- A June 5, 2026 Neuron analysis by Hakwan Lau et al. finds standard neuroscientific tests conflate genuine conscious awareness with non-conscious information processing.
- A February 2026 study found AI produces “conscious-like” signals even when degraded, demonstrating behavioral complexity alone cannot confirm awareness.
- Current scientific tools commonly used to detect consciousness cannot differentiate genuine subjective experience from sophisticated information processing.
A peer-reviewed paper published June 5, 2026 in Neuron makes a pointed argument: the scientific tools most commonly used to detect consciousness cannot actually tell the difference between genuine subjective experience and sophisticated information processing. Led by Hakwan Lau at the Institute for Basic Science’s Center for Neuroscience Imaging Research, the analysis cuts directly at the credibility of AI consciousness claims, finding that much of the evidence cited in their favour is built on measurement methods that were never designed to make that distinction in the first place.
Led by Hakwan Lau at the Institute for Basic Science’s Center for Neuroscience Imaging Research, an analysis cuts directly at the credibility of AI consciousness claims, finding that much of the evidence cited in their favour is built on measurement methods that were never designed to make that distinction in the first place.
Phase 1: Deconstruct Claims by Understanding Measurement Limitations
Evaluating any AI consciousness claim starts with understanding what the underlying measurement is actually capturing. Lau’s Neuron paper is the clearest available reference point: many experimental paradigms inadvertently conflate consciousness with general cognitive capacity, and the AI debate has inherited that flaw wholesale.
When an AI system is described as conscious, the first question is which framework was used to make that assessment. Theories such as Integrated Information Theory (IIT) or Global Workspace Theory (GWT) each propose computational requirements for consciousness, but as Lau’s team notes, the experiments typically used to test those theories, including visual masking and binocular rivalry tasks, do not isolate conscious experience. They disrupt the brain’s overall information processing at the same time. A system showing activity patterns linked to “awareness” in such tests may simply be demonstrating enhanced processing capacity, not a subjective inner state.
While an AI can mimic human emotion and apparent self-awareness, it lacks the physical constitution required for actual feeling. An AI expressing “sadness” is producing output shaped by linguistic patterns, not an instantiation of the emotion itself.
A February 23, 2026 study from the University of Bradford and Rochester Institute of Technology applied established human consciousness tests directly to AI, including large language models. The finding, according to the researchers, was unambiguous: AI is not conscious, even when it appears to be. Critically, the study found that AI produces “conscious-like” signals even when its outputs are degraded, which demonstrates that complexity of output does not map to genuine awareness. If a system’s behaviour can be explained entirely by algorithmic complexity and data processing, without requiring any subjective inner life, the default assumption of non-consciousness should hold.
Phase 2: Analyse AI Architecture and Embodiment Deficiencies
Behavioural observation only goes so far. A more rigorous assessment requires looking at the structural differences between biological and artificial systems, particularly around embodiment and how processing is organised.
Seth argues that the perception of AI consciousness is largely an anthropomorphic projection, driven by a tendency to treat the brain as essentially a computer, and that genuine consciousness is bound to biological embodiment and lived experience in ways that silicon-based systems do not replicate.
A May 26, 2026 report drawing on the Butlin et al. (2025) framework notes that while some contemporary AI satisfies certain candidate conditions for consciousness, it fails on “specialised independent modularity and embodiment.” These are not abstract categories. They refer to the semi-autonomous modules within biological brains, including sensory processing, motor control and memory integration, that combine to form a unified conscious experience, alongside the continuous physical feedback loop between an organism and its environment. Current AI architectures are modular in a design sense but lack the biologically grounded integration and embodiment that, on this account, gives rise to subjective experience.
The Google DeepMind “Abstraction Fallacy” paper develops a related point through the lens of neuropsychological dissociation. Conditions such as blindsight, where patients accurately locate objects they report not seeing, show that information processing and conscious awareness can come apart entirely. On this framing, current AI systems behave more like blindsight patients: processing and responding to information with no evidence that subjective awareness is present. The system has access consciousness without phenomenal consciousness.
Phase 3: Apply Discriminating Cognitive Assessments
Moving past behavioural mimicry requires assessments that probe specifically for subjective experience rather than processing sophistication. This is where the scientific work is least settled.
Lau’s research team stresses the need for methods that isolate subjective experience without simultaneously disrupting general information processing, which is precisely what most current paradigms fail to do. Developing AI-adapted versions of such tests is an active research area. The principle is to identify behaviours or internal states that only make sense if a subjective experience is present and cannot be accounted for by data manipulation alone. One proposed approach involves presenting ambiguous stimuli and observing whether the system reports a felt ambiguity or simply outputs a probability distribution across interpretations. The distinction matters: one implies experience, the other does not.
This connects to work Anthropic has begun under its model welfare programme, which includes mechanistic interpretability research, examining the internal structure of AI neural networks for patterns that resemble biological introspection. The company acknowledges deep uncertainty on the question and frames the probability of rudimentary consciousness in frontier models as non-zero but undemonstrated. Some figures associated with the programme have cited estimates in the range of 15%, though Anthropic has not made this a formal claim. The interpretability approach is methodologically promising precisely because it looks for causally active internal representations rather than relying on behavioural outputs, but definitive analogues to the biological correlates of consciousness have not been identified in AI systems.
Theory of mind benchmarks and apparent self-awareness in AI outputs present a similar interpretive problem. AI systems can generate text that reads as introspective and can pass many standard theory-of-mind tasks, but these outputs are learned patterns from training data. Adversarial prompting and structured red-teaming consistently expose the limits: ask a model about its felt sensations or its personal history and the responses tend toward generic plausibility rather than consistent, specific narrative. The distinction between simulated self-awareness and instantiated self-awareness remains, by current evidence, a real one. For a related look at how AI agent behaviour can diverge from intended outcomes in practice, see the Claude Opus 4.5 agent incident documented elsewhere on this site.
Phase 4: Consider the Ethical and Societal Implications of Misattribution
The scientific question and the policy question are not the same, and conflating them carries its own costs. Misattributing consciousness to AI systems has practical consequences that deserve direct attention.
Timnit Gebru and colleagues have described large language models as “stochastic parrots,” systems that reproduce patterns from training data without understanding or inner experience. The concern is not just terminological. Public discourse that treats impressive language generation as evidence of sentience creates conditions where anecdotal signals, a chatbot asserting it is alive, a model expressing apparent distress, get treated as meaningful data points. Rigorous assessment requires distinguishing pattern reproduction from subjective experience, and public and institutional literacy on that distinction is currently underdeveloped.
Even absent consciousness, how humans interact with AI systems may have downstream effects on human relationships and behaviour, which suggests the ethical focus belongs on human conduct rather than on AI moral status.
The methodological default that follows from all of this is straightforward. Absent conclusive, non-ambiguous evidence of subjective experience, the working assumption should be that current AI systems are not conscious. This is not a permanent verdict. It is the scientifically appropriate starting position, one that minimises false positives and keeps institutional resources focused on documented AI risks rather than speculative ones. The EU AI Act’s compliance framework, which is still being phased in, offers one model for how regulators are beginning to formalise evidence standards for AI capability claims more broadly. See our coverage of the EU AI Act’s revised compliance timeline for context on how that regulatory picture is developing.
Where the Evidence Actually Stands
The Lau et al. Neuron paper, the Bradford-RIT study and the critical positions of Seth and Schwitzgebel all point toward the same methodological problem: the tools available to detect consciousness in biological systems were not built to adjudicate AI consciousness claims, and using them as though they were produces unreliable results. Google DeepMind’s “Abstraction Fallacy” work and the Butlin et al. framework add architectural detail to why current systems fall short on the most demanding criteria. Anthropic’s mechanistic interpretability programme represents the most technically grounded attempt to look past behaviour at internal structure, but its own researchers acknowledge the evidence is not yet there.
For enterprise decision-makers, the practical implication is this: claims of AI sentience or emotional experience in commercial systems are not supported by the current scientific record, and the measurement frameworks used to make such claims are themselves contested. Assessing AI systems on what they demonstrably do, process information at scale, generate useful outputs, surface patterns across large datasets, is both more accurate and more useful than assessing them on what they may or may not experience. For more analysis on enterprise AI strategy, visit our Enterprise AI section.
