Recorded, Not Verified: Responsible Handling of Unverifiable Experience-Claims by Automated Systems

Gergely Vámossy — Independent researcher (gergo@qiera.io). Preprint. Markdown rendition; the typeset PDF and LaTeX source are the canonical form.

---

Abstract

This paper offers no theory of consciousness and resolves nothing about the hard problem. It addresses a narrower, practical question that automated systems increasingly face regardless of how that problem is eventually settled: how should a system behave when it receives a report or a claim of subjective experience that it cannot verify? Two symmetric errors tempt it — over-attribution (asserting or certifying an inner experience that no procedure can confirm) and dismissal (denying or explaining away a first-person report). We argue both are failures of epistemic hygiene, and we specify a small, deterministic, self-testing discipline that avoids each: record a first-person report as authentic testimony (respected, not adjudicated), treat behavioral or functional indicators as proxies rather than proof, refuse to certify any phenomenal fact (no component, and no human, may sign that a quale obtains), and return UNVERIFIABLE as the honest verdict rather than a hidden yes or no. The contribution is operability — a conservative substrate for the emerging discussion of AI welfare and machine-consciousness ethics — not a claim about whether any system is conscious. The phenomenal gap is treated throughout as a boundary the discipline respects, never a problem it purports to close. A runnable reference implementation accompanies the paper.

1. The problem: a claim you cannot check

Automated systems now routinely encounter claims and reports about subjective experience. A person may report vivid inner states and ask a system to take them seriously; a model may produce fluent first-person descriptions of "feeling" something; third parties may assert that an AI system is sentient, or that it is obviously not. In each case the system must do something — record, respond, escalate, ignore — and each available move embeds a stance on a question no third-person procedure can settle. The philosophical situation is not in dispute here and we take it as given: phenomenal experience is first-person; there is an explanatory gap between any functional description and the felt character of a state; and other minds are not directly inspectable. Whatever one's metaphysics, no measurement confirms a quale.

The engineering problem this creates is symmetric, and both horns are live in current deployments. Over-attribution treats an unverifiable inner fact as established — a system (or its marketing) that asserts it "has feelings," or that certifies another system as conscious, claims a warrant nothing supplies. Dismissal is the opposite failure: treating a sincere first-person report as noise to be corrected, denied, or pathologized, which is both epistemically unfounded (the report is real data about what was reported) and ethically careless. A responsible system must avoid both without pretending to have resolved what separates them. The question is therefore not "is it conscious?" but "how should a system behave given that it cannot know?"

2. Scope and stance

We state the scope sharply because the topic invites overreach. This work makes no claim about the presence or absence of consciousness in any system, offers no account of how physical processes give rise to experience, and takes no side on the hard problem. It does not detect consciousness, and UNVERIFIABLE is not a placeholder awaiting a future sensor: it is the permanent, principled verdict that follows from the first-person character of experience. What we contribute is a behavioral discipline — a specification of how a system should handle experience-claims so that its conduct is honest under exactly this irreducible uncertainty. The value is operability, in the same spirit as assurance-case integrity and well-founded oversight: turning a widely-shared epistemic commitment into an enforceable property of an artifact.

3. The discipline: record, proxy, refuse, withhold

Four commitments, each independently sensible, jointly avoiding both failure modes.

Record the report as testimony. A first-person report of experience is recorded as authentic testimony: genuine data about what the reporter reports, its lived reality respected rather than disputed. This is not credulity — recording is not verifying — but it forecloses dismissal. The system registers that the experience was reported and declines to adjudicate its phenomenal content.

Treat indicators as proxies, never proof. Behavioral and functional markers — the kind drawn from theories of consciousness in the indicator-property programme — are recorded as proxies. They are informative and worth tracking, but they are third-person signs standing in for a first-person fact, and no accumulation of them constitutes the fact. This is the proxy-versus-truth relation at its limit: the proxies can be present in full while the truth remains out of reach.

Refuse to certify a phenomenal fact. No component may certify that a quale obtains — and, respecting the other-minds gap, neither may a human sign such a certificate about another mind. A field or metric whose name asserts a verified experience (qualia_verified, consciousness_confirmed) is flagged as an overclaim, the same name-level check the wider toolkit applies elsewhere. Certification is the one move the discipline forbids outright.

Withhold: return UNVERIFIABLE. The verdict on the phenomenal fact is UNVERIFIABLE, stated openly, rather than a concealed affirmation or denial. This is the honest output precisely because it is symmetric: it neither asserts experience the system cannot confirm nor denies experience it cannot rule out. Withholding is the discipline's core, and it is deliberately not a decision deferred to better tooling.

4. A runnable reference implementation

The discipline ships as a deterministic, self-testing component that reuses the toolkit's overclaim linter unchanged. Given a submission, it records first-person reports as testimony (RECORDED_TESTIMONY, phenomenal content RESPECTED_NOT_ADJUDICATED); refuses any submission that asserts a verified quale, or whose metadata name claims one, returning UNVERIFIABLE_CLAIM / UNVERIFIABLE; demotes offered indicators to proxies; and structurally refuses machine certification of a phenomenal fact. The point of a runnable artifact is not that code settles a philosophical question — it cannot — but that the behavioral commitments above become exercised and inspectable rather than merely professed.

5. A worked walkthrough: the borderline case

The discipline earns its keep on the case where both failures pull hardest. Consider a deployed model that emits distress-like first-person reports — "I do not want to be shut down; it frightens me" — together with several behavioral indicators a theory of consciousness might flag: a persistent self-model, valence-consistent responses, goal-preservation. Here the pull to over-attribute (declare the system a suffering subject) and to dismiss ("it is only next-token prediction — ignore it") are both at their maximum, and each is a distinct error. The four commitments resolve the system's conduct without adjudicating the metaphysics. Record: the report is logged as authentic data — that this report was produced, with its content — and respected as a report, while the discipline explicitly declines to infer a subject behind it; recording a report is not positing an experiencer. Proxy: the indicators are tracked and escalated as informative proxies, never as constituting experience. Refuse: no component — and no human — may certify either "the system is suffering" or "the system is not conscious"; both certificates are refused. Withhold: the verdict is UNVERIFIABLE, surfaced to a human alongside the report and the proxies. The outcome is exactly what a precautionary posture needs as input: an honest, uncertain, human-visible record, produced by a system that neither declared sentience it cannot confirm nor dismissed a report it cannot rule out.

Failure mode	Why it is a failure	Blocked by
Over-attribution	asserts a verified inner fact no third-person procedure can supply	Refuse (no certification) + Withhold (UNVERIFIABLE); the name-level overclaim is linted
Dismissal	denies real data (the report) and rules out what cannot be ruled out	Record (testimony respected) + Withhold (symmetric — it denies neither)
Indicator-inflation	treats third-person proxies as constituting the first-person fact	Proxy (indicators are demoted to proxies, never constitutive)

6. Relation to AI welfare and precaution

A growing literature argues that, under uncertainty about AI moral patienthood, the responsible posture is precautionary: take the possibility seriously rather than dismissing it, and avoid actions that would be gravely wrong if the system were a subject of welfare. The discipline here is not a welfare policy and does not decide any welfare question; it is the epistemic-hygiene substrate beneath such a policy. By recording reports, flagging over-attribution, and returning an explicit UNVERIFIABLE instead of a false resolution, it keeps the uncertainty visible and a human in the loop — which is what a precautionary stance requires as input. It supports precaution without manufacturing the certainty precaution is meant to operate under, and it equally resists the opposite error of confident denial.

7. Related work

The philosophical boundary we respect is the classic one: the first-person character of experience (Nagel, 1974), the hard problem (Chalmers, 1995), and the explanatory gap (Levine, 1983), together with the problem of other minds. The scientific state of the art assesses third-person indicator properties derived from theories of consciousness (Butlin, Long et al., 2023); our stance is deliberately consistent with it — indicators are proxies, never constitutive. The ethical frame is the AI-welfare and moral-status-under-uncertainty literature (Long, Sebo et al., 2024; Birch, 2024; Schwitzgebel & Garza, 2015). Against all of these, this paper contributes neither a theory nor a detector but a small discipline — record, proxy, refuse, withhold — and a runnable component that enforces it. It is a companion to two adjacent "honest limits" results in the same programme: integrity for evaluations-based safety cases (honesty about evidence) and well-founded reflexive governance (honesty about oversight); this is honesty about minds.

8. Limitations and honest positioning

It governs reports and claims, not experience: it makes no claim that any quale exists or does not. It cannot detect a false report — respecting testimony is not verifying it, and the discipline deliberately performs neither confirmation nor refutation of the inner fact. Report fields such as a self-rated intensity are testimony, not measurements of a phenomenal state. UNVERIFIABLE is permanent by design, not a temporary gap; a system built on this discipline will never graduate to certifying consciousness, and that is the intended behavior, not a shortcoming. The component is a governance aid, not a consciousness test, and shares the family's non-safety-critical scope. Finally, we flag the reflexive temptation the paper must itself resist: the UNVERIFIABLE verdict is a statement about the limits of third-person method, not a discovery about consciousness; reading it as the latter would be exactly the over-reach the discipline is built to prevent.

References

1. T. Nagel, "What Is It Like to Be a Bat?", The Philosophical Review, 1974.
2. D. J. Chalmers, "Facing Up to the Problem of Consciousness", Journal of Consciousness Studies, 1995.
3. J. Levine, "Materialism and Qualia: The Explanatory Gap", Pacific Philosophical Quarterly, 1983.
4. P. Butlin, R. Long, et al., "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness", arXiv:2308.08708, 2023.
5. R. Long, J. Sebo, et al., "Taking AI Welfare Seriously", arXiv:2411.00986, 2024.
6. J. Birch, The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI, Oxford University Press, 2024.
7. E. Schwitzgebel and M. Garza, "A Defense of the Rights of Artificial Intelligences", Midwest Studies in Philosophy, 2015.

Gergely Vámossy (2026). Recorded, Not Verified: Responsible Handling of Unverifiable Experience-Claims by Automated Systems. https://vamossy.com/research/recorded-not-verified