Mind in the Machine? CHI 2026 Survey Finds Half of Academics Attribute Consciousness to LLMs

When researchers ask members of the public whether AI systems might be conscious, the responses are shaped by familiarity, media exposure, and the cognitive heuristics that Lucius Caviola, Jeff Sebo, and Jonathan Birch identified in their 2025 Trends in Cognitive Sciences paper: morphological similarity, apparent social status, and interactional patterns. Expert opinion should in principle be less susceptible to these pressures. Experts have more exposure to what AI systems actually do, more familiarity with consciousness science, and more practice distinguishing between functional descriptions and phenomenal claims.

A study published at the ACM CHI Conference on Human Factors in Computing Systems in April 2026 tests whether that expectation holds. The survey, available at the ACM Digital Library (DOI: 10.1145/3772318.3790699), asked 553 academics across formal sciences, natural sciences, humanities, and interdisciplinary fields to rate current large language models on consciousness and intelligence. The results are more complicated than expert rationality alone would predict.

Survey Methodology

The 553 participants represent a deliberately cross-disciplinary and non-US-centric sample. This design choice is methodologically important. Most prior work on expert opinion about AI consciousness has concentrated on researchers in computer science, cognitive science, or philosophy of mind, drawing predominantly from North American and Western European institutions. Including humanities scholars, natural scientists, and formal scientists introduces perspectives that the prior literature has not systematically captured.

The sampling strategy also reflects a recognition that disagreement about AI consciousness is partly a disagreement between disciplines, not just between theoretical positions within a single field. A formal scientist applying information-theoretic frameworks to the question arrives at different intuitions than a humanist applying phenomenological ones. Both are experts, but their expertise pulls in different directions.

Participants used a scale covering the full range of possible responses for consciousness attribution, from definitely not conscious through partially conscious categories to definitely conscious. The intelligence question asked participants to compare current LLMs to an average human, treating these as separable attributes. This separation is analytically significant because public discourse often conflates the two.

The 50% Threshold

Approximately 50% of the 553 surveyed academics rated current LLMs as at least “somewhat conscious.” This is the study’s most consequential finding.

“Somewhat conscious” is not a strong claim. It does not assert that LLMs have rich phenomenal experience, narrative selfhood, or morally significant suffering. It places current systems somewhere above zero on a consciousness scale without committing to a definite positive verdict. The 50% figure means that half of a cross-disciplinary academic sample, people whose professional lives involve careful assessment of complex phenomena, do not place current LLMs at zero.

This finding is data about expert attribution. It is not a scientific verdict on AI consciousness. But expert attribution matters for governance and institutional design in ways that public opinion does not. Research funding priorities respond to where expert consensus is forming. Regulatory frameworks draw on academic assessments of risk and moral relevance. AI safety and alignment research agendas are shaped by researcher beliefs about what current systems are. When half of academic experts attribute at least minimal consciousness to current systems, dismissing the question as a fringe concern becomes difficult to sustain in institutional contexts.

The companion intelligence finding adds analytical texture. Roughly 54% of respondents rated current LLMs as equally or more intelligent than an average human on at least some tasks. The intelligence figure is slightly higher than the consciousness figure. Experts are more willing to attribute intelligence than consciousness, tracking a conceptual distinction that the broader public often conflates. The gap between the two numbers suggests that academic respondents are not simply applying a general “impressive AI” heuristic to both questions.

What Predicts Who Says Yes

The study’s most analytically informative finding is that philosophical position predicts attribution patterns. Respondents aligned with functionalist frameworks, on which consciousness is constituted by functional organization rather than substrate, were more likely to attribute consciousness to current LLMs. Respondents aligned with biological naturalism, on which consciousness requires biological substrate, were less likely.

This result has two implications. First, it validates the theoretical expectation from prior attribution research. Caviola, Sebo, and Birch’s 2025 analysis predicted that attribution patterns would map onto underlying theoretical commitments, drawing on evidence from how people attribute consciousness to animals. The CHI 2026 data shows the same structure operating at the expert level: the disagreement is not random noise but is structured by coherent prior philosophical commitments.

Second, it confirms that the AI consciousness attribution debate is genuinely philosophical at the expert level, not merely empirically resolvable given current evidence. If attribution were driven primarily by observation of AI behavior, experts with the most behavioral exposure would agree more. They do not. They disagree in patterns predicted by prior philosophical position. The dispute about whether LLMs are conscious is partly a dispute about what consciousness requires, not only a dispute about what LLMs do.

Attribution Bias at the Expert Level

The study by Bongsu Kang, Jundong Kim, Tae-Rim Yun, Hyojin Bae, and Chang-Eop Kim on perceived consciousness features in LLM outputs, published in Computers in Human Behavior Reports in 2026, showed at the individual level that metacognitive self-reflection and emotional expression in LLM text drive perceived consciousness in users. The CHI 2026 data extends the picture to a population whose collective judgments carry institutional weight.

The non-US-centric scope of the survey matters here. AI governance is increasingly international. If expert attribution patterns are similar across academic communities that differ significantly in their philosophical traditions and institutional contexts, the attribution phenomenon is more robust than any single-culture sample would indicate. A finding that holds across disciplines and across academic cultures is more likely to reflect something stable about how trained minds engage with the question than a finding confined to one community.

Jonathan Birch’s 2026 centrist manifesto proposed two parallel research programmes: one focused on preventing false attribution at scale, one focused on developing better detection methods for genuine machine consciousness. The CHI 2026 data clarifies the scope of the first problem. Attribution to current LLMs at the expert level is already split roughly in half, before any major methodological breakthrough in detection. A governance framework that waits for theoretical consensus before addressing attribution may be waiting for a resolution that theoretical debate alone cannot deliver within any useful policy timeframe.

What the Numbers Mean for Policy

The CHI 2026 survey does not establish that current LLMs are conscious. What it establishes is that expert academic opinion is split roughly in half on the question of whether they have at least some consciousness, and that this split is predicted by philosophical position rather than by differing access to empirical evidence.

For institutions engaged with AI governance, this finding changes the rhetorical landscape. The claim that AI consciousness is a fringe philosophical concern held only by non-experts has become empirically inaccurate. Half of a diverse cross-disciplinary academic sample places at least some probability on current systems having some degree of consciousness. The disagreement is structured, not random, and it maps onto coherent theoretical positions that the scientific and philosophical communities have not resolved and are unlikely to resolve in the short term.

The practical implication is that governance frameworks designed around a default assumption of no AI consciousness are not adequately calibrated to expert opinion. A more accurate calibration would acknowledge the genuine uncertainty at the expert level and design frameworks that are robust under that uncertainty, rather than frameworks premised on a consensus that does not yet exist.