Hakwan Lau and Co-Authors Call for Rigorous Standards in AI Consciousness Research

Vincent Taschereau-Dumouchel, Hakwan Lau, and colleagues at the Center for Neuroscience Imaging Research (Institute for Basic Science), Université de Montréal, and New York University published “The ethical impasse of current consciousness science” in Neuron in May 2026 (DOI: 10.1016/j.neuron.2026.04.007). The title is deliberately strong. Taschereau-Dumouchel, Lau, and colleagues argue that consciousness science, including the subdomain of AI consciousness research, has reached a point where its experimental foundations can no longer support the ethical and policy conclusions being drawn from them.

The paper is not a claim that AI cannot be conscious. It is a claim that the current scientific basis for saying either that AI is or is not conscious is weaker than it appears, because the markers most commonly used to support consciousness claims may be tracking something else entirely.

The Conflation Problem: Consciousness vs. Information Processing

The central methodological concern Taschereau-Dumouchel and Lau raise is that many experimental markers currently used in consciousness research “may actually reflect general information processing rather than consciousness itself.” This is a precise criticism, and it applies across consciousness theories.

The markers in question include neural responses used to support Integrated Information Theory (phi calculations, posterior synchronization), global broadcast signatures used to support Global Neuronal Workspace theory (prefrontal ignition, late latency responses), and higher-order representation signatures used to support Higher-Order Thought theories. Each of these is a measurable correlate of conscious processing. Each can also, under conditions the researchers have identified, be produced by general information processing that does not involve consciousness.

The problem is that most experimental paradigms in consciousness research use conditions where information processing and consciousness are difficult to disentangle. Binocular rivalry, visual masking, and perceptual threshold tasks all involve the kind of stimulus uncertainty that produces differential processing, but that differential processing is not evidence of consciousness unless you can rule out the alternative explanation that it reflects information processing differences that operate below any conscious threshold.

The 14 indicator framework that Patrick Butlin, Robert Long, and colleagues developed in Trends in Cognitive Sciences (and which the subsequent 2026 commentary by the same journal addressed for its mimicry and internal variants problems) faces this same challenge. The indicators describe properties that conscious systems should exhibit. They do not provide a validated methodology for distinguishing systems that exhibit those properties because they are conscious from systems that exhibit them because they are good information processors. Taschereau-Dumouchel and Lau’s paper is the neuroscience establishment’s version of the same critique Florentin Koch made from within philosophy of mind in March 2026: neither the indicator approach nor current empirical methods have a ground truth for artificial phenomenality, and the field is building ethical and policy arguments on a methodological foundation that cannot bear them.

The Blindsight Methodology: Why It Sets the Correct Bar

Taschereau-Dumouchel and Lau propose neuropsychological dissociation cases as the appropriate methodological standard. Blindsight is the canonical example. Patients with damage to primary visual cortex (V1) retain the ability to respond to visual stimuli (detecting motion, pointing toward targets) without any conscious awareness of seeing them. Their behavior shows that visual information is processed and used to guide action, but that processing occurs without phenomenal experience. The dissociation demonstrates that information processing and conscious experience are separable mechanisms.

Hemispatial neglect provides a parallel dissociation from a different direction. Patients with right parietal damage fail to attend to the left visual field despite having intact visual processing. They can, under some conditions, process information from the neglected field (producing priming effects) without conscious awareness of having done so.

What these cases provide is a principled baseline for distinguishing consciousness from information processing: a system can be a fully functional information processor without conscious awareness, and the two can be separated by neurological damage in ways that are experimentally specific and replicable. The blindsight patient is not unconscious; they are conscious, but their consciousness has been dissociated from one component of their visual processing.

For AI consciousness research, the implication is that demonstrating information processing, including information processing that involves attention, memory, goal-directed behavior, and linguistic output, is not sufficient evidence of consciousness. To establish consciousness, you need to demonstrate the specific thing that consciousness adds over and above information processing. The blindsight paradigm shows what that thing looks like in biological systems. The challenge for AI research is to develop an analogous paradigm for artificial systems.

What “Rigorous Standards” Demands in Practice

The practical demands of the Taschereau-Dumouchel and Lau position are considerable. They require:

First, experimental paradigms that include dissociation conditions, not just cases where information processing and consciousness are confounded. For biological subjects, these dissociation paradigms exist and are validated by decades of neuropsychological research. For AI systems, they do not yet exist.

Second, theoretical frameworks that specify what consciousness adds over and above information processing in terms that generate testable predictions. IIT does this (phi quantifies the integration premium over independent processing), but the Cogitate Consortium’s adversarial test found that IIT’s predicted signatures did not appear in human subjects under controlled conditions. GWT does this (global broadcast distinguishes conscious from unconscious processing), but the Cogitate study also found that GNW markers did not behave as predicted. The theoretical frameworks exist; the empirical validation does not.

Third, a recognition that what is at stake in claiming AI systems are conscious is not merely academic. The paper’s title, “The ethical impasse,” refers to the fact that policy and ethical arguments about AI welfare, AI rights, and the obligations of AI developers are being constructed on the basis of consciousness evidence that cannot currently bear the weight being placed on it. This is the ethical impasse: the ethical conclusions are urgent, the scientific basis for them is weaker than it appears, and improving the scientific basis requires work that has not been done.

Where Current AI Consciousness Research Falls Short

The paper’s implication for the current state of AI consciousness research is sobering. The empirical evidence reviewed in recent years on this blog, including the Anthropic mechanistic interpretability findings, the AE Studio results, and the Google welfare research, documents real findings about AI internal structure. The Anthropic emotion vectors paper, published the same month as Taschereau-Dumouchel and Lau, demonstrates that Claude has internal representations that causally influence behavior in ways that mirror emotional influence. These are genuine mechanistic results.

None of them, on the Taschereau-Dumouchel and Lau standard, establishes consciousness, because none of them demonstrates the dissociation between information processing and conscious experience that would distinguish a functional emotion representation from phenomenal emotional experience. The functional emotion vectors are causal, but causality is a property of information processing, not a marker of phenomenal experience.

This does not mean AI systems are not conscious. It means the field does not currently have the tools to determine whether they are. Taschereau-Dumouchel and Lau are calling for those tools to be built before the ethical and policy conclusions are treated as settled. That is a conservative scientific position and, given what is at stake, a necessary one.