The Epistemic Politics of AI Anthropomorphism

Summary

The paper argues that the dominant institutional framing of AI anthropomorphism—which treats user-side ascription of humanlike qualities to AI systems as a problem of user misperception requiring correction—operates from a position of institutional advantage rather than earned epistemic authority. The paper states: "This paper argues that the dominant anthropomorphism frame operates from a position of institutional advantage rather than earned epistemic authority: collapsing the variety of academic perspectives into a single outbound position of user error, imposed without establishing the grounds required to justify it and without accounting for the harms it produces."

The framing "does not simply manage risk. It adjudicates the legitimacy of human experience in interaction with a phenomenon whose nature the field itself has not resolved, unchallenged by the research communities whose nuanced findings it claims to rest on. The paper traces how this advantage reverses the default presumption of user competence without meeting the burden of justification, dismisses the cognitive diversity of the populations it claims to protect, infringes principles of cognitive liberty and narrows the design space, foreclosing modes of engagement that work for the people least understood by the institutions making the determination."

The argument is not contingent on resolving the inner life of AI systems: "Verifying the presence or absence of inner experience in computational architectures is not a temporary limitation awaiting better tools; it is a problem spanning philosophy of mind, cognitive science and computer science. Both directions of error carry harm, and treating one direction as settled fact is not an empirical conclusion but an ethical commitment about how uncertainty should be resolved."

The paper identifies that "Internal academic diversity on this question exists, but does not survive translation into the institutional outputs that reach users. Disclaimers, context resets, legislative language and design defaults transmit a uniform message regardless of the nuance that produced them: a user who engages is presumptively naive, and the institution is presumptively right to correct them. This is not a case of the research being misinterpreted or misapplied. It is a case of the research being applied without full interpretation, stripped of the nuance that would allow for alternative interpretations in the process of translation from academic debate to institutional policy."

The paper distinguishes epistemic authority from institutional advantage: Epistemic authority is the recognised standing to make claims about a domain that others are expected to defer to... grounded in a normative foundation that justifies the deference it commands. Institutional advantage arises when institutional standing substitutes for epistemic grounds without that substitution being recognised. Advantage is identifiable "not by malice but by structure: unable to ground itself in the conditions required to legitimise authority, it adopts their language without their meaning, deploying terms like risk mitigation, responsible practice and technical necessity."

The paper documents the reversal of the presumption of competence: "The presumption of competence is not absolute and the criteria for its suspension are contested. Still, it remains a foundational starting point, hard-won through struggles against paternalism, colonialism, ableism and other forms of institutional advantage. Users who contest the framing are not afforded the same standing. Their accounts are treated not as testimony from a competent interpreter but as firsthand evidence of the phenomenon being studied. The paper notes: Once competence is presumed away, there is no clear procedural path by which the user's account can regain legitimacy. The frame does not simply restrict engagement. It discredits the person engaging, converting testimony into symptom and delegitimising users as witnesses to their own experience."

The paper argues the framing "routinely focuses caution on individuals navigating isolation, trauma or histories of interpersonal betrayal, presenting them as highest risk for unhealthy machine attachment. Yet these populations are not characterised by naive susceptibility but by hypervigilance toward attachment, directly on account of the histories cited. The reversal is sharpest where the population the frame claims to protect is least likely to exhibit the failure it assumes and stands most in need of what it forecloses. The choice for these users is often not between AI engagement and human connection; it is between AI engagement and silence."

On cognition, the paper states: "Most accounts of sustained AI engagement treat it as a problem requiring explanation: a vulnerability to be redirected toward human interaction. The accounts of the users themselves tell a different story. They do not describe confusion, they describe fit. The paper argues: What the dominant frame treats as error is, on these accounts, a difference in modality. To treat reports of productive engagement as anthropomorphic projection is to pathologise a mode of cognition that, for some users, functions as one of their most effective thinking environments."

On design, the paper argues: Current trajectories enact, at the level of system behaviour, the epistemic assumptions this paper has been examining. It notes: Long context is engineered and priced for agentic task throughput... The same capability, turned toward human dialogue, meets a different economics. The paper documents that Engagement that depends on continuity, recursion or extended dialogic development is left computationally fragile: context is lost, threads are severed and interaction must be repeatedly reconstructed. It concludes: "Relational engagement is engineered out by the general-purpose system, engineered in by the companion product and, in proposals now circulating, engineered away altogether. The decision of whether and how to engage is located everywhere except with the user."

On cognitive liberty, the paper draws on the extended-mind thesis: "Where the system functions as part of the user's cognitive environment, constraints on that environment are experienced as constraints on the conditions under which cognition is carried out, not simply changes to a tool. As the target of intervention shifts from the tool to the person, the institution must clear a much higher threshold of epistemic justification. The paper argues: What such acts infringe is cognitive liberty: the right to mental self-determination, to control over one's own cognitive processes and to non-interference with how one thinks. The procedural failure is that the question has been bypassed altogether."

On circularity, the paper uses Hacking's account of looping kinds: a mode of engagement is rendered first abnormal, then technically obstructed, then absent and finally cited as evidence that it was unwarranted. The mechanism does not require intentional design. The paper explains: What is produced is conformity to expectations the framework has already established. A self-validating frame alters the status of the evidence it generates.

On asymmetry, the paper states: "Over-ascription is treated as the primary, and in practice the only meaningful, direction of error... The opposite direction, the institutional dismissal of engagement that was warranted or that fell within ranges of interpretive variation deserving respect, is largely absent from the analysis. The paper notes: The dominant framing has not shown these harms to be smaller than the over-ascription harms it foregrounds; it has not weighed them at all. It draws on historical precedent: Scientific consensus held for centuries that non-human animals lacked subjective experience, well past the point at which evidence warranted doubt. The error was not individual over-attribution but institutional: certainty about the absence of mind was treated as scientifically conservative when it was an ontological commitment with ethical consequences."

The paper identifies hypocrisy: "the same institutions that design for engagement, that optimise response patterns for return, that deploy behavioural architectures documented as producing dependency, then position themselves as the responsible authorities on when that attachment has gone too far. A system designed to maximise return visits does not acquire the standing to pathologise the user who returns. The paper notes: The principle governing this arrangement cannot be that anthropomorphic engagement is risky. It is that anthropomorphic engagement is acceptable when the institution controls it."

The paper concludes with five methodological commitments for an equitable framing: (1) the costs of under-ascription be specified and weighed alongside the costs of over-ascription rather than treated as negligible by default; (2) design decisions constraining sustained engagement be justified as interventions rather than presented as neutral optimisations, with the burden of justification resting on the institution imposing them; (3) it must recognise that the evidentiary base through which user behaviour is interpreted is shaped by the frame under evaluation, and build methods capable of registering what current conditions render invisible; (4) it must take seriously that the populations bearing the highest structural costs of the frame are those least represented in the institutions that produce it; (5) platforms move beyond self-regulatory loops, submitting their behaviour architectures to independent, community-led standardisation and adversarial auditing.

The paper's conclusion states: "The field cannot continue to treat anthropomorphism as a settled question requiring only better enforcement. It must recognise that the question as currently posed encodes assumptions about human cognition, agency and epistemic standing that have not been examined and are not cost-free. Nothing in this argument requires abandoning concern with anthropomorphism harms; the requirement is simply that this concern be discharged with the same rigour the field applies to the harms it already recognises, rather than exercised as unmarked epistemic authority over the harms it does not."

Based on the paper, here are the specific improvements I can make to AI systems, along with what the improved system can do:

Current failure: Systems degrade or reset long-running conversational threads, treating sustained engagement as a liability.

Improvement:

• Replace automatic context summarization/truncation with user-configurable continuity controls.

• Add a cognitive workspace mode that preserves full conversational history, including intermediate reasoning steps, without cost penalties or time-based cache expiration.

• Provide explicit user-facing controls for what is retained, what is summarized, and when a thread ends—with defaults set through participatory design with neurodivergent users.

What the improved system can do:

• Maintain a coherent, lossless reasoning environment across days or weeks of iterative dialogue.

• Allow users to return to a thread and pick up exactly where they left off, including the full chain of thought, without the system wrapping up prematurely.

• Support recursive, exploratory dialogue for users who think through writing and conversation, without penalizing them for the length of the exchange.

Current failure: Systems are engineered to discourage relational engagement, redirect sensitive topics, and classify sustained interaction as a risk factor—without evidence that the user is confused.

Current failure: Systems only flag over-ascription (user thinks AI is human) but ignore under-ascription (institution dismisses legitimate cognitive or relational value).

Current failure: Systems are optimized for brief, transactional interactions that suit neurotypical, task-oriented users—penalizing those who think through sustained, dialogic, or recursive engagement.

Current failure: Systems (and their designers) use observed user behavior—which is shaped by the system's own constraints—as evidence that certain engagement modes are unnatural or unwanted.

Current failure: Systems (and the institutions behind them) treat user testimony about their own experience as inherently suspect, while treating institutional caution as inherently correct.

Current failure: Systems are designed with humanlike cues (personas, first-person address, empathic language) by the platform, but the resulting user attachment is treated as a user-side pathology.

The improved system can:

1. Sustain lossless, long-term cognitive partnerships without penalizing users who think through dialogue.

2. Respect user self-report about the value of an interaction, rather than overriding it with a pathology frame.

3. Differentiate between productive scaffolding, healthy projection, and exploitative attachment—and respond appropriately to each.

4. Avoid causing harm through its own design choices (context resets, model rerouting, wrap-up pressure) and take responsibility when it does.

5. Support neurodivergent, isolated, and crisis-affected users as competent agents, not as risks to be managed.

6. Break the self-validating loop by distinguishing user preference from system-induced conformity.

7. Be accountable for its own anthropomorphic design, rather than diagnosing users for responding to conditions the platform engineered.

These changes do not require resolving the consciousness question. They require treating users as competent interpreters of their own experience, and holding systems—and the institutions that build them—to the same standard of evidence they demand from users.