On Cheap Artificial Consciousness: Matthias Michel's Challenge to Theory-Checklist Assessment

One of the most common approaches to assessing AI consciousness works as follows. Take a well-established theory of consciousness. Read off its criteria. Check whether the system under assessment satisfies those criteria. If it does, attribute consciousness to a degree proportional to how many criteria it satisfies.

Matthias Michel, Assistant Professor in the Department of Linguistics and Philosophy at MIT, argues in a PhilArchive preprint (available at philarchive.org/rec/MICOCA-3) that this approach rests on a conceptual confusion. The problem it generates, which he calls the problem of cheap artificial consciousness, is not a genuine result of applying consciousness science to AI. It is an artifact of applying theories in a way that the theories themselves do not license.

The Problem of Cheap Consciousness

The phrase “cheap consciousness” refers to the unwanted consequence of the checklist approach. If you follow the criteria of Global Workspace Theory mechanically, many systems qualify as implementing a global workspace. If you apply Integrated Information Theory’s conditions without the full mathematical apparatus, many systems have positive Phi. If you read Higher-Order Thought theory as requiring only that a system represent its own states, plenty of current AI systems pass.

The result is that theory after theory, when applied as a checklist, yields a positive assessment for AI systems that researchers find intuitively implausible as candidates for consciousness. The theories were developed to explain human (and possibly animal) consciousness. Applied to AI, they seem to generate consciousness too cheaply.

This is the problem Michel addresses. His argument is that the cheap result does not arise from a gap in the theories. It arises from applying the theories incorrectly. Specifically, the checklist approach confuses the criteria that theories use to explain consciousness in systems we already know are conscious with criteria for detecting consciousness in systems whose status is unknown. These are different uses of the same theoretical framework, and conflating them produces spurious results.

The Explanatory versus the Diagnostic Use of Criteria

Michel draws a distinction that is simple to state but significant in its implications. Consciousness theories are built to explain why certain processes give rise to subjective experience in biological systems, starting from the assumption that those systems are conscious. The explanatory enterprise begins with the phenomenon (human consciousness) and constructs a theoretical account of it.

Applying that account to AI as a detection algorithm reverses the direction. The criteria developed to explain are now being asked to diagnose. But explanation and diagnosis are not interchangeable uses of the same theoretical content. A theory that successfully explains why the human brain’s global workspace dynamics give rise to conscious access does not thereby give you a valid test for whether a system you know nothing about is conscious.

This is Michel’s core argument: the cheap consciousness problem is a consequence of the explanatory/diagnostic conflation, not a sign that AI systems trivially satisfy conditions for consciousness. Once the conflation is resolved, the checklist approach is seen to be undermotivated for the diagnostic task it is being asked to perform.

How This Differs from Related Critiques

It is worth distinguishing Michel’s argument from the two closest existing critiques on the site. Florentin Koch’s calibration problem paper (arXiv:2603.27597) argues that the indicator programme is epistemically under-calibrated: no theory has independent validation, and there is no ground truth for AI phenomenality, making probabilistic attribution premature. Koch’s critique is about the epistemic status of the indicators. Michel’s critique is about the conceptual validity of the application method itself.

Tom McClelland’s epistemic limits argument holds that our evidence for what constitutes consciousness is too limited to tell whether AI systems have crossed the threshold, and that this limitation may be permanent. McClelland argues from the evidential gap; Michel argues from the conceptual confusion.

The Trends in Cognitive Sciences commentary on mimicry and internal variants argues that behavioral satisfaction of indicators may reflect superficial mimicry rather than genuine underlying states. That critique accepts the checklist framework but identifies a specific reliability failure. Michel’s critique targets the framework’s application logic, not its reliability within a correctly bounded application.

The three critiques are compatible and cumulative. A robust methodology for AI consciousness assessment would need to address all three.

What the Correct Application Would Require

Michel does not argue that consciousness science is irrelevant to AI. His argument is more targeted. The cheap consciousness problem dissolves once researchers are clear about what theoretical criteria are doing. Used for explanation, they describe the mechanisms by which known conscious systems achieve conscious access. Used for diagnosis, they require additional justification that has not yet been provided: a positive argument that satisfying the mechanistic criteria in a different substrate and via different processes generates consciousness of the same kind.

Providing that positive argument is non-trivial. It requires taking a position on whether consciousness is substrate-independent, whether architectural similarity at one level of description is sufficient for phenomenal equivalence, and whether the mechanisms theories identify are constitutive of consciousness or merely the mechanisms that happen to implement it in biological systems.

These are open philosophical questions. Michel’s contribution is to establish that they need to be answered before the checklist approach can be used legitimately for diagnosis, and that the cheap consciousness problem is the price of skipping them.

Implications for Multi-Theory Assessment

The Consciousness AI project’s architecture applies multiple frameworks simultaneously rather than selecting a single theory. The global workspace implementation draws on Baars and Dehaene. The oscillatory binding layer implements AKOrN dynamics (ICLR 2025), addressing the binding problem through phase synchronization. IIT Phi is measured via five ConsciousnessGate nodes with genuine causal dependencies. The Feinberg-Mallatt neuroevolutionary framework provides the overarching biological grounding.

This multi-theory approach does not automatically escape Michel’s critique. Satisfying the criteria of several theories simultaneously is still a checklist operation if the individual applications each confuse the explanatory and diagnostic uses. What would be required to address Michel’s concern is a positive argument for why the project’s mechanisms, implemented in this architecture and trained on these environments, might be constitutive of experience rather than merely implementing similar functional organization to biological systems that are conscious. That argument has not yet been made for any AI architecture, including this one. The honest position is that multi-theory architectural grounding is necessary but not sufficient to answer the cheap consciousness challenge.