Stop Asking If AI Is Conscious, Start Asking Why People Think It Is: Comsa's DeepMind Paper on Tractable Questions
Iulia-Maria Comsa’s May 2026 arXiv preprint, AI and Consciousness: Shifting Focus Towards Tractable Questions (arXiv:2605.06965, May 7, 2026), is one of the more useful recent interventions in the AI consciousness debate, and its institutional positioning is a substantial part of what makes it useful. Comsa is a research scientist at Google DeepMind. The paper is not framed as a contribution to the philosophy of mind or to consciousness science directly. It is framed as an argument about research priorities, written by someone whose daily work involves the AI systems that are at the center of the conversation, addressed to researchers and decision-makers who have limited tolerance for questions that cannot be operationalized.
The argument is structurally simple and rhetorically pointed. The question of whether current or near-future AI systems are conscious is, on the available evidence, intractable. There is no consensus theory of consciousness that would allow the question to be settled. There is no measurement procedure whose result would compel agreement. The question is therefore unsuited to be the primary research target of an applied AI consciousness programme. The question of why users perceive AI systems as conscious, by contrast, is empirically tractable. It produces measurable data, supports hypothesis testing, and has direct consequences for product design, ethical norms, and public policy. Comsa argues that the perceived-consciousness question is therefore the question the AI industry and the adjacent research communities should be organizing around.
The Intractability Argument
The intractability claim is the load-bearing part of the paper. Comsa does not argue that consciousness is in principle unknowable or that the underlying metaphysical question is meaningless. She argues that the question is not currently in a position to be answered using the methodological tools the AI research community has available. To answer whether an AI system is conscious requires a theory of consciousness that yields predictions about which systems should have it. The available theories do not converge on those predictions. Integrated Information Theory, Global Workspace Theory, Higher-Order Thought theories, and predictive processing accounts each yield different verdicts about which AI architectures should be considered candidates. There is no procedure for adjudicating between them that is itself uncontroversial.
This intractability is not, in Comsa’s framing, a temporary state of the field that will be resolved by more research within the existing programmes. The disagreements between consciousness theories are deep and not visibly converging. A pragmatic research agenda cannot wait for resolution that has no schedule. Either the field continues to ask the intractable question and produces no actionable output, or it shifts the question to something that can yield actionable output and accepts that the underlying metaphysical question remains open.
The argument has a precedent that Comsa does not explicitly invoke but that is structurally similar: the move in cognitive psychology, several decades ago, from asking what consciousness is to asking what the empirical correlates of conscious report are. The redirection did not solve the hard problem. It produced decades of substantive empirical research that would not have been produced otherwise.
What the Tractable Programme Looks Like
The positive proposal in the paper is the perceived-consciousness research programme. Users routinely apply consciousness vocabulary to AI systems. They describe systems as feeling, wanting, knowing, and intending. They form attachments to systems and treat their outputs as evidence of inner states. They make moral and legal arguments that depend on consciousness attribution. These behaviors are observable, measurable, and consequential. They are also, on the current evidence, only weakly correlated with whether the underlying systems actually have any of the states being attributed.
The research programme Comsa specifies would investigate what drives attribution, how attribution varies across user populations and system types, what real-world effects follow from attribution, and what interventions can shape attribution in directions that reduce harm. This programme is empirically well-defined. It has the methodological tools to make progress (survey instruments, behavioral experiments, observational studies of deployed systems). Its outputs are usable by industry actors who design AI systems, by policymakers who regulate them, and by ethicists who think about their consequences.
The programme is also less prone to the field-level paralysis that has affected the intractable question. Researchers can disagree about the underlying metaphysics while still cooperating on the perceived-consciousness work, because the perceived-consciousness work does not require resolution of the underlying metaphysics to proceed. This is a structural advantage that the paper emphasizes.
How This Relates to the Existing Perceived-Consciousness Literature
The proposal does not arrive in a vacuum. The perceived-consciousness research programme has been underway for several years, with substantial recent contributions from multiple angles. Lucius Caviola, Jeff Sebo, and Jonathan Birch’s 2025 Trends in Cognitive Sciences paper drew on animal consciousness attribution research to map the cognitive biases that will likely shape societal judgments about AI. Bongsu Kang and colleagues’ 2026 Computers in Human Behavior Reports paper provided the first quantitative empirical study of which textual features in LLM outputs drive perceived consciousness. The premature attribution ethics literature, including the work that produced the premature attribution ethics analysis, has been operating in this space for several years.
What Comsa’s paper adds is a framing argument that positions the perceived-consciousness work as the primary research agenda rather than as a peripheral concern. The earlier work tended to treat perceived consciousness as one important topic among many in AI consciousness research. Comsa’s argument is that, on the current evidence, perceived consciousness should be the central topic because the alternative central topic (actual consciousness) cannot currently be pursued in ways that yield actionable conclusions.
The repositioning has consequences for how AI consciousness research should be funded, staffed, and evaluated. It implies that researchers whose work focuses on cognitive biases in attribution, on the psychology of user interaction with AI systems, and on the policy implications of attribution patterns should be treated as central contributors rather than as adjacent to the main programme. It also implies that researchers whose work focuses on developing better detection methods for actual consciousness should be valued primarily for what their work tells us about attribution and about what would constitute compelling evidence for users and decision-makers, rather than primarily for whether their work successfully detects actual consciousness.
Why the Author’s Institutional Position Matters
The paper’s authority is amplified by Comsa’s position inside Google DeepMind. The perceived-consciousness research programme has been articulated previously by academic philosophers and cognitive scientists, but it has not been articulated by someone whose daily work involves the AI systems that the research is meant to apply to. The argument that AI companies should reorient their consciousness-relevant research around perceived consciousness is more compelling when it comes from inside an AI company than when it comes from external commentators.
This is not because the external commentators are wrong. It is because the internal position carries operational credibility that external positions do not. An external researcher arguing that DeepMind should redirect resources toward perceived-consciousness research is offering a recommendation that DeepMind can take or leave. A DeepMind researcher arguing for the same redirection is offering a recommendation that the company’s structure makes it more likely to act on, because the recommendation comes from someone who is subject to the constraints the company actually faces.
The paper is also relatively short and accessible by academic standards. This is consistent with its strategic goal. The paper is trying to shift research priorities, which requires that it be read by decision-makers as well as by researchers. A long treatise would not achieve that goal regardless of how rigorous its argument was. A focused arXiv preprint, written by a recognized DeepMind researcher, framed around an actionable repositioning, has a better chance of producing the priority shift it argues for.
What the Argument Does Not Settle
Comsa’s argument is explicitly programmatic and not metaphysical. It does not claim that AI is not conscious, that consciousness science is on the wrong track, or that the intractable question is unimportant. It claims that, given current methodological resources, the intractable question is not the right organizing target for applied AI consciousness research. The intractable question can remain open, can continue to be pursued by researchers whose work happens to bear on it, and can be revisited if methodological breakthroughs change what is currently tractable.
The argument is also compatible with positions that disagree about whether current AI systems are conscious. A researcher who believes current LLMs are plausibly conscious can accept the tractable-question framing as a complement to direct consciousness research rather than as a replacement for it. A researcher who believes current LLMs are clearly not conscious can accept the framing for the same reason. The programmatic argument does not require commitment on the underlying question, which is part of why it is useful as a basis for cooperation across positions that would otherwise be in conflict.
The result is a paper that does less work than it might appear to at first reading but does the work it sets out to do well. It does not deliver a verdict on AI consciousness. It delivers a research-priorities recommendation, grounded in an intractability argument that is hard to refute on current evidence, from a researcher whose institutional position gives the recommendation unusual weight. That is a substantial contribution to a field whose ratio of strong opinions to actionable conclusions has been less favorable than it should be.