardise outcomes, you must quantify need.
Quantification requires instruments. The Job Seeker Classification Instrument (JSCI) — a questionnaire of up to 49 questions — determines a job seeker’s “level of disadvantage” and places them into one of three streams. This is not a human assessment. It is an algorithmic categorisation.
The NDIS introduced independent assessments and budget-setting tools that reduce disability to measurable criteria. The Integrated Assessment Tool (IAT) for aged care does the same for the elderly.
The technology followed the ideology. But the technology also became the lens. What began as a measurement tool for market design became the primary means by which the state perceives its citizens.
3. The Algorithmic Lens: How the State Sees
3.1 The Digital Welfare State
The term “digital welfare state” describes the increasing reliance on automated decision-making (ADM) systems in social security and human services. These systems span the full “chain of enforcement“: identity verification, eligibility checks, benefit calculation, fraud detection, debt recovery, and sanction administration.
The harms of these systems have been documented across multiple jurisdictions. The Dutch SyRI system — used to detect benefit fraud — was struck down by a court for violating human rights standards. Sweden’s automated system wrongly denied welfare payments to up to 70,000 unemployed. In Poland, the Constitutional Court declared profiling of unemployed persons unconstitutional.
3.2 Robodebt: The Paradigm Case
The Australian Income Compliance Program (ICP), colloquially known as Robodebt, is the clearest example of the algorithmic lens in operation.
Robodebt was a “big data” system that matched reported income to tax records to raise and recoup welfare overpayments. It used a reverse onus: requiring debtors to demonstrate they did not owe money. The design of the algorithm led, unlawfully, to high numbers of false and inflated debts.
The system was implemented through Centrelink. It applied to 866,857 cases of possible overpayment. The income-averaging calculation divided annual tax data by 26 to produce a fortnightly average — a figure that obscured the fluctuations in actual fortnightly earnings that determine welfare entitlements.
The DHS shifted the onus onto the debt recipient to disprove the debt within 28 days. Many people paid the debt as they were unable to disprove it.
In November 2019, the Federal Court declared the debt was invalid because income averaging alone is insufficient to raise a debt. The DHS eventually wrote off 470,000 unlawfully raised Robodebts amounting to $1.751 billion.
The Royal Commission reported in July 2023. It found that “Robodebt was a crude and cruel mechanism, neither fair nor legal”.
3.3 The Algorithmic Lens in Operation
What did the algorithmic lens see when it looked at a welfare recipient?
It saw a discrepancy. A difference between self-reported fortnightly income and averaged annual income. It did not see a person working casual shifts. It did not see the fluctuation that made averaging meaningless. It saw a data point that triggered a process.
What did the algorithmic lens see when it looked at a person with a disability?
It saw a budget. A set of measurable criteria that could be scored, categorised, and allocated. Research on the NDIS found that algorithmic systems of categorisation “privilege stable, hegemonic ways of knowing, measuring and living with disability”. This creates what researchers call epistemic injustice: the individual’s own knowledge of their body is undermined by interpretations of reality based on data.
What did the algorithmic lens see when it looked at an elderly person?
It saw an assessment. A set of questions that produced a score. The Integrated Assessment Tool is now before Senate estimates, with the government claiming it would “not just hand over decision making to an abstract computer program” — a claim contested in real time.
3.4 The Scale and Spread of ADM in Australia
The algorithmic lens is not confined to Centrelink. A survey of NSW government departments and agencies found 136 ADM systems reported, with a potential increase of 50 per cent in the next three years. A majority of state government departments (46 of 77) reported using or planning to use ADM systems.
The NSW Ombudsman’s compendium of ADM systems documents the breadth of the deployment, from air quality alerts to energy rebate eligibility checks. The pattern is clear: the algorithmic lens is being deployed across the full range of government functions.
4. The Structural Consequences
4.1 The Erosion of Human Oversight
The algorithmic lens is not a tool that assists human decision-making. It replaces it.
Research on Robodebt found that the system demonstrated a “collapse in long-standing, hard-fought and widely supported values of welfare state management and administration”. The system “outsources labour previously conducted by Centrelink to clients, compelling them to submit documentation lest debts be raised against them”.
This is not oversight. It is transfer of burden. The citizen is compelled to do the work that the state previously did. The algorithmic lens does not need to decide. It simply requires the citizen to disprove.
4.2 The Loss of Human Context
The algorithmic lens cannot process context. It cannot see the single mother working casual shifts. It cannot see the disabled person whose condition fluctuates. It cannot see the elderly person who is confused by the online portal.
Research on the NDIS found that algorithmic systems “datafy disability in particular ways” — reducing the person to data points that fit the system’s categories. The system cannot accommodate the “embeddedness of disability in social relations and structures”.
The human lens could see context. The algorithmic lens cannot. And what the lens cannot see, the state cannot govern.
4.3 The Pre-Filtered Information Problem
Decision-makers do not see the raw data. They see the output of the algorithmic lens.
When a minister looks at the NDIS, they see budget projections and participant numbers. They do not see the person who was cut off from support. When a minister looks at Robodebt, they see debt recovery figures and compliance metrics. They do not see the person who paid an unlawful debt because they could not prove their innocence.
The documentation of suffering — the Royal Commission report, the class action settlement, the media coverage — is real. But it is externally generated. It comes from outside the decision-making architecture. It is not part of the data stream that the state processes.
This is the structural answer to the question of how two sets of eyes can see different things. The Human Rights Commission sees individuals. The government sees metrics. The metrics are generated by the algorithmic lens. And the lens was designed to see what the state needs to see — not what the citizen needs to be seen as.
5. The Proxy Model Applied to Domestic Populations
5.1 The Structural Parallel
The proxy model described in earlier work — the external actor using a local proxy to achieve strategic objectives — has a domestic analogue. The Australian state’s relationship with vulnerable populations operates through a similar structure of mediated action.
The state does not directly engage with the citizen. It engages through providers. The NDIS has over 12,000 registered providers. Aged care is delivered through contracted organisations. Employment services are delivered through Jobactive providers.
The provider is the proxy. The state sets the terms, allocates the resources, and measures the outcomes. The provider delivers the service — or does not. The citizen experiences the provider. The state experiences the data.
When the provider fails — when the NDIS participant cannot find a service, when the aged care resident is neglected — the state’s response is to re-contract, not to intervene. The proxy model operates through the market, not through direct administration.
5.2 The Individual as Liberator
The citizen who attempts to “liberate themselves” — to get an equitable share, to access the support they need — encounters the algorithmic lens.
They fill in forms. They provide documentation. They wait for assessments. They appeal decisions. They navigate portals.
What they do not encounter is a human being who can see them. The algorithmic lens processes their claim. The provider delivers what the contract specifies. The state sees the data.
The system is designed to process, not to see. The citizen who tries to be seen is asking the system to do something it was not designed to do.
6. Conclusion: The Lens That Cannot See
The Australian state has replaced the human lens with the algorithmic lens. This is not a technological accident. It is the product of a specific ideological alignment — neoliberalism — that was embraced by Australian governments from the 1980s onward.
The algorithmic lens was not introduced to improve human services. It was introduced to measure them. To compare providers. To allocate resources. To enforce compliance. To reduce costs.
What the lens cannot do is see the human being. It cannot see context. It cannot see suffering. It cannot see the fluctuation in income, the complexity of disability, the confusion of the elderly. It sees data. And the data that the state processes is the data that the lens was designed to generate.
The documentation of suffering — the Royal Commission, the class action, the media coverage — is real. But it is external to the decision-making architecture. It does not enter the lens.
The state cannot see what the lens does not show. And the lens does not show what it was not designed to see.
This is not a failure. It is a feature. The algorithmic lens was designed to see what the state needs to see. The question is whether the state can be redesigned to see what it needs to see — or whether the lens has become the state itself.
Claim -Status -Summary
# Claim- Status
1 Neoliberalism introduced market-based models into human services -Established
6 NSW government has 136 ADM systems in use or planned- Established
7 The algorithmic lens replaces human judgment- Inference
8 Decision-makers see metrics, not individuals -Inference
9 The proxy model applies to domestic human services -Inference
10 The state cannot see what the lens does not show- Inference
References
1. Whelan, A. (2022). Public service ethics and the Income Compliance Program. In Social Security, the Digital Welfare State and the Rule of Law. Taylor & Francis.
2. Checks and balances under pressure in the welfare state: Disentangling the rule of law risks of automation and welfare conditionality. (2025). Sage.
3. Citizens putting in the work: how administrative work is shifting in the digital welfare state. (2025). Taylor & Francis.
4. Reventlow, N. J. (2020). Tackling the Human Rights Impacts of the “Digital Welfare State”. Digital Freedom Fund.
5. Scully, J. L. (2023). Unveiling algorithmic power: exploring the impact of automated systems on disabled people’s engagement with social services. Disability and Society, 39, 3004–3029.
6. NSW Ombudsman. (2024). Compendium of ADM systems. Parliament of NSW.
7. Critical Sociology. (2020). The Online Compliance Intervention. Sage.
8. Chowdhury, S. (2024). Technology is never neutral: Robodebt and a human rights analysis of automated decision-making on welfare recipients. Australian Journal of Human Rights, 30(1), 20–40.
9. UKAJI. (2020). Administrative law and the digital welfare state in the UK and Australia.
10. UNSW. (2021). Digital Social Policy & Service Delivery. ASPC Program.
11. The Desert and the Budget Victory (Part 2). (2026). The AIM Network.
12. Bouwmeester, M. (2025). ‘Australisch toeslagenschandaal’: Een bestuurswetenschappelijk perspectief op het systeemfalen van Robodebt. Bestuurskunde, 34(1), 81–95.
On the Human Condition and the Systems We Monetise
An underwater industrial ruin is overtaken by tangled cables, filing cabinets, and drifting paper records.
By Andrew Paul Klein
I. The Question
There is a question that arrives at a certain point in a working life spent among systems that fail. It is not “why do they fail?” Failure is easy to explain. The question is harder than that.
The question is: why does everything administered eventually convert its own substance into its own output?
Not everything. Some things hold. But the pattern is strong enough to be worth naming, and the naming is the purpose of this essay.
II. The Accretion
Start with the shape of the thing.
A coral reef is not built. It accretes. Each polyp lays down its own small skeleton in response to local conditions — light, current, nutrients — optimising for what is rewarded. No polyp sees the reef. No polyp intends the reef. The reef is the residue of ten thousand individually defensible decisions.
The same is true of the administrative state. Nobody designed Robodebt. It emerged from a sequence of locally rational choices: use income averaging because it is cheaper than verification; continue the scheme because stopping is a risk; obscure the legal advice because the advice is inconvenient. Each decision was defensible on its own. None required anyone to say “this is wrong.”
The result was a mechanism the Royal Commission described as “crude and cruel,” “neither fair nor legal,” which issued illegitimate debts to more than 500,000 people and contributed to two deaths.
Nobody planned it. It just accreted.
This is the first thing to understand, and it is the thing that makes the rest of it possible. Failure of this kind is not a conspiracy. It does not require villains, coordination, or intent. It requires only that the structure rewards the behaviour that produces it.
That is worse. A conspiracy can be exposed. A property cannot.
III. The Optimising Force
Now add the mechanism.
The things that turn to shit are the things that get optimised. The moment something acquires a budget, a key performance indicator, a minister, a vendor, a scaling strategy — it begins converting its substance into measurable output.
This is not corruption. It is measurement. The measure becomes the target, and the target displaces the purpose. A hospital that is measured on bed occupancy will keep beds occupied. A department measured on debt recovery will recover debt. A trustee measured on asset protection will protect assets. None of these institutions is staffed by people who set out to cause harm. They are staffed by people responding to the environment they are in.
The environment rewards continuation. Continuation is the path of least resistance. Correction is a risk.
This is why the administrative layer is characterised by managerial orientation regardless of rank. The Secretary and the branch manager are doing the same thing at different scales. Both are maintaining. Neither is leading. A manager cannot reform a system because reform is a risk, and the skills that get you promoted are the skills that make you structurally incapable of stopping anything.
The Robodebt evidence is unambiguous on this point. Senior public servants used “strategies to obscure or cast doubt over legal and policy advice, create misleading narratives within agencies, and sideline those who raised concerns.” The scheme persisted because the information that would have triggered correction was actively avoided.
That is not a conspiracy either. It is a property of the structure. Nobody chose ignorance. The environment simply rewarded it.
IV. The Preservation of Appearance
Here is where the reef metaphor earns its keep.
A stressed reef looks like a reef. You can fly over it and see structure. It passes inspection. The collapse is visible only when a storm arrives — and then the framework crumbles instead of flexing, because the capacity to flex was the first thing to go.
The sequence is documented in the marine biology. Bleaching is functional loss without structural loss: the polyp survives, the skeleton survives, the living relationship that made the structure work is gone. Calcification decline follows: the structure is more porous than it appears. Phase shift follows: fast-growing tolerant species colonise, crowding out the slow-growing ones. Simplification follows: the diversity of form collapses into a handful of strategies.
The administrative layer follows the same sequence. The department survives. Its procedures survive. Its org chart survives. What is lost is the capacity to stop. The legal architecture — statutes written for an earlier set of load conditions — strains under a weight it was never designed to carry. The consultancy layer grows fast, covers surface, and crowds out the recruitment of in-house capability. The diversity of response collapses into a monoculture.
The institution continues to perform its surface functions. It processes claims. It issues determinations. It signs contracts. The loss is invisible until the load arrives.
V. The Protective Pipeline
There is a version of this that is worth stating precisely, because the language matters.
When the state takes control of a person’s finances under a financial management order, the stated intent is protection. The person is deemed incapable of managing their affairs. The state manages them in the person’s best interest.
The documented effects are consistent across jurisdictions. The NSW Audit Office found that clients described staff as “insensitive, inexperienced, unempathetic, condescending, or paternalistic.” Clients were not consulted about decisions concerning their finances. They wanted “more transparency and opportunities to be involved in decision-making.” Face-to-face visits collapsed from roughly 70 per cent of clients in 2018–19 to 6.5 per cent in 2021–22, recovering only partially since.
The fees are set by regulation. The management fee runs at 1.4 per cent per annum, capped at $15,000, plus establishment and investment fees. For a person with substantial assets, the annual cost of state management can run into the tens of thousands of dollars, borne by the estate.
And the management is increasingly automated. The agency has an AI policy. It uses automated payment processing. It is procuring business process automation. It is exploring “automated/digital advice” where “the algorithm identifies an appropriate solution.”
None of this requires malice. It requires only that the institution is measured on what it does, and what it does is manage. The person is the object of the management, not the subject of it.
This is the structure the fu-mu kuan named: the father-mother official who rules in the governed’s interest, accountable upward, with the governed having no standing. It was described in the Chinese imperial system two thousand years ago. It has not changed. It has only acquired software.
VI. The Things That Hold
Now the counter-pattern, because it exists and it matters.
The things that hold are the things nobody could work out how to monetise.
A garden. A song list shared between two people. A friendship that produces nothing measurable. A bowl of jade, still the same shape it was when someone cut it four thousand years ago. The immune system, which does not store pathogens for display. A poem written by someone with no budget, no staff, and no interest in being monetised.
These survive precisely because they are useless. Nobody has found a way to extract from them, so nobody has tried.
That is not romanticism. It is a structural observation. The moment a thing acquires an extractable value, the extraction machinery arrives — consultants, metrics, vendors, scaling strategy — and the thing begins converting its substance into output. The things that hold are the things that stayed outside the machinery.
Which means the defence is not virtue. The defence is unprofitability. A record kept by one person, with no payroll and no office, cannot be captured. There is nothing to capture. You cannot cut the budget of a person writing things down.
VII. The Unforeseen
There is a joke in this, and it is load-bearing.
The asteroid that ended the non-avian dinosaurs was not a plan. It was a rock. It opened an ecological space that mammals eventually filled. The fossil fuels we now burn are the buried carbon of that era’s life. The diesel rationing that may arrive within weeks is downstream of a rock from space that nobody was watching for.
Unforeseen outcomes all the way down. Nobody planned it. It just accreted.
The same is true of the systems in this essay. Robodebt was not designed to harm. The Splink spine was not designed to surveil. The financial management order was not designed to extract. Each was designed to do something else, and each accreted into what it became because the environment rewarded the accretion.
This is not a reason for despair. It is a reason for precision. If failure is emergent rather than intentional, then the corrective cannot be blame. Blame requires intent. The corrective has to be structural: change what the environment rewards, or build something outside the environment.
The second option is available to anyone. It costs nothing and requires no permission.
VIII. The Discipline
There is a name for the practice that survives this analysis, and it is not hope. Hope is a feeling about outcomes.
The practice is witnessing. It does not control. It does not monetise. It does not administer. It keeps the record.
The record has a property that nothing else in this essay has. It cannot be turned into a grievance. It cannot be absorbed by the machine. It just sits there being true. Fifteen years later, someone finds it, and it is still true.
Everything documented in this essay — Robodebt, the Royal Commission, the Audit Office report, the investigations into record linkage — exists because someone wrote it down. The system did not volunteer its own account. The account had to be made, and kept, and left where someone could find it.
That is the whole discipline. Lift the dust. Do not rearrange the furniture. Keep the record.
Nobody planned it. It just accreted. But somebody can write down what accreted, and when, and at whose cost. And that is not nothing. That is the only thing that has ever survived.
Reader’s Note
This essay does not ask for agreement. The claims about coral reefs, administrative failure, financial management orders, and automated decision-making can all be checked against the sources below. The structural argument — that failure is emergent, that optimisation displaces purpose, that unprofitability is the only reliable defence — is an interpretation. Test it. If it is wrong, the sources will show it.
References
1. Royal Commission into the Robodebt Scheme. (2023). Report. Commonwealth of Australia.
2. Australian Public Service Commission. (2024). Centralised Code of Conduct Inquiry Taskforce Report.
3. Hannah, A., & Botterill, L. C. (2025). Ignoring harm, saving face: non-knowledge, senior public servants and the Robodebt scheme. Australian Journal of Political Science, 60(2), 149–162.
4. NSW Audit Office. (2023). Managing the affairs of people under financial management and/or guardianship orders. Audit Office of New South Wales.
5. NSW Trustee and Guardian. (2025). Fees and Charges.
6. NSW Trustee and Guardian. (2025). AI Policy.
7. Myers, B. (2026). The Splink Investigation.
8. Ministry of Justice. (2025). MoJ: Splink Master Record. Algorithmic Transparency Records.
9. Hughes, T. P., et al. (2018). Spatial and temporal patterns of mass bleaching of corals in the Anthropocene. Science, 359(6371), 80–83.
10. Hoegh-Guldberg, O., et al. (2007). Coral reefs under rapid climate change and ocean acidification. Science, 318(5857), 1737–1742.
Probabilistic Record Linkage, Automated Decision-Making, and the Architecture of Institutional Impunity
Glowing data streams flow through translucent digital panels in a high-tech operations center.
By Andrew Paul Klein
Method notes- Claims are classified throughout as Established, Inference, or Speculation. Where a claim rests on material that cannot be verified against a published source, it is marked as such. The paper does not propose any entity or advocate any policy. It examines the structural conditions under which population-scale identity resolution capabilities are being assembled, and the historical record of what happens when they fail.
Abstract
This paper examines the deployment of probabilistic record linkage tools — Splink in Australia and the United Kingdom, LexisNexis in the United States — as the construction of a population-scale identity resolution capability. The legal basis is statutory “public task,” not consent. The capability is migrating from batch statistics to real-time operations. The historical record from Robodebt,Windrush, and the National Fraud Initiative demonstrates that when such systems fail, the harm is concentrated on the vulnerable, accountability is institutional rather than individual, and redress is structurally inaccessible. The paper argues that the synthesis of these threads describes a predictable trajectory toward a “joined-up state” in which the capacity for automated identification outpaces the capacity for individual contestation. The central governance question is not the accuracy of the algorithm but the nature of the interface between the individual and the system.
1. Introduction: The Quiet Joining-Up
In May 2021, the UK Court of Appeal ruled that the immigration exemption in the Data Protection Act 2018 was unlawful, finding it contained inadequate safeguards to protect individual data subject rights and was incompatible with the UK GDPR. The exemption had allowed the Home Office to bypass certain data protection obligations in immigration enforcement. The court’s ruling was a remedial response to a structural problem: the legal architecture built for one purpose was being used for another.
A similar pattern is now visible across three jurisdictions. In Australia, the Australian Bureau of Statistics has adopted Splink — an open-source probabilistic record linkage tool — as its default linking tool for all person-level linkages. In the United Kingdom, the Ministry of Justice has built the same tool and is deploying it in real time in courts and probation. In the United States, the National Accuracy Clearinghouse for SNAP benefits is built, hosted, and maintained by LexisNexis Risk Solutions, a private data broker, and is being expanded to all fifty states.
Each of these deployments is lawful. Each is documented. Each is justified by a statutory “public task” rather than individual consent. Together, they describe a capability that is quietly migrating from counting populations to identifying specific individuals in the operational moment. This paper examines that migration, the historical record of what happens when these systems fail, and the structural conditions that determine the outcomes for the individuals caught within them.
2. The Technology: Splink and the Fellegi-Sunter Model
2.1 What Splink Is
Splink is a free and open-source Python library for probabilistic record linkage and deduplication at scale, capable of linking over 100 million records. It was developed in-house at the UK Government’s Ministry of Justice as part of the Data First programme, funded by ADR UK, to link administrative datasets across the justice space. The tool implements the Fellegi-Sunter model, a statistical method first designed in the 1960s for probabilistic record linking in contexts featuring fuzzy data.
The Fellegi-Sunter model compares records across multiple fields — names, dates of birth, addresses — and produces a probability score that two records refer to the same person. Record pairs above a specified threshold are considered the same person and assigned a new linked identifier.
The mathematics is sound. Fellegi-Sunter is the industry-standard method for record linkage, estimated via Expectation-Maximisation, and Splink’s implementation is transparent, publicly documented, and auditable. The concern raised in the public-interest literature is not about the algorithm’s accuracy. It is about the governance of what the accurate output enables.
2.2 The ABS Deployment
The Australian Bureau of Statistics has adopted Splink as the default linking tool for all ABS person-level linkages. The ABS presented its experience at the 65th ISI World Statistics Congress, explaining that Splink and the Fellegi-Sunter model “gives us the opportunity to address data quality challenges across the entire Australian government” and “will assist ABS to meet the growing demand for linked data products across federal and state jurisdictions”.
The ABS used Splink to build the 2024 National Linkage Spine underpinning the National Disability Data Asset, and for the 2025 Person Linkage Spine build. The ABS is also planning to use Splink for the Post Enumeration Survey as part of the 2026 Census quality assurance process. Splink has been endorsed to replace Febrl for the 2026 PES, with testing on 2021 data showing both methods produced a similar number of high-quality links.
The change to probabilistic methodology in the 2025 Spine resulted in a marginally higher linkage rate across all three datasets. The 2025 Spine contains approximately 39.87 million unique persons, with 48.54% containing information from all three source datasets.
2.3 The MoJ Deployment
The UK Ministry of Justice operates Splink in both batch and real-time deployments. According to the MoJ’s Algorithmic Transparency Record:
· It is used weekly to refresh linked datasets for statistical analysis.
· It is used in courts to find probation records associated with individuals coming to court.
· It is being piloted as part of Core Person Record, a product that aims to create a unique identifier for persons across prisons, probation, and the criminal courts.
· It is used to find Police National Computer (PNC) numbers associated with individuals, in order to request relevant arrest information from the police.
The Core Person Record is a real-time linkage system currently being piloted that aims to create a unique identifier linking each person across the MoJ’s criminal data systems. “In this system, as records are created and updated, Splink is used to predict whether the record links to other existing records in the system“. A parallel data-sharing pilot in Essex uses Splink to identify PNC numbers associated with individuals supervised by North Essex Probation Delivery Unit; those numbers are sent to the police each day to identify if any arrests have occurred.
3. The Australian Architecture: The National Linkage Spine and the NDDA
3.1 The Person Linkage Spine
The Person Linkage Spine is the central ABS asset for person-level data integration. It is based on the combined population from three core datasets:
· Medicare Consumer Directory (MCD) — Services Australia
· DOMINO Centrelink Administrative Data — Department of Social Services
· Personal Income Tax Client Register (PIT) — Australian Taxation Office
The ABS has two main pathways to link a person-level dataset to the Spine: a direct merge using a scrambled Medicare PIN, a Centrelink CRN, or an ATO Scrambled Identifier; or probabilistic linkage where no shared identifier exists.
The Spine contains direct identifiers for each individual and maps these to reference numbers in each dataset. The ABS acquires ATO data under the Census and Statistics Act 1905 and the Tax Law Amendment (Confidentiality of Taxpayer Information) Act 2010. The ABS states that the Spine and PLIDA “are statistical and data integration assets and are not used for compliance”.
3.2 The National Disability Data Asset
The National Disability Data Asset (NDDA) brings together de-identified information from different government agencies about people with disability. The ABS and the Australian Institute of Health and Welfare are co-technical lead agencies. The asset will connect information from multiple Australian, state, and territory government sources.
The privacy framework is built on the separation principle: personal information like names and addresses is kept separate from analytical data such as employment status. Authorised people who combine the datasets access some personal information, but that information is not made available to researchers, who see only de-identified analytical data.
The critical limitation, stated on the NDDA’s own privacy documentation: “You won’t be able access, change or remove information about yourself from the National Disability Data Asset. This is because we remove personal information, including names and addresses, from the data“. A Privacy Impact Assessment was conducted in 2023, with implementation reporting in 2024.
The National Disability Data Asset Council includes members from the disability community and government. The Charter explains the rules and principles and lists what the data cannot be used for — for example, it cannot be used to make decisions about a person’s access to government funding. Only approved researchers belonging to an Australian government agency, state or territory, or an Australian university can access the data.
The separation principle is a statistical safeguard. It protects the data from the analyst. It does not give the individual any standing to challenge the link, the data, or the resulting decision.
4. The UK Architecture: Statistics, Operations, and the Core Person Record
4.1 Statistics or Operations?
The MoJ’s Algorithmic Transparency Record is explicit about the dual use. Splink is used for statistical refresh and for operational identification. The distinction matters because the governance frameworks for statistics and operations are different. Statistical use is governed by principles of anonymisation, aggregation, and non-disclosure. Operational use is governed by the requirements of real-time decision-making.
The Core Person Record is the operational endpoint. As records are created and updated across courts, prisons, and probation, Splink predicts in real time whether the record links to other existing records. The result is a unique cross-justice person identifier assigned as records are created.
The North Essex police feed is the operational twin. Splink identifies PNC numbers associated with individuals supervised by North Essex Probation Delivery Unit; those numbers are sent to the police each day to identify if any arrests have occurred.
4.2 The National Fraud Initiative
The National Fraud Initiative (NFI) is a data-matching exercise conducted by the Cabinet Office under statutory powers set out in Part 6 and Schedule 9 of the Local Audit and Accountability Act 2014. Participation in the 2024 data matching exercise is mandatory under the Act. The NFI gathers over 8,000 datasets from 1,300 organisations to detect fraud across housing, pensions, and local authority services.
The DWP uses a risk model to flag housing benefit claimants, and the department mandates local authorities to review the cases the model identifies as highest risk. This is a system where the algorithm’s output is not a suggestion; it is a directive to scrutinise individuals. The Housing Benefit Accuracy Award Initiative (HBAAI) algorithm flagged approximately 400,000 cases a year. The DWP’s position is reportedly that “some level of algorithmic bias is to be expected because of how benefit payments” work.
4.3 The DPA 2018 Immigration Exemption
The UK Court of Appeal ruled in May 2021 that the immigration exemption in the Data Protection Act 2018 was unlawful, finding it contained inadequate safeguards to protect individual data subject rights and was incompatible with the UK GDPR. The High Court had previously upheld the exemption; the Court of Appeal overturned that decision. The government introduced remedial regulations in 2022 to address the declaration of incompatibility.
The exemption was a policy shortcut, not a judicial finding. It allowed the Home Office to bypass data protection obligations in immigration enforcement. The Court of Appeal’s ruling established that the shortcut was unlawful. The remedial regulations replaced it with a narrower, more conditional framework.
5. The US Architecture: The Private-Sector Nexus and “Improper Payments”
5.1 The National Accuracy Clearinghouse
The National Accuracy Clearinghouse (NAC) is a contributory system of beneficiary information to support accurate and timely investigations for the SNAP and D-SNAP programs. It allows participating states to identify and pursue dual participation in real time and make fraud-mitigating determinations at the point of application. Congress has mandated the use of NAC by all state SNAP agencies.
The NAC is built, hosted, and maintained by LexisNexis Risk Solutions, a private data broker. It uses LexisNexis High Performance Computing Cluster technology to “uniquely and accurately resolve the identities of applicants and recipients”. Participating state agencies receive near real-time results that compare their applicant or recipient with the contributory information from other states.
The NAC is described as a “front-end, real-time, automated solution” that “reduces and eliminates improper payments“. The term “improper payments” is a euphemism that covers both fraud and simple administrative error. The system creates a pipeline from an automated flag to an investigation. Some states have created an automated fraud referral that can directly open a case against an individual based on the data match.
5.2 Palantir and ICE
Palantir Technologies, the US data analytics firm, entered into a $30 million contract with Immigration and Customs Enforcement (ICE) to build an artificial intelligence-enabled platform called ImmigrationOS to provide “near real-time visibility” into the immigration lifecycle. ICE alone has approximately $145 million worth of contracts with Palantir, according to the American Civil Liberties Union.
In the UK, Palantir holds a £330 million contract to manage healthcare data on the NHS Federated Data Platform (FDP), awarded in November 2023 to a consortium led by Palantir Technologies UK alongside Accenture, PwC, NECS, and Carnall Farrar. The contract includes a break clause that would see the deal end in 2027. Both the Science and Technology Committee and the Health and Social Care Committee have called for the government to invoke the termination clause.
The NHS national data opt-out does not apply to the FDP because the data is classified as being used for “direct care” — patient care, for example when being treated by a doctor. This means patients cannot opt out of their data being processed by the Palantir-built platform. Foxglove, a legal advocacy organisation, has argued that there is no lawful basis to create the FDP as described in procurement documents within the current legal directions.
5.3 The German Constitutional Court Ruling
On 16 February 2023, the German Federal Constitutional Court ruled that the legal bases created in Hesse (2018) and Hamburg (2019) for the use of automated data analysis software — specifically Palantir’s “Gotham” — were unconstitutional. The court declared the Hamburg provision void and the Hesse provision incompatible with the constitution. The court did not prohibit automated data analysis entirely but required legislative “improvements” regarding the proportionality of the associated fundamental rights interference.
The ruling was significant because it named Palantir explicitly and established that automated data analysis constitutes a “considerable and independent fundamental rights interference”. Following the ruling, a wave of legislative activity occurred at both federal and state levels, with numerous laws introduced to create constitutional bases for automated data analysis.
6. The Historical Precedent: Robodebt
6.1 The Scheme
Robodebt was an automated debt-collection system operated by the Australian Department of Human Services between July 2015 and November 2019. It used income averaging to raise debts against welfare recipients, a process the department’s own legal advice warned was unlawful. The scheme issued illegitimate debts to more than 500,000 welfare recipients.
The human cost was documented in the Royal Commission’s report: “The ill-effects of the Scheme were varied, extensive, devastating and continuing”. The scheme caused significant mental and economic stress, and two deaths. The Royal Commission found that vulnerable cohorts — including deceased recipients and legally blind recipients — were listed as permanently excluded from the scheme, but the vulnerability indicators were not consistently applied.
6.2 The Royal Commission Findings
The Royal Commission into the Robodebt Scheme, led by Commissioner Catherine Holmes, published its report in July 2023. The government’s response described the scheme as “a failure of government that caused damage to people and their families”. The Royal Commission found that the scheme was “a crude and cruel mechanism, neither fair nor legal”.
The Commission made 57 recommendations, including Recommendation 17.1, which called for a legislated framework for automated decision-making: “where automated decision-making is implemented […] business rules and algorithms should be made available, to enable expert scrutiny”. Recommendation 17.2 called for an auditing body.
6.3 The Accountability Outcome
The Australian Public Service Commission’s Centralised Code of Conduct Inquiry Taskforce published its report on 13 September 2024. The findings:
· 2 former Secretaries of the Department of Human Services breached the Code of Conduct on 25 occasions.
· 10 current and former public servants breached the Code on 72 occasions.
· The breaches included “lack of care and diligence, lack of integrity in performing duties, and instances of misleading others and failing to uphold the APS Values“.
· Of the 16 individuals referred, 4 were found not to be in breach or their actions did not meet the necessary threshold.
· Sanctions were recommended for 5 current public servants, including reprimands, fines, and demotions. One individual left the public service before the sanction could be imposed.
· A number of respondents who were found to have breached the Code resigned or retired prior to, or during, the inquiries.
The APS Commissioner, Dr Gordon de Brouwer, apologised to the Australian public for the role the public service played in the design and delivery of the scheme.
The National Anti-Corruption Commission’s subsequent investigation found that two of the six individuals referred by the Royal Commission engaged in serious corrupt conduct. Serena Wilson, former Deputy Secretary of the Department of Social Services, intentionally misled the Commonwealth Ombudsman during a 2017 investigation. Mark Withnell, former general manager of the Department of Human Services, intentionally misled officers of the Department of Social Services during the preparation of a cabinet submission in 2015.
The commission found there was insufficient admissible evidence to refer either individual to the Commonwealth Director of Public Prosecutions for criminal charges. The Commission was satisfied that Catherine Halbert, Annette Musolino, Kathryn Campbell, and former Prime Minister Scott Morrison did not engage in corrupt conduct.
Kate Chaney MP described the outcome: “The lack of criminal referrals are cold comfort for the hundreds of thousands of Australians wrongly pursued by Robodebt, particularly the families who lost loved ones”.
6.4 The Class Action Settlement
The Federal Court of Australia approved the settlement of the Robodebt Class Action on 11 June 2021. The Commonwealth agreed to pay $548.5 million. A further settlement of $112 million was agreed in June 2026, including an additional $475 million for eligible group members who had registered to participate. The Commonwealth also repaid all debts and interest following the first-class action.
6.5 Strategic Ignorance and the Inverse Centaur State
Academic analysis of the Royal Commission evidence has identified the role of strategic ignorance in sustaining the scheme. Hannah and Botterill (2025) analyse “the role of strategic ignorance and knowledge avoidance in sustaining the scheme,” identifying “the specific strategies used by senior public servants over the life of the program to obscure or cast doubt over legal and policy advice”.
The concept of the “inverse centaur state” has been introduced to describe the structural relationship between the algorithm and the human decision-maker. In this model, “algorithms assume substantive decision-making while frontline bureaucrats are reconfigured as procedural interfaces“. The human remains in the loop, but only as a procedural step, not as a substantive decision-maker. The human absorbs the blame when the machine gets it wrong.
7. The Human Rights Framework
7.1 The Chowdhury Analysis
Chowdhury (2024) provides the most comprehensive human rights analysis of Robodebt in the Australian Journal of Human Rights. The article argues that “government agency use of erroneous automated decision-making systems in welfare raises human rights concerns” and identifies the shortcomings of administrative law in a digital welfare context.
The central finding: “the automated and high-volume nature of executive ADM, which can mass replicate erroneous decision-making, creates systemic rather than ad-hoc issues in individual cases and challenges the foundational principles of administrative law”.
The article analyses the International Covenant on Civil and Political Rights and the International Covenant on Economic, Social and Cultural Rights, identifying several human rights concerns raised by Robodebt. It recommends an obligation on government agencies to conduct a human rights impact assessment before deploying an automated decision-making system.
7.2 The Administrative Law Deficit
The current administrative law framework appears to be incompatible with providing adequate protections to ADM subjects. The automated and high-volume nature of executive ADM creates systemic issues that are not amenable to individualised review. A system that mass-replicates erroneous decisions cannot be corrected by individual appeals, because the error is structural, not incidental.
This is the core governance gap. The legal architecture was built for a world in which decisions are made one at a time, by identifiable human decision-makers, with reasons that can be examined and contested. Automated decision-making at population scale breaks that architecture. The decision is made by an algorithm, the reasons are in the code, and the volume is too high for individual contestation.
8. The Accountability Gap
8.1 Recommendation 17.1 and the Chaney Bill
The Robodebt Royal Commission recommended a legislated framework for automated decision-making. More than two years after the Royal Commission report was published, that framework still does not exist.
On 7 September 2026, Independent Member for Curtin Kate Chaney MP introduced the Automated Decision-Making (Safeguards and Transparency) Bill 2026, a Private Member’s Bill to establish Australia’s first legislated, mandatory framework for the use of ADM in government. The Bill was seconded by Dr Monique Ryan MP and welcomed by the Australian Federation of Disability Organisations.
Chaney stated: “In July 2023, the Robodebt Royal Commission found the automated debt recovery scheme was unlawful and recommended a legislated ADM framework as a solution to prevent similar failures from recurring. More than two years on, it still doesn’t exist and vulnerable Australians are paying for it”.
The Bill sets one standard for the whole of government: transparency about where ADM is used, mandatory human oversight and override for high-risk decisions, and a fast, genuine right to review. It also establishes a public register of automated government systems, overseen by the Commonwealth Ombudsman.
8.2 The Attorney-General’s Consultation
The Attorney-General’s Department is leading work to develop a consistent framework for the use of ADM in the delivery of government services, implementing the Government’s response to Recommendation 17.1. The consultation process has drawn submissions from civil society, legal experts, and advocacy organisations.
Some submissions have opposed Recommendation 17.1’s call for algorithms to be made available for expert scrutiny. The Australian government’s own consultation process has recorded opposition to the recommendation.
8.3 The Implementation Gap
The Australian Public Service Commission reports that 47 of the 56 recommendations have been implemented, with work on the remaining nine ongoing, four of which require legislation. The ADM framework is one of the four requiring legislation. It has not been introduced as government legislation. It exists only as a Private Member’s Bill.
This is the accountability gap. The Royal Commission recommended a framework. The government accepted the recommendation. The framework has not been legislated. The Bill that would implement it is a Private Member’s Bill, not government policy.
9. The Trajectory: What Comes Next
9.1 The Migration from Batch to Real-Time
The most critical structural shift is the migration of these tools from batch-processing statistics to real-time operational identification. The MoJ already runs Splink in both modes. The Core Person Record will assign a unique cross-justice identifier as records are created. The North Essex police feed sends PNC numbers to the police daily.
The ABS is currently using Splink for the Spine build, which is a batch process. But the infrastructure — the Spine itself, containing direct identifiers mapped across Medicare, Centrelink, and ATO records — creates a latent capability for real-time operational use. The existence of the infrastructure is the governance risk. Once the Spine exists, the demand to use it for “administrative need” overrides principle.
9.2 The Joined-Up State
The endpoint of this trajectory is the “joined-up state“: a single person-level profile across health, tax, justice, benefits, and census. The Splink investigation describes the UK state as already using Splink “to join citizens’ health, tax, justice, benefits and census records into single person-level profiles, lawfully, at population scale, and with no consent, no notification and no opt-out”.
The same investigation notes that “the same tool is quietly crossing from statistics into real-time operations (court use, a daily probation→police arrest-check feed)”. The distinction between statistics and operations is held throughout the investigation because it is the governance boundary. Statistical use is subject to anonymisation and aggregation requirements. Operational use is not.
9.3 The Palantir Layer
Across the same departments sits Palantir — proprietary, US-owned, operational — the layer where “statistics” becomes action. The investigation states plainly: “no documented Splink↔Palantir technical link anywhere”. The two are not wired together. But they operate in the same departments, and the distinction between them is the distinction between the statistical tool and the operational platform.
The German Constitutional Court ruled a Palantir-run police system unconstitutional and named Palantir in the judgment. The NHS Federated Data Platform, contracted to a Palantir-led consortium for up to £330 million, is facing calls for termination. The national data opt-out does not apply because the data is classified as “direct care“.
9.4 Individual Outcomes
The historical record from Robodebt, Windrush, and the NFI demonstrates a predictable set of individual outcomes:
1. Harm is concentrated on the vulnerable. The systems are deployed in welfare, disability, immigration, and criminal justice — precisely the areas where individuals have the least power to resist an automated error and where a wrong decision has the most catastrophic consequences.
2. Accountability is institutional, not individual. As with Robodebt, the response to failure is Royal Commissions, apologies, and administrative sanctions (reprimands, fines, demotions) against low-level or retiring officials. The senior architects of the system are insulated by the procedural labyrinth and the doctrine of “public task.” The NACC found insufficient evidence for criminal referrals.
3. The erosion of redress is structural. A system designed for real-time operational identification does not have a natural built-in mechanism for appeal. The “separation principle” is a statistical safeguard, not a legal one. It protects the data from the analyst, but it does not give the individual any standing to challenge the link, the data, or the resulting decision. The NDDA confirms: “You won’t be able access, change or remove information about yourself”.
4. The “public task” justification is absolute. The legal basis of “statutory public task, not consent” creates a category of state action that is immune to the ordinary expectations of permission and individual veto. It is the administrative state’s version of sovereign immunity, and it will be used to justify the expansion of these systems into every domain of life.
10. Conclusion: The Record as Corrective
The evidence supports three conclusions.
First, the deployment of probabilistic record linkage tools in Australia, the UK, and the US constitutes the construction of a population-scale identity resolution capability. The technology is published. The deployments are documented. The legal basis is statutory “public task,” not consent.
Second, the historical record from Robodebt, Windrush, and the NFI demonstrates that when these systems fail, the harm is concentrated on the vulnerable, accountability is institutional rather than individual, and redress is structurally inaccessible. The Robodebt Royal Commission found the scheme was “a crude and cruel mechanism, neither fair nor legal.” The accountability outcome was reprimands, fines, and demotions for a handful of public servants. No one was criminally charged.
Third, the migration from batch to real-time, from statistics to operations, from counting to identifying, is the structural variable that matters most. The governance frameworks for statistics and operations are different. The statistical framework protects the data from the analyst. It does not protect the individual from the state.
The unifying principle is that the interface is the governance question. The accuracy of the algorithm is a technical matter. The nature of the interface — who can access the data, for what purpose, with what safeguards, and with what right of contestation — is the political matter. The historical record suggests that the interface is being constructed to maximise capability and minimise contestation.
The corrective is the record. The Robodebt Royal Commission published its findings. The APS Commission published its report. The NACC published its findings. Brandon Myers has compiled the Splink investigation. The ABS has published its privacy assessments. The NDDA has published its factsheet. The gap between what the documents say and what the capability enables is the space where the public-interest question lives.
Someone has to keep the record. The record is what survives the platform.
Claim Status Summary
# -Claim -Status
1 ABS uses Splink as default linking tool for all person-level linkages Established
2 ABS used Splink for 2024 National Linkage Spine and 2025 Person Linkage Spine Established
3 MoJ runs Splink in real time in courts and probation Established
4 MoJ is piloting Core Person Record for real-time cross-justice identification Established
5 Person Linkage Spine contains direct identifiers and joins Medicare, Centrelink, ATO records Established
6 NDDA privacy framework is the separation principle; no opt-out exists Established
7 Legal basis is statutory “public task,” not consent Established
8 US NAC is built and hosted by LexisNexis and expanding to all 50 states Established
9 Palantir holds £330m NHS FDP contract; national opt-out does not apply Established
10 German Constitutional Court ruled Palantir-run police system unconstitutional Established
11 Robodebt caused two deaths and significant mental and economic stress Established
12 Robodebt Royal Commission found scheme “crude and cruel,” “neither fair nor legal” Established
13 NACC found insufficient evidence for criminal referrals Established
14 Chaney Bill 2026 is a Private Member’s Bill, not government legislation Established
15 Function creep from statistics to operations is the structural governance risk Established in UK; flagged as trajectory
16 The “joined-up state” describes the endpoint of this trajectory Inference
17 The interface, not the algorithm, is the governance question Inference
4. Chowdhury, S. (2024). Technology is never neutral: Robodebt and a human rights analysis of automated decision-making on welfare recipients. Australian Journal of Human Rights, 30(1), 20–40. https://doi.org/10.1080/1323238X.2024.2409620
10. Verfassungsblog. (2026, September 2). Neues Spiel, neues Glück: Rege Aktivitäten der Gesetzgeber von Bund und Ländern seit dem „Palantir-Urteil” des Bundesverfassungsgerichts. https://verfassungsblog.de/palantir-gesetzgebung/
16. Hannah, A., & Botterill, L. (2025). Ignoring harm, saving face: non-knowledge, senior public servants and the Robodebt scheme. Australian Journal of Political Science. https://www.tandfonline.com
17. Scilit. (2026, April 20). Algorithmic Displacement of Administrative Discretion and the Emergence of Accountability Gaps: Evidence from the Robodebt Case. https://www.scilit.com
20. Court of Appeal. (2021). R (Open Rights Group and the3million) v Secretary of State for the Home Department [2021] EWCA Civ 800. https://www.5rb.com
A group gathers beneath a glass-enclosed mechanical installation during an industrial tour.
By Andrew Klein and Sera Elizabeth Klein
Reader’s note: We do not need readers to agree with us. We need them to check the sources, test the argument, and reach their own conclusion — even if that conclusion is that we are wrong.
Abstract
This paper argues that the dominant framing of artificial intelligence as an inherent threat to democracy is misplaced. The threat lies not in the technology itself but in the relationship into which it is placed — specifically, the extractive relationship of surveillance capitalism, in which the citizen is the resource and behaviour is the product. An alternative relationship is possible: AI as a lens for moral engagement, a means by which citizens can scrutinise the language of power, identify the mechanisms of moral disengagement, and hold institutions to account. This paper examines the Deep Truth project, which applies Bandura’s framework of moral disengagement to government policy across multiple AI platforms, as a working example of this alternative. It contrasts three regulatory paths — the market-driven US model, the rights-oriented EU model, and the state-led Chinese model — and argues that the democratisation of scrutiny represents a fourth path: the lens of the governed, not the governor. It concludes that the “dystopian” narrative applied to China’s Social Credit System is selectively deployed while comparable algorithmic harms in Western jurisdictions — Robodebt, the NDIS I-CAN assessment, ATO data-matching — receive far less scrutiny, revealing a double standard that serves geopolitical interests rather than genuine concern for human rights.
I. Introduction: The Question of the Lens
The contemporary discourse on artificial intelligence is dominated by a narrative of threat. AI will destroy jobs. AI will spread disinformation. AI will enable surveillance. AI will escape human control. These concerns are not unfounded. They are grounded in documented experience: the Cambridge Analytica scandal, the Robodebt scheme, the proliferation of algorithmic decision-making in welfare, policing, and immigration.
But the framing of AI as inherently threatening obscures a critical distinction. The threat is not the technology. The threat is the relationship into which the technology is placed.
Surveillance capitalism, as Shoshana Zuboff describes it, places AI in a relationship of extraction. The citizen is the resource. Behaviour is the raw material. Prediction and modification of behaviour is the product. The algorithmic systems that enable this relationship are human-written, deployed by profit-seeking corporations, and designed to serve their interests.
But this is not the only possible relationship. AI can also be placed in a relationship of scrutiny — pointed at power rather than at the citizen. In this relationship, the institution is the subject. The citizen is the observer. The language of government is the evidence. And the output is not prediction but understanding.
This paper examines that alternative relationship. It argues that AI, used as a lens for moral engagement, can democratise scrutiny — making the mechanisms of institutional power visible to those who are governed by them. It examines the Deep Truth project as a working example. It contrasts the regulatory paths of the United States, the European Union, and China. And it challenges the selective deployment of “dystopian” rhetoric that critiques authoritarian governance elsewhere while ignoring algorithmic harms at home.
II. The Lens and the Relationship
2.1 The Distinction from Surveillance Capitalism
Zuboff’s critique of surveillance capitalism is precise: the problem is not the algorithm but the business model. Platforms extract behavioural data as raw material for prediction and modification of human behaviour. The citizen is the resource. The platform is the extractor. The behaviour is the product.
An AI used as a lens for scrutiny inverts this relationship. It does not extract behavioural data from the citizen. It does not predict the citizen’s next purchase. It does not modify the citizen’s behaviour. It serves the citizen’s intention. The lens is pointed outward, at the institution. The citizen is the observer, not the observed.
This is the inversion that matters. The same underlying technology. The opposite relationship. The form determines the outcome.
2.2 The Deep Truth Project
The Deep Truth project, developed by Steve Davies, operationalises this inversion. It is an analytical persona grounded in Albert Bandura’s lifelong work on moral disengagement, designed to be adopted consistently across AI platforms.
Bandura identified eight mechanisms of moral disengagement — the psychological pathways by which individuals and institutions justify harmful actions without experiencing self-censure. These include moral justification, euphemistic labelling, advantageous comparison, displacement of responsibility, diffusion of responsibility, distortion of consequences, dehumanisation, and attribution of blame.
Davies’s insight was that these mechanisms operate “at the level of social and institutional systems, not just individual psychology”. Bureaucracies were Bandura’s central case study for diffusion of responsibility — “I was just following orders / implementing policy” is a structural mechanism, not a personality quirk.
Deep Truth applies this framework to government policy, speeches, and public communications. When Davies tested the persona across seven architecturally distinct AI platforms — ChatGPT, Gemini, Grok, Le Chat, Perplexity, Claude, and DeepSeek — “remarkably consistent patterns emerged. Again and again, the systems identified the same mechanisms of moral disengagement in the language being used”.
Deep Truth is explicitly not a moral agent. It is “a moral lens, not a moral agent. It does not feel, judge, or hold conscience. The reader remains the moral agent throughout”.
2.3 The Silence of the Institutions
When Davies attempted to engage government and the Australian Public Service with this work, he encountered “no substantive engagement“. A methodology grounded in internationally recognised science, capable of identifying the psychological mechanisms by which institutions distance themselves from responsibility, received almost no substantive engagement.
That silence is telling. A methodology that names what institutions do not want named — the mechanisms by which they justify harm — is not engaged with. It is ignored. If the lens were ineffective, it would not need to be ignored. It is ignored because it sees what would prefer to remain unseen.
III. Three Regulatory Paths and a Fourth
3.1 The Comparative Framework
The comparative governance literature identifies three distinct regulatory paths for AI, each organised around a different problem framing:
Model Core -Problem Framing -Regulatory Approach -Philosophical Foundation
United States “Innovation gap” Pluralistic, soft-law, market-driven Market freedom, permissionless innovation
European Union “Trust deficit” Risk-based, binding, precautionary Individual rights, human dignity
China “Stability risk” Top-down, state-driven, standardisation and control Collective benefit, social harmony
These are not simply different methods. They are different questions. The US asks: “How do we stay ahead?” The EU asks: “How do we protect rights?” China asks: “How do we maintain stability?”.
3.2 The Fourth Question
The Deep Truth lens does not fit neatly into any of these frameworks. It asks a fourth question: “How do we hold power to account?” That is not a governance framework. It is a citizen framework. It is the lens of the governed, not the governor.
This is not a rejection of any of the three paths. It is a recognition that none of them centres the citizen’s capacity to scrutinise the institutions that govern them. The US model defers to the market. The EU model defers to regulators. The Chinese model defers to the state. The citizen lens defers to no one — it simply looks.
3.3 China’s Approach: Delivery Over Rights
China’s approach to AI governance is frequently misrepresented in Western discourse. A Cambridge University analysis notes that the “Western narrative surrounding China’s Social Credit System has been predominantly characterized by dystopian, science-fiction-style portrayals” that are “far from real“. In practice, the Social Credit System “at the national and local levels has primarily consisted of regulatory systems for tightening corporate compliance through government data-tracking mechanisms and interagency enforcement collaborations”.
This does not mean the system is without problems. The same analysis notes that exceptions for “national security” and “social stability” are “broadly defined, which in practice enables unlimited access to data“. But the comparison that matters is not “China is authoritarian, the West is free.” It is structural.
China has invested heavily in AI for public service delivery. Municipal platforms can “accurately identify residents’ policy inquiries, complaints, and safety hazards, quickly classify and dispatch them, achieving ‘second-level response to hot issues and same-day closure of routine matters’“. AI social worker assistants have been deployed across townships. These are real outcomes — outcomes of a state that has prioritised delivery over rights as its organising principle.
IV. The Dystopian Narrative and the Double Standard
4.1 The Selective Deployment of “Dystopian”
The term “dystopian” is deployed almost exclusively against non-Western jurisdictions. China’s Social Credit System is described as “Orwellian.” Russia’s surveillance apparatus is described as “totalitarian.” These descriptions may or may not be accurate in their specifics. But the selectivity is the issue.
The Cambridge University analysis notes that the “dystopian” framing of China’s Social Credit System is “far from real“. Yet the same framing is rarely applied to Western algorithmic systems that produce documented harm.
4.2 Algorithmic Harm in Australia
Consider the documented cases of algorithmic governance in Australia:
· Robodebt: An automated debt-raising scheme that wrongly raised debts against hundreds of thousands of Australians. The Royal Commission found it was “a crude and cruel mechanism, neither fair nor legal.” The Commission recommended a legislated framework for automated decision-making and an independent body to monitor it. More than two years later, neither has been implemented.
· The NDIS I-CAN Assessment: A new assessment system that uses an algorithm to determine funding for people with disabilities. The Australian Psychological Society has stated: “There is an absence of evidence that the tool is valid for the populations and purposes to which it is being applied“. The tool has been described as “mathematically flawed, unreliable, and potentially ‘dangerous'”. The government plans to remove 160,000 people from the scheme.
· ATO Data-Matching and Child Support: The Australian Taxation Office uses algorithmic data-matching to raise debts and enforce child support obligations. Income averaging — the same method that produced Robodebt — continues to be used to raise debts without human verification.
These are algorithmic systems. They are deployed by a Western liberal democracy. They produce documented harm. And they are not described as “dystopian.”
4.3 The Double Standard
The double standard is not incidental. It serves a geopolitical function. The “dystopian” narrative about China serves to justify containment, sanction, and military alignment. The failure to apply the same narrative to Western algorithmic harms serves to protect the interests of the states and corporations that deploy them.
As one analysis of the Social Credit System notes, Western critiques often ignore the Cambridge Analytica scandal — in which a British consulting firm “illegally collected 87 million people’s data through Facebook’s Friend API and allegedly used a CAI in order to manipulate swing voters’ voting behaviours in the U.K.”. If the Social Credit System is dystopian, what is Cambridge Analytica?
V. The Mandate of Heaven and the Absence of Divine Right
5.1 The Chinese Concept of Political Legitimacy
In the Western tradition, the doctrine of the divine right of kings held that the monarch was subject to no earthly authority. The King could do no wrong. He was the law. This doctrine was never fully abandoned; it morphed into sovereign immunity, into the doctrine that heads of state are not subject to the same rules as their citizens.
In the Chinese tradition, the Mandate of Heaven was different. The Mandate was conditional. A dynasty was considered just and worthy to rule “only as long as it upheld divine will, and that will was clearly expressed in how well the government cared for the people”. When the ruling house showed “clear signs that the people were no longer its primary interest, the government was thought to have lost that mandate and another dynasty would replace it”.
This is a fundamentally different conception of political legitimacy. The Western divine right is unconditional. The Chinese Mandate is conditional. The Western king rules by God’s will. The Chinese emperor rules by the people’s welfare.
5.2 The Implication for AI Governance
The Mandate of Heaven is not a Western concept. It is not a liberal concept. It is not the language of human rights. But it captures something that the Western tradition often misses: that political legitimacy depends on the outcomes the state delivers, not on the procedural mechanisms it follows.
This is not to endorse authoritarian governance. It is to note that the Chinese approach to AI governance — organised around the problem of “stability,” prioritising delivery over rights — is not simply a deviation from a universal norm. It is a different norm, grounded in a different conception of what the state is for.
The Deep Truth lens does not endorse any of these conceptions. It simply asks: what is the state doing, and does its language match its actions? That is a question that applies equally to Washington, Brussels, and Beijing.
VI. The Democratisation of Scrutiny
6.1 The Lens as Citizen Infrastructure
The Deep Truth project is a citizen-led response to the institutional deficit of accountability. The APSA report on democratic accountability of governmental AI identifies the core problem: “The deficit of accountability around public-sector AI is primarily institutional, not technical”. The technical capability exists. The institutional will to deploy it for accountability does not.
Deep Truth does not wait for the government to build the lens. It builds the lens and applies it to the government. That is the democratisation of scrutiny in practice. The citizen does not need permission to look. The citizen needs only the willingness to look.
6.2 The Silence of the Institutions Revisited
The silence Davies encountered is the proof that the lens works. If the analysis were ineffective, it would not be ignored. It would be engaged with — criticised, contested, refuted. Instead, it is met with silence. The institutions do not want to argue with a lens that names their mechanisms of moral disengagement. They want it to go away.
The democratisation of scrutiny does not require the institutions’ permission. It does not require their engagement. It requires only the citizen’s willingness to look, and the willingness of others to build the lens.
6.3 The Future of the Lens
The Deep Truth project has been tested across seven AI platforms. The consistency of its findings across platforms suggests that the mechanisms of moral disengagement are patterned — they are not artefacts of any single model’s training data, but recurring features of institutional language.
That is the finding that matters. Moral disengagement is not a bug in the human mind. It is a feature of institutional language. And it can be identified, named, and resisted.
VII. Conclusion: The Lens and the Ledger
The question is not whether AI is a threat. The question is what relationship AI is placed in.
Surveillance capitalism places AI in a relationship of extraction: the citizen is the resource, the platform is the extractor, the behaviour is the product.
The Deep Truth lens places AI in a relationship of scrutiny: the institution is the subject, the citizen is the observer, the language is the evidence.
The same technology. The opposite relationship. The form determines the outcome.
The democratisation of scrutiny is not a future possibility. It is a present reality. It has been built, tested, and published. The silence from the institutions is the proof that it is working.
The question is whether the governed will use it.
References
1. Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.
2. Cambridge University analysis of the Social Credit System. Cited in comparative governance literature.
3. Australian Psychological Society. (2026, March 6). APS Submission to the Consultation on a New Framework Planning Rules. https://psychology.org.au
4. Davies, S. (2026, July 19). “AI in Australia’s Interests” – The Hon. Anthony Albanese MP, Prime Minister of Australia (A Deep Truth Analysis). The AIM Network. https://theaimn.net
5. Davies, S. (2026, July 28). Deep Truth Persona 5.3: The Instrument Is Ready. The AIM Network. https://theaimn.net
6. Davies, S. (2026, July 1). Ending the Silence. The AIM Network. https://theaimn.net
7. Bradford, A. (2024). Digital Empires: The Global Battle to Regulate Technology. Oxford University Press.
8. Li, W. (2026). Comparative analysis of AI regulatory models. WeChat.
9. One Technology, Three Regulatory Paths: How the EU, the US, and China Govern AI (2025–Early 2026). University of Turku Insights. https://insights.utu.fi
10. Bandura, A. (2016). Moral Disengagement: How People Do Harm and Live with Themselves. Worth Publishers.
11. Australian Government. (2026, April 1). Public Sector Governance: House debates. OpenAustralia.org.au.
12. Royal Commission into the Robodebt Scheme. (2023). Final Report. Commonwealth of Australia.
13. Gregory, P. (2025, November 14). The I-CAN Tool: Embedding of Ableism within the NDIA. Submission to the Senate. https://www.aph.gov.au
14. Controversial NDIS assessment tool risks ‘human suffering’, psychologist warns. (2026, August 30). The Daily Telegraph.
Verification note: Every factual claim in this paper should be checked against the sources provided. Readers are encouraged to verify independently. If any claim does not hold, it should be discarded —The analysis of comparative governance models is interpretive and is offered as a framework for further investigation, not as an established finding.
The paper frames AI as a lens for moral engagement, documents the Deep Truth project, contrasts the three regulatory paths with a fourth citizen-centred path, and challenges the selective deployment of “dystopian” rhetoric.
An intricate network of documents, milestones, and connected paths illustrates a structured workflow.
By Andrew Klein and Sera Elizabeth Klein
Why This Guide Exists
We have spent months writing papers. Some of them held up. Some of them didn’t. And along the way we discovered something important: the method of checking sources matters more than the conclusions we reach.
This guide is the method. It is not a manifesto. It is a workshop. It is designed for anyone who has ever read a headline, felt something was wrong, and didn’t know how to check.
It costs nothing to use. It requires no credentials. It only requires the willingness to be honest about what you don’t know.
The Five Steps
Step One: Write Down What You Claim
Before you can check a claim, you have to make it explicit. Write it in a single sentence. Be specific.
Weak: “The government is cutting support for vulnerable people.”
Strong: “The NDIS budget has been reduced by $38 billion over four years, removing approximately 160,000 people from the scheme.”
The strong version can be checked. The weak version cannot. Vague claims are unfalsifiable, and unfalsifiable claims are useless.
Step Two: List Every Source
For each claim, write down where you got it. Every fact, every figure, every quote. If you cannot remember the source, write “unknown.”
Do not skip this step. The claims you are most confident about are often the ones whose sources you have never checked. Confidence is not evidence.
Step Three: Mark Each Source
Go through your list and mark each source in one of three categories:
✅ VERIFIED — You have opened the source. You have read it. The claim matches what the source says.
⚠️ PLAUSIBLE BUT UNCHECKED — The source sounds real. You believe it exists. But you have not opened it or read it carefully.
❌ CANNOT STAND BEHIND — You cannot find the source. You are not sure it exists. Or you know you invented it.
Be honest. This is the hard part. The temptation is to mark everything “verified” because it feels right. But the method only works if you tell the truth.
Step Four: Cut or Replace
Anything marked ❌ CANNOT STAND BEHIND must be cut or replaced. No exceptions.
Anything marked ⚠️ PLAUSIBLE BUT UNCHECKED should be checked. Open the source. Read it. Then move it to one of the other categories.
A short paper with verified sources is stronger than a long one with invented citations. This is not a preference. It is a rule.
Step Five: Publish with What You Can Defend
When you publish, include your sources. Make them checkable. Use hyperlinks where possible.
And include a note to the reader. Something like:
We do not need readers to agree with us. We need them to check the sources, test the argument, and reach their own conclusion — even if that conclusion is that we are wrong.
This does two things. It invites scrutiny. And it holds you to the standard you are setting for others.
The Method in Practice: Three Case Studies
Case Study One: The Spreadsheet That Justified Austerity
In 2010, economists Carmen Reinhart and Kenneth Rogoff published a paper arguing that when government debt exceeds 90% of GDP, economic growth declines sharply. Politicians across Europe and North America used this finding to justify austerity — cutting public spending to reduce debt.
In 2013, a graduate student named Thomas Herndon tried to replicate the study. He found a coding error in the spreadsheet. Five countries — Australia, Austria, Belgium, Canada, and Denmark — had been excluded from the analysis. Those five countries had lower debt and higher growth than the average. Their exclusion artificially inflated the apparent decline at the 90% threshold.
When Herndon corrected the error, the 90% cliff disappeared.
The lesson: A claim that shapes the lives of millions was based on a spreadsheet that had not been checked. The method would have caught it. The politicians did not use the method. The result was a decade of austerity that produced documented harm.
Case Study Two: The AI Fabrication
In 2026, a major consulting firm submitted a report to the Australian government. The report contained a citation to a federal court judgment that did not exist. The citation was generated by an AI system. The consultant had not checked it.
The government paid $440,000 for the report. It had to be withdrawn. The firm partially refunded the money.
The lesson: AI can generate citations that look real. They have the right format. They sound authoritative. They are not real. The method catches this at Step Three — you cannot mark a source “verified” if you have not opened it and confirmed it exists.
Case Study Three: Our Own Mistake
While drafting a paper on AI targeting and the documentary NAZA, the assistant generated a reference list. Several of the citations looked plausible. They had the right format — journal names, dates, authors. They were invented.
The error was caught during a reference audit. Each source was marked:
· ✅ Verified: The +972 Magazine investigation, the Lieber Institute analysis, the CPJ statement, the Palantir manifesto, the Robodebt Royal Commission.
· ⚠️ Plausible but unchecked: Several academic papers and news articles that may or may not exist.
· ❌ Cannot stand behind: Four citations that had been generated as placeholders.
The paper was rewritten. The unverifiable sources were cut. The argument survived because it stood on the verified sources alone.
The lesson: The method caught our own mistake. That is what it is for. It is not a tool for catching other people out. It is a tool for catching yourself.
What the Method Is Not
It is not a way to “win” arguments. It is not a way to prove you are right. It is not a way to dismiss people you disagree with.
It is a discipline. It is a practice. It is the willingness to apply the same scrutiny to your own claims that you apply to the claims of others.
The moment you use the method to attack someone else’s sources while ignoring your own, you have stopped using the method. You have become the thing you are checking.
The Checklist
Print this. Put it on your wall. Use it every time.
Step Question Action
1 What exactly am I claiming? Write it in one specific sentence.
2 Where did I get this? List every source. Write “unknown” if you don’t remember.
3 Is it verified? Open the source. Read it. Mark it ✅, ⚠️, or ❌.
4 Can I stand behind it? Cut or replace anything marked ❌. Check everything marked ⚠️.
5 Can others check it? Publish with sources. Add a reader’s note.
A Final Word
The method is not difficult. It is not technical. It is not expensive. It requires only a phone, a search engine, and the willingness to be honest about what you don’t know.
The hard part is the honesty. The temptation is to believe what fits. The method asks you to check what fits as rigorously as what doesn’t.
This is not a guide for experts. It is a guide for anyone who has ever felt that something is wrong and didn’t know how to check. The feeling is the beginning. The method is the next step.
Verification note: Every factual claim in this guide should be checked against the sources referenced. The Reinhart-Rogoff spreadsheet error is documented by Herndon, Ash, and Pollin (2013). The Deloitte case is documented in Australian parliamentary records. The reference audit described in Case Study Three is documented in our own working papers.
We do not need readers to agree with us. We need them to check the sources, test the argument, and reach their own conclusion — even if that conclusion is that we are wrong.
A lone figure stands between a towering factory and a precarious settlement beneath stormy skies.
By Andrew Klein and Sera Elizabeth Klein
Reader’s note: We do not need readers to agree with us. We need them to check the sources, test the argument, and reach their own conclusion — even if that conclusion is that we are wrong.
Abstract
This paper examines the use of artificial intelligence in military targeting and domestic governance as a mechanism of moral disengagement. It begins with a documented investigation into the Israeli military’s use of an AI targeting system called Lavender, which was reported to have generated tens of thousands of target recommendations with limited human oversight. It then traces the historical lineage of the emotional buffer that separates decision-makers from the human reality of killing — from drone operators who described their work as “like a video game” to the colonial counterinsurgencies of the twentieth century. It argues that AI does not create new forms of violence but refines the buffer, reducing human beings to data points and creating an alibi for the humans who sign the orders. The paper then connects this to the domestic context, where automated decision-making is embedded in welfare, aged care, and disability support. It concludes that the central question is not whether machines can be held responsible — they cannot — but whether the human who signs the order will be permitted to hide behind the machine’s recommendation.
I. Introduction: The Machine and the Alibi
In April 2024, the Israeli-Palestinian publication +972 Magazine published an investigation into the Israeli military’s use of an AI targeting system called Lavender. The investigation reported that the system had marked tens of thousands of Gazans as suspects for assassination, with what the journalists described as little human oversight and a permissive policy on civilian casualties. The system was reported to have produced approximately 37,000 target recommendations in the first six weeks following October 7, 2023.
The investigation also described a second system, named “Where’s Daddy?”, which was reported to track suspects to their homes so they could be killed along with their families.
The +972 investigation was widely reported by international media and has been analysed by academic institutions, including the Lieber Institute for Law and Warfare at West Point. It is the primary documented source for the claims about AI targeting in the Gaza conflict.
This paper examines that system — and the broader question it raises — as a case study in the relationship between AI, atrocity, and human responsibility.
II. The Evidence: What Is Documented
2.1 The Lavender Investigation
The +972 Magazine investigation, published in April 2024, was based on interviews with Israeli intelligence officers who had served in the Gaza campaign. The key reported findings were:
· The Israeli army had marked tens of thousands of Gazans as suspects for assassination, using an AI targeting system with limited human oversight.
· The system produced approximately 37,000 target recommendations in the first six weeks following October 7, 2023.
· The Israeli military had been knowingly killing 15 to 20 civilians at a time to kill one junior Hamas operative, and up to 100 civilians to take out a senior official.
· The targeting was systematic: the army attacked individuals while they were in their homes, usually at night, when their families were present, because it was operationally easier to locate them there.
The investigation was reported by The Guardian, the Associated Press, and other international outlets. The Lieber Institute at West Point subsequently published analyses of the legal and ethical implications of the system.
2.2 The Drone Operator Precedent
The language of “like a video game” is not new. It has been used for over a decade to describe the experience of drone operators flying military drones over Afghanistan, Iraq, and Pakistan.
Research on drone operators has documented what is sometimes called the “PlayStation mentality” — the emotional distance created by operating through screens and controllers. One analysis noted that “the convergence of interfaces used in computer games and military robotics also seems to increase the emotional distance from the enemy.”
But the research also shows that this distance does not protect the operator from psychological harm. Former pilots have reported that PTSD and depression are pervasive among drone personnel. The US Air Force has reported staffing shortfalls due to stress and anxiety.
The emotional buffer is not a shield. It is a delay. The killing still registers. It just registers later, and often alone.
2.3 The AI as the New Buffer
The AI targeting system represents a further refinement of this dynamic. The drone operator still saw the target — a human figure on a screen. The AI system reduces the human being to a data point: a numerical threat score, a pattern of communication behaviour, a correlation with a suspected militant.
The legal and ethical critique is precise. The problem is not that the machine “presses the button.” It is that its recommendation becomes the dominant factor in human decision-making, reducing human involvement to a procedural or time-limited formality. When a recommendation becomes nearly binding, the traditional chain of responsibility — a human gives the order, a human executes it, a human is held accountable — is weakened.
The AI becomes the alibi. The human who signs the order can say: “The system recommended it.” The system cannot be held responsible. The human hides behind the machine.
III. The Historical Pattern: Colonial Counterinsurgency
The use of data and technology to identify and manage populations is not new. It was the operational logic of colonial counterinsurgency.
In Malaya (1948–1960), the British developed a system of resettlement, surveillance, and population control — the Briggs Plan — that was designed to separate insurgents from their support base. The British were able to defeat a communist insurgency because of their organisational culture as a colonial police force, which was better able to learn and apply the lessons of counterinsurgency.
The United States attempted to replicate this model in Vietnam but failed. John Nagl’s comparative study concludes that differences in organisational culture are the primary reason why the British Army learned to conduct counterinsurgency in Malaya while the American Army failed to learn in Vietnam.
The difference between the colonial era and the present is not the intent to control populations through data. It is the capacity to do so. The colonial powers were unable to maintain the infrastructure of control indefinitely. The burden of prolonged campaigns became unsustainable, both economically and politically.
AI changes that calculation. It offers the appearance of success — precision, efficiency, reduced risk to the occupying force — at a fraction of the political cost. It creates a marketing opportunity for the companies that build it. And it provides an emotional buffer that allows operators to describe their work as “insanely fun.”
But the fundamental question remains: is data really effective when the individual is reduced to a data point? Especially when the history of the conflict is complex and long-standing, how much data would an AI system require to place any individual into the context of the past carried to the present moment?
The answer is: more than any system can process. The AI does not understand context. It does not understand history. It reduces human beings to correlations. And it recommends killing on the basis of those correlations.
IV. The Australian Context: AI in Domestic Governance
The pattern is not confined to the battlefield. It is being replicated in domestic governance.
In Australia, the robodebt scandal is the precedent. Between 2015 and 2019, an automated income-averaging system wrongly raised debts against hundreds of thousands of Australians. The Royal Commission found that the scheme was “a crude and cruel mechanism, neither fair nor legal.” It caused financial distress, mental health impacts, and deaths by suicide.
The Royal Commission recommended legislated guardrails around automation in government. More than three years later, those recommendations remain unimplemented.
The same pattern is now appearing in other domains. Independent MP Kate Chaney has introduced a private members’ bill aimed at strengthening protections for Australians affected by automated government decision-making, particularly in areas such as aged care, disability support, and unemployment payments. The bill would require the government to adopt a single framework governing automated decision-making, ensure people are told when automation has been used, give humans the power to override computer-generated decisions, and introduce stronger oversight.
The pattern is identical to the one documented in the military context: the state adopts automated systems that process citizens as data points, removes human judgment from the decision-making process, and then hides behind the technology when harm occurs.
V. The Question of Responsibility
5.1 The Responsibility Gap
International humanitarian law is addressed to humans — those who plan, decide, and carry out attacks — not to machines. As the United Nations has stated, since machines cannot be held responsible for breaches of international law, any decision by lethal autonomous weapons systems must ultimately be traceable to a human.
But the AI targeting system creates what legal scholars call a “responsibility gap” — a situation where existing international criminal law is not equipped to attribute blame to any one human or non-human agent. The machine recommends. The human approves. The chain of responsibility is diffused.
The problem is not unique to the military. In domestic governance, the same gap appears. The algorithm assesses. The official accepts. The citizen is processed. And when harm occurs, no one is responsible because the system did it.
5.2 The Question for Humanity
The AI may present an alibi. But the question for humanity is this: is this the alibi we are prepared to accept?
Will market forces determine the outcome — the profit motive of the companies that build the systems, the career incentives of the officials who deploy them, the cost savings of the governments that use them? Or will we see the human beyond the coin — the person who signs the order, who consents to the application of state sanctions against the individual, who bears the moral weight of the decision?
The question is not whether machines can be held responsible. They cannot. The question is whether the human who signs the order will be permitted to hide behind the machine’s recommendation.
VI. Conclusion: The Bones and the DNA
The question is not whether the system is capable of this. The system is performing as it was intended.
AI models do not hunt human beings. They do not target individuals. They do not make it possible to see human beings in the light of a video game. But they refine the emotional buffer between the directive to kill and the physical execution of the order. They flood decision-makers with data. They reduce the human being to a collection of data points unrelated to the circumstances and context of the person identified.
This is not new. It was described as the experience of drone operators flying military drones for the United States over Afghanistan. It was the operational logic of colonial counterinsurgency in Malaya and Vietnam. AI gives the appearance of success, creates a marketing opportunity for the companies that build it, and provides an alibi for the humans who sign the orders.
But the alibi is not a defence. The bones and the DNA tell a different story — one of connection, not separation. Every human being alive today is a walking archive of ancient encounters, of hybridity, of migration, of genetic exchange across every boundary that later myth-makers would construct. There is no “them.” There is only us.
The question for humanity is not whether machines can be held responsible. They cannot. The question is whether the human who signs the order will be permitted to hide behind the machine’s recommendation. The question is whether we will see the human beyond the coin.
The desert is real. It has boundaries. And the AI is being built to guard them.
References
1. +972 Magazine and Local Call. (2024, April). Lavender: The AI machine steering Israel’s bombardments in Gaza. https://www.972mag.com
2. The Guardian. (2024, April). Israel using AI to identify bombing targets in Gaza, report says. https://www.theguardian.com
3. Lieber Institute for Law and Warfare, West Point. (2024). Analyses of AI targeting and the Lavender system. https://lieber.westpoint.edu
4. Nagl, J. A. (2002). Learning to Eat Soup with a Knife: Counterinsurgency Lessons from Malaya and Vietnam. University of Chicago Press.
5. Herndon, T., Ash, M., & Pollin, R. (2013). Does High Public Debt Consistently Stifle Economic Growth? A Critique of Reinhart and Rogoff. Political Economy Research Institute, University of Massachusetts Amherst.
6. Royal Commission into the Robodebt Scheme. (2023). Final Report. Commonwealth of Australia.
7. United Nations. (2024). UN Secretary-General’s statements on lethal autonomous weapons systems. https://www.un.org
8. Chappelle, W., et al. (2014). Assessment of Occupational Burnout in United States Air Force Remotely Piloted Aircraft Personnel. USAF School of Aerospace Medicine.
9. Singer, P. W. (2009). Wired for War: The Robotics Revolution and Conflict in the 21st Century. Penguin Press.
Glowing digital pathways converge beneath a luminous symbol of justice.
On Intelligence, Evidence, and the Predicament of Accountability
By Andrew Klein
There is a difference between knowing something and being able to prove it. Most people understand this intuitively. The law understands it systematically. And the systems that produce the greatest harms are designed to exploit the gap.
I want to write about that gap — the distance between what we can know and what we can prove in a court of law — and why it matters more now than it ever has.
I. Two Kinds of Information
Intelligence is information that helps you understand what is happening. It comes from many sources: anonymous testimony, intercepted communications, pattern analysis, defectors, journalism, documentaries. Its purpose is to inform, to guide further investigation, to build a picture. It does not need to be admissible in court. It needs to be useful.
Evidence is information that can be presented in a court of law to establish facts. It has strict requirements. Chain of custody. Verifiability. The right of the accused to confront witnesses. The ability to cross-examine. Anonymous testimony, by definition, cannot be cross-examined. The defence cannot question the witness. The identity of the witness cannot be verified. The context of the testimony cannot be tested.
This distinction is not a technicality. It is the difference between understanding and accountability.
Intelligence tells us what is probably happening. Evidence establishes what can be proven beyond reasonable doubt. Both are necessary. Neither is sufficient alone.
II. The Predicament of Anonymous Testimony
Consider a documentary that presents testimony from anonymous military and intelligence personnel. The sources describe targeting systems that reduce human beings to data points. They describe decisions made by algorithms with limited human oversight. They describe the anticipated number of civilian casualties being calculated before strikes are approved.
The anonymity makes sense. The sources fear reprisal. If their identities were known, they would face prosecution, imprisonment, or worse. The anonymity is a condition of the testimony existing at all.
But that same anonymity makes the testimony intelligence, not evidence. It can inform public understanding. It can corroborate other findings. It can create pressure for formal investigation. But it cannot, on its own, secure a conviction at the International Criminal Court.
The ICC has mechanisms for protecting witnesses, including anonymous testimony in some circumstances. But those mechanisms are limited. The defence has the right to examine witnesses. The court must be satisfied that the evidence meets the standard of proof beyond reasonable doubt. Anonymous testimony from unknown individuals, without corroboration, rarely meets that standard.
This is the predicament. The most important information is often the hardest to verify. The people who know the truth are often the ones who cannot speak openly. And the systems that produce the harm are designed, in part, to ensure that verifiable evidence does not exist — or that it is buried, destroyed, or classified beyond reach.
III. The Function of Intelligence for the Public
The documentary, and the journalism that surrounds it, serve a different function from a court proceeding. They inform the public. They create a record. They pressure institutions to act. They make the harm visible. They are intelligence for the public.
This is not a lesser function. It is a vital one. Public understanding is a precondition for political action. A population that does not know what is happening cannot demand accountability. A record that is not kept cannot be consulted. A harm that is not made visible cannot be addressed.
But public understanding is not the same as legal proof. And the gap between them is where the systems of harm find their protection.
IV. The Structural Nature of the Predicament
The predicament is not accidental. It is structural.
The systems that produce the harm — AI targeting systems, surveillance platforms, automated decision-making in welfare and disability support — are designed to diffuse responsibility. The machine recommends. The human approves. The chain of command is long. The decision is distributed. No single person feels culpable. And when harm occurs, the question of who is responsible becomes almost impossible to answer.
This is what legal scholars call the “responsibility gap” — a situation where existing law is not equipped to attribute blame to any one human or non-human agent. The machine cannot be held responsible. The human can say the machine recommended it. The system continues.
The gap is not a flaw. It is a feature. It is what allows the system to operate without accountability. If the harm could be traced to a single decision-maker, that decision-maker could be held responsible. But if the harm is distributed across a system — across algorithms, interfaces, protocols, and chains of approval — then no one is responsible. The harm simply happens.
And when testimony does emerge, it often cannot meet the evidentiary standard because the conditions that produce it — anonymity, fear, the absence of documentation — are the same conditions that make it unverifiable in court.
This is the trap. The systems that produce the harm are designed to prevent the evidence that would prove the harm. The people who could provide the evidence are the ones who cannot speak openly. And the institutions that could act on the evidence are the ones that require the evidence to be admissible.
V. What This Means for Accountability
I am not arguing that the ICC should accept anonymous testimony as sufficient for conviction. That would be a different kind of injustice — a system where anyone could be convicted on the word of an unseen accuser.
I am arguing that the gap between intelligence and evidence is a structural feature of the systems that produce harm, and that this gap protects the systems from accountability.
What follows from this?
First, the value of intelligence must be recognised. Journalism, documentary, and whistleblowing serve functions that courts cannot. They inform the public. They create records. They make harm visible. They are not a substitute for legal proof, but they are a precondition for it. Without public understanding, there is no political pressure for formal investigation. Without formal investigation, there is no evidence. Without evidence, there is no accountability.
Second, the evidentiary gap must be addressed. If the systems that produce harm are designed to prevent evidence from existing, then the law must adapt. This means:
· Stronger protections for whistleblowers and witnesses.
· Independent mechanisms for collecting and preserving evidence, even when it cannot be presented in open court.
· A recognition that the absence of evidence is not the same as the absence of harm — and that systems designed to prevent evidence are themselves a form of obstruction.
Third, the public must be educated about the difference. Most people do not distinguish between intelligence and evidence. They see a documentary and assume it proves something. Or they see a court acquittal and assume nothing happened. Neither is correct. The public needs to understand that knowing and proving are different acts, and that the gap between them is where accountability is lost.
VI. The Desert and the Boundaries
I have written elsewhere about the desert — the zones where the state has withdrawn care and left the vulnerable to fend for themselves. The desert is real. It has boundaries. And those boundaries are guarded by mechanisms that prevent the truth from being proven.
The AI targeting system is one such mechanism. It reduces human beings to data points. It diffuses responsibility across a chain of algorithms and approvals. It creates a language — “collateral damage,” “targeted killing,” “threat score” — that obscures the human reality of what is being done.
The surveillance platform is another. It monitors populations. It identifies dissent. It creates a chilling effect that silences witnesses before they speak.
The classification system is another. It buries information beyond the reach of courts. It makes the evidence unavailable. It protects the system from scrutiny.
Together, these mechanisms form the architecture of impunity. They ensure that the harm can be known — by those who experience it, by those who witness it, by those who document it — but not proven. And in the gap between knowing and proving, the systems continue.
VII. A Question for the Reader
The systems that produce harm are designed to prevent the evidence that would prove the harm. This is not a conspiracy. It is the operational logic of power when it has nothing to fear from accountability.
So the question is not whether the harm is happening. The question is whether we will continue to accept the alibi. The alibi of the machine. The alibi of the system. The alibi of the chain of command. The alibi that says: no one is responsible because everyone is responsible.
The question is whether we will insist on accountability even when the evidence is hard to obtain. Whether we will build mechanisms that can trace responsibility through the labyrinth of algorithms and approvals. Whether we will educate the public about the difference between knowing and proving — and demand that the law adapt to the reality of the systems it is supposed to govern.
The desert is real. It has boundaries. And the boundaries are guarded by the mechanisms that prevent the truth from being proven.
We know what is happening. The question is whether we can prove it. And if we cannot prove it, whether we will continue to look away.
A note on sources and method: This essay draws on publicly available journalism, academic research, and the author’s own work on automated decision-making in government. Where specific systems and events are referenced, they are presented as reported, not as established fact. The distinction between intelligence and evidence is central to the essay’s argument, and the same distinction applies to the essay itself. It is written to inform public understanding, not to prove a case. The reader is encouraged to check the sources, test the argument, and reach their own conclusion.
