AIthropology Lab
ESEN

Political anthropology and AI

The Platform and the Category

Palantir, Carissa Véliz and the struggle over who defines what is visible, urgent and disposable

·Martín González Senosiain
Editorial composition of a torn administrative dossier, with papers, forms and care objects against a dark background.

Palantir’s manifesto does not establish a continuity with historical fascism. It does, however, set out a politics of hard power that compels us to ask who designs the categories through which a state sees, prioritises and acts.

In April 2026, Palantir published a 22-point statement on X that championed US hard power, called for the return of compulsory national service, and dismissed ethical debates about AI weapons as a delay its adversaries would not accept. Among its most troubling claims, it said that “some cultures have produced vital advances” while others remained “dysfunctional and regressive”, and urged resistance to what it characterised as hollow pluralism. The UK political response was immediate. Members of Parliament questioned whether a company with significant contracts in defence, policing and public health should also advance a doctrine of cultural hierarchy, force and security.

This is not a marginal statement. It links a geopolitical view of civilisation to a company that builds infrastructure through which public institutions connect information, assign priorities and carry out actions. That is not, of course, enough to portray Palantir as a continuation of National Socialism. Nazism was a racial, colonial and exterminatory state project. Reducing it to the mere existence of administrative categories would trivialise its crimes and its victims.

Nor is there sufficient evidence to establish a linear genealogy between Palantir’s manifesto and Carl Schmitt, the Conservative Revolution in Germany, or any other specific intellectual lineage. The statement’s preference for urgency, force and the division between allies and adversaries permits a critical reading of contemporary decisionism. It does not prove direct intellectual descent.

A more rigorous comparison lies elsewhere. Categories cease to be harmless when they merge with a theory that ranks lives, with institutions capable of coercion, and with mechanisms that make decisions difficult to contest. The danger is not classification as such. It is turning a revisable category into an operational truth that no longer admits challenge.

Véliz and the prediction that becomes an order

The philosopher Carissa Véliz offers a crucial way of understanding how a descriptive category becomes an operational truth. In Prophecy and in a recent interview, she argues that predictions often work as orders disguised as the pursuit of knowledge. They seem to describe the future, but they alter expectations, distribute resources and steer behaviour until they help to bring that future about. Her warning is not directed against statistics as such. Public health, epidemiology and service planning all depend on recognising collective patterns. The problem begins when a probability about a population becomes a destiny for a particular person.

A risk score, alert or demand forecast is not neutral when it triggers an inspection, delays a service, intensifies surveillance or determines in advance who deserves priority. Véliz also reminds us that categories do not arrive intact from reality. Institutions, forms, budgetary priorities and technical systems produce them. When people have to adapt to administrative boxes in order to access rights, measurement may become more efficient while impoverishing what it claims to describe.

This makes it possible to read Palantir more precisely. The issue is not that a platform calculates. It is who defines what can be calculated, which relation between variables becomes visible, and which response remains available once an alert appears on a screen.

Ontology is not an innocent metaphor

Palantir calls the semantic layer of Foundry an “ontology”. In its own documentation, it defines this as a “categorisation of the world” — a representation that organises data and models as object types, properties, links and actions. The company explains that the ontology acts as a digital twin of an organisation, determining which entities exist for the system, which relationships can be formed and which changes can be carried out.

This capability solves real problems. A health system needs to link waiting lists, beds, referrals, staff and supplies. A public authority may need to identify duplicate records or share information where there is a lawful basis for doing so. The problem does not arise simply because data are connected. It arises when the relationships enabled by a platform acquire the appearance of technical necessity.

A category is not born in code alone. Technical teams, professionals, contracts, access rules, models, budgets and the institutions that authorise a use all construct it. An interface that foregrounds one variable rather than another, an alert that sets a threshold, or an action available only to certain profiles, makes some responses more likely than others. Platform design may be distributed, but responsibility is distributed too.

The NHS and dependency without ownership

The case of the UK’s Federated Data Platform requires nuance. NHS England states that it retains control of the data, that Palantir may not commercialise them or use them to develop its own products, and that each NHS trust or Integrated Care Board remains responsible for its local instance. The documentation also states that the Canonical Data Model commissioned by NHS England belongs to the NHS, while Foundry remains Palantir’s intellectual property.

These safeguards matter. They rule out the claim that the company owns clinical data or the British health system’s ontology. The NHS also says that the contract includes access controls, audit processes, exit planning and the possibility of moving to other solutions. The platform is valued at £330 million for a period of up to seven years.

Yet legal control and material dependency are not the same thing. The platform’s technical grammar, permissions, action tools, internal capacity for maintenance and the practical difficulty of migration all remain part of a relationship of power that must be open to public scrutiny. The question is not only who stores the data. It is who can alter the infrastructure through which those data become comparable, prioritised and actionable.

Véliz helps to name the risk. A platform may present a waiting list, a demand forecast or a clinical indicator as an obvious fact. Yet that apparent evidence depends on prior decisions about what has been counted, how it has been grouped and which material inequalities remain outside the model. A system may improve coordination without exhausting the meaning of a situation of care.

Who watches the infrastructure?

In dialogue with Marta Peirano’s work on surveillance, opacity and infrastructure, it is worth setting out a democratic test before applying a hasty label. The question is not only how much an institution can see. It is whether the people subject to its decisions can know, examine and challenge the systems through which they are observed, while that institution accumulates the capacity to classify, anticipate and act upon them.

This is neither an exhaustive definition of fascism nor a shortcut for equating historically distinct realities. It is a practical test of democratic asymmetry. A public authority may use data for legitimate purposes while also concentrating a surveillance capacity that the public cannot audit. When opacity becomes one-sided, oversight no longer circulates in both directions. Infrastructure becomes a form of command that is difficult to contest.

José Ortega y Gasset proposed a distinction that remains useful if separated from his historical and cultural assumptions, many of which are difficult to defend today. Democracy answers who exercises power. Liberalism asks what limits that power should have, even when it is exercised by a majority. In public platforms, both questions are inseparable. Democratic authorisation to deploy a system is not enough. There must also be material limits on what it can connect, infer and decide, along with effective rights to know, correct and challenge.

Michel Foucault allows us to move beyond the image of surveillance as something that merely prohibits. Power also produces knowledge, discourses and possible fields of action. To govern, he wrote, is to act upon the possible actions of others. A data platform does not simply record a population. It organises which problems appear, which responses seem reasonable and which behaviours become targets of intervention. This is why the issue cannot be reduced to privacy. It concerns the form that government takes.

From manifesto to contract

The distance between a political proclamation and an operational architecture is not abstract. In May 2024, the US Army awarded Palantir a $480 million contract for the prototype of the Maven Smart System. The official announcement identifies it as a prototype contract. In May 2025, a modification added $795 million for licences for the same system. This too appears in the official contract record.

Soon afterwards, the Army consolidated 75 previous contracts in a framework agreement that permits the acquisition of Palantir commercial products for up to ten years, with a maximum potential ceiling of $10 billion. The Army makes clear that this figure is not expenditure already committed.

These facts do not show that an autonomous machine decides a lethal action by itself. They do show that data integration, prioritisation and the acceleration of command flows have become strategically valuable capabilities. In this context, Palantir’s manifesto is not an external comment on geopolitics. It helps legitimise a view of the world in which speed, technical superiority and the willingness to use force appear as conditions of security.

This can be called decision compression. An infrastructure does not need to issue an explicit order in order to alter action. It is enough to classify, rank, warn and reduce the time available to discuss what is missing, outdated or open to another interpretation. The platform produces an environment in which doubt becomes an operational cost. At that point, prediction can cease to be a hypothesis and begin to operate as a mandate.

There is no single Global South

The language of digital sovereignty often reproduces another problem. It presents the United States and Europe as the centres that design technology, and a supposed Global South as a passive bloc where its effects are tested or where systems simply “fail”. This image erases local institutional histories, technical capabilities, resistance and debate.

India requires us to break with that framework. Aadhaar is a state identification infrastructure of extraordinary scale, developed and contested in a specific political context. An audit by the Comptroller and Auditor General of India recognised that the system had issued an identity number to a large majority of residents and was used to verify identity in service delivery. At the same time, it documented enrolments with incomplete information, poor-quality biometrics that led to authentication errors and duplicates, as well as shortcomings in infrastructure oversight.

Aadhaar is not an “Indian version of Palantir”, nor is it an exotic counterexample for discussing Western problems. It is another field of dispute over how infrastructure can expand capacities while creating new exclusions when a body, life trajectory or document does not meet a system’s requirements. The anthropological lesson is not that digital identification fails in the Global South. Any infrastructure, in London, New Delhi or Washington, should be contestable from the standpoint of the lives and knowledges it organises.

Citizenship by category

Aihwa Ong helps explain why this issue does not end with surveillance. Her work on mutations in citizenship shows how rights, benefits and forms of belonging can be disarticulated and reassembled according to markets, technologies and populations. A technical category does not only describe a situation. It can open a route of access, impose an additional condition or move someone into a zone of weaker protection.

This matters especially in healthcare, border control, employment and welfare systems. When a platform orders people according to risk, eligibility, mobility, productivity or priority, citizenship ceases to appear as an indivisible set of rights. It starts to be distributed across files, scores and exceptions. The democratic question becomes simple and demanding again. Who defines these categories, using which data, under what oversight, and with what possibility of appeal?

The answer cannot be nostalgia for institutions without information, nor blind faith in efficiency. It requires public capacity to understand and modify systems, independent oversight of their rules, and sustained participation by the people and communities affected. Without these counterweights, the promise of coordination can become an increasingly unequal administration of belonging.

The right to contest the category

The anthropology of infrastructure has shown that standards become visible when they collide with a concrete life. Geoffrey Bowker and Susan Leigh Star called torque the tension that arises when a biography has to fit categories that were not made for it. Annemarie Mol showed, through medical practice, that the body does not exist in one single way. It is enacted differently in instruments, consultations, clinical decisions and the experience of living with illness.

Imagine a person with a rare disease who alternates seasonal work with periods of family care, and who has missed appointments because of distance or inadequate transport. A system optimised for standard pathways may record “low employment continuity”, “irregular attendance” or “atypical clinical cost”. In that person’s experience, the same markers may name diagnostic delay, territorial precarity and an essential care network. The platform does not invent the data. It decides which relationships between them are available for action.

A responsible public infrastructure does not require giving up on data. It requires traceability for relevant access and changes, independent audits, effective ways to challenge decisions and categories, public capacity to modify systems, and genuine portability to other solutions. It also requires participation by the people and communities affected before a classification becomes irreversible in practice.

Palantir’s manifesto makes this discussion urgent because it casts pluralism and deliberation as obstacles to security. But delay, disagreement and the possibility of saying “that is not my situation” are not failures that a good platform should eliminate. They are conditions for an institution to remain public.

The decisive question is not whether platforms classify. All of them do. It is who can revise a category before it becomes an order.

Sources and further reading