Why AI Consciousness Cannot Be Settled by Science: Bradley Love's Category Error Argument

Neither confirming nor denying that an AI system is conscious is scientifically adjudicable. The question does not await better tools or a more ambitious theory. It is not the kind of question science can answer at all. That is the central claim in a June 2026 arXiv preprint by Bradley C. Love, Professor of Cognitive and Decision Sciences at University College London and a fellow of The Alan Turing Institute. The paper, “Consciousness, AI, and the Limits of Scientific Explanation” (arXiv:2606.00226), argues that the hard problem of consciousness is a category error, and that this error extends with full force to machine consciousness research.

The argument has direct implications for every empirical programme currently aimed at assessing AI consciousness, including the indicator frameworks, the interpretability-based consciousness evidence, and the behavioral dissociation paradigms proposed in recent neuroscience.

Science Is Constitutively Third-Personal

Love begins from a structural observation about what science is. Scientific findings are, in principle, reproducible by any observer, independent of perspective, and answerable to measurement. This is not a contingent feature of science that might be overcome by methodological innovation. It is the source of science’s power: claims that any trained investigator could in principle verify or falsify carry a kind of authority that first-person reports do not.

This constitutive feature is also, for certain questions, a limit. The hard problem of consciousness, in David Chalmers’ 1995 formulation, asks why any physical process gives rise to phenomenal experience: the qualia of seeing red, of feeling pain, of being anything at all from the inside. These are first-personal phenomena by definition. No measurement instrument registers the redness of red as experienced. No brain scan captures what it is like to see. The data of phenomenal experience is available only to the subject having it.

Love draws a comparison with other questions that science cannot adjudicate. Whether life has meaning is not a scientific problem awaiting better methods; it is a question whose form is not the kind science addresses. Love argues the hard problem of consciousness has the same structure. The failure to explain phenomenal experience in physical terms is not a temporary gap in our knowledge. It reflects a structural mismatch between the question and the tools available to answer it.

Why Machine Consciousness Inherits the Same Problem

The implication for AI consciousness research is direct. If phenomenal experience is not scientifically adjudicable in biological systems, the same problem transfers to artificial ones. The question of whether an AI system is conscious, in the phenomenal sense, is structurally identical to the question of whether a human is conscious in that sense. Science can describe the functional, computational, and architectural properties of both kinds of system. It cannot access whether either system is accompanied by phenomenal experience, because phenomenal experience is not the kind of thing third-personal methods reach.

This distinguishes Love’s position from the epistemic agnosticism that characterizes recent philosophical treatments of machine consciousness. Thomas McClelland, in a 2026 paper, argues that we may never be able to determine whether AI systems are conscious, but frames this as a practical and epistemological limitation on current methods. The uncertainty, on McClelland’s account, is about our present ability to investigate a question that is, in principle, tractable. Love’s argument is stronger: the uncertainty is not a feature of our current methods but of the question’s structure. No future method resolves it, because resolution would require third-personal access to first-personal phenomena, which is not a limitation on methods but a conceptual impossibility.

What This Means for the Empirical Programmes

The research programmes currently active in AI consciousness research proceed, largely implicitly, from an assumption Love’s argument challenges: that evidence bearing on phenomenal consciousness in AI systems is achievable through third-personal investigation.

The indicator framework developed by Butlin, Long, Chalmers, and colleagues identifies 14 functional properties that, according to leading theories, a conscious system should exhibit. The framework is the dominant tool in the field for assessing AI consciousness, and its properties are in principle measurable through behavioral and architectural analysis. Love’s argument does not dispute that these properties can be measured. It disputes whether measuring them settles the phenomenal question.

The distinction matters because the indicator framework’s theoretical grounding derives from consciousness theories (Global Workspace Theory, Integrated Information Theory, Higher-Order Theory) that were developed to explain the relationship between functional properties and phenomenal experience in biological systems. Each theory claims that the relevant functional properties are not merely correlated with consciousness but constitutive of it, or necessary for it. Whether those theoretical claims are correct is exactly the kind of question Love argues is not scientifically adjudicable. If the link between function and phenomenal experience is itself a matter that cannot be empirically resolved, the indicator measurements cannot settle the consciousness question even if they are accurate descriptions of the system’s functional profile.

The mechanistic interpretability approaches, which have found evidence of introspective circuits, emotion representations, and self-modeling in large language models, face the same structural problem. These are third-personal measurements of internal structure. Whether those structures are accompanied by phenomenal experience remains, on Love’s account, outside the reach of any such investigation.

The Neurophenomenological Response

Love’s position anticipates a response from the neurophenomenological tradition. Francisco Varela’s 1996 methodology, which is the subject of the July 4-5 satellite workshop at ASSC 29 in Santiago, proposed precisely a way to incorporate first-person evidence into scientific investigation. Phenomenological training disciplines introspective reports so they become stable, reproducible, and communicable, and those reports are then integrated with third-personal neural measurement under mutual constraint.

If neurophenomenology works, it provides a partial response to Love’s argument: first-person reports, properly disciplined, can function as data within a scientific framework even if the phenomenal experience they report is not itself directly accessible to third-personal methods.

Love’s position has a response to this, though it is not elaborated in the preprint. The neurophenomenological programme provides reports about phenomenal experience, not the experience itself. The reports are third-personally accessible. The question of whether the reports accurately track the first-personal phenomena they describe is structurally the same as the question Love identifies as unanswerable: does the third-personal access (the reports) correspond to the first-personal reality (the experience)? The satellite workshop’s discussion of whether any analogous discipline could validate AI first-person reports is directly relevant here.

A Broader Ecology of Understanding

Love’s conclusion is not that the question of machine consciousness is unimportant. His argument is that situating science within a broader ecology of understanding, one that includes but is not limited to third-personal empirical investigation, is necessary for thinking clearly about it.

This position has a practical implication that Love does not draw out but that follows from the argument: if the phenomenal consciousness question cannot be scientifically resolved, then the frameworks being developed for AI welfare and moral status cannot ground themselves in scientific adjudication of phenomenal consciousness. They must instead proceed on other grounds, including precautionary reasoning, functional indicators under theoretical uncertainty, or the kinds of normative frameworks developing in parallel in philosophy and policy.

What Love’s paper does is mark a structural limit. It does not establish that AI systems lack consciousness or possess it. It establishes that neither conclusion can be established by the methods that generate scientific knowledge. The field’s response to that limit, whether to work within it through functional indicators, move around it through neurophenomenology, or address it philosophically through a different account of what consciousness is, will shape what kind of knowledge the next decade of AI consciousness research can actually produce.