Even Human-Like AI Fails the Consciousness Test in Public Minds

Artificial intelligence systems now write poetry, negotiate, console, and deceive with apparent fluency, and public conversation increasingly borrows the vocabulary of minds to describe what they do. Yet a new study from Ludwig-Maximilians-Universität München suggests that beneath this loose talk lies a strikingly firm intuition: people refuse to grant AI the one property that matters most in our folk metaphysics of mind. Even when an artificial agent behaves in exactly the same way as a human being, under identical circumstances, observers consistently judge it to be less conscious. The research, led by Dr. Louis Longin of LMU’s Chair of Philosophy of Mind together with Dr. Bahador Bahrami, Professor Ophelia Deroy and other collaborators, was published in the journal Cognition and offers one of the most direct experimental probes to date of how ordinary people distribute mental states between machines and people.

The methodological innovation at the heart of the study lies in its controlled comparison. Previous investigations of AI mind perception typically asked general questions, such as whether an AI system has feelings, intentions, or experiences, without holding behaviour constant across the artificial and human cases. That left open the possibility that any reluctance to attribute minds to machines simply reflected differences in what the machines were described as doing. Longin and his colleagues closed this loophole. As Longin explains, their study is the first to directly compare how people attribute the very same mental states to AI and to humans behaving in exactly the same way, under identical circumstances. Any difference in judgement that survives this design cannot be explained by behaviour alone, and must instead reflect something about the category to which the agent belongs.

The experiments involved nearly 1,100 participants, a substantial sample for this kind of vignette-based research. Some participants read short scenarios in which artificial agents were more or less responsive to their surroundings, for example by reacting to sounds in the environment or to the emotional states of other characters. A second group read the same scenarios, but with human protagonists in place of the artificial ones. After reading, all participants rated how conscious or aware the relevant protagonist was of its surroundings. This manipulation of responsiveness was crucial: it allowed the researchers to trace how judgements of consciousness scale with behavioural sensitivity, and to test whether the same behavioural gradient produces the same psychological gradient for machines and for people.

The results revealed a double dissociation in how the two key terms behave. For both artificial and human agents, the more responsive the protagonist was to what was happening around it, the more conscious and aware it was judged to be. Responsiveness, in other words, drives mental state attributions upward across the board. But when the researchers compared attributions of consciousness specifically, a clear gap emerged between AI and humans, and the gap was not a marginal one. Even the most responsive artificial agent was judged to be less conscious than the least responsive human. Behaviour, however human-like, could not carry an artificial system across the line into consciousness as far as participants were concerned.

Attributions of awareness told a very different story. When participants were asked about awareness rather than consciousness, the pattern of judgements followed much the same trajectory for AI as it did for humans. The two terms, often treated as near synonyms in everyday speech, apparently do different psychological work. Awareness, a less loaded and more technical descriptor, seems to track what an agent does: if a system registers sounds, responds to emotional cues, and adjusts its behaviour accordingly, people are willing to say it is aware. Consciousness, by contrast, appears to be bound up with something deeper, namely the question of whether there is subjective experience, something it is like to be the agent. That question, for most participants, is settled not by behaviour but by what the agent is.

This asymmetry speaks directly to a worry that has been growing in public discussion: that as AI systems behave in ever more human-like ways, people will begin to treat them as if they genuinely possessed human-like mental states, with all the moral and social consequences that follow. Longin argues that the findings point to something far more nuanced than a simple slide toward anthropomorphism. People are quite willing to say that an AI notices things and is aware of its surroundings, he notes, but they clearly draw a line when it comes to the term consciousness. Across the experiments, participants consistently attributed less consciousness to AI than to humans, even when the behaviour described was identical. The word conscious may sometimes be used loosely to characterise what AI agents display, but the underlying distinction between humans and machines remains intact.

The researchers interpret this pattern in terms of essentialism in mental state attributions, a theme reflected in the article’s title, AI is not as conscious as humans: Essentialism in mental state attributions to artificial systems. On this reading, people do not treat mental states as properties that can be freely distributed to any system that behaves appropriately. Instead, they treat consciousness as tied to an essence, something intrinsic to being human, that no amount of behavioural mimicry can confer on an artefact. This would explain why the behavioural gradient, which reliably raises consciousness judgements for both kinds of agent, cannot close the gap between them: the gap is not located on the behavioural dimension at all. It is a categorical judgement layered on top of, and largely insulated from, the evidence participants are given.

The contrast with awareness suggests that this essentialist barrier is selective rather than absolute. Less emotionally charged, more functional terms are applied to AI and humans in much the same way, because they describe capacities and performances rather than inner experiences. Deroy, a philosopher and senior author of the study, points out that we may be using mental terms to refer to AI simply because we lack better words. After all, AI is trained to act and speak like a human, so such descriptions seem natural. Her observation reframes the debate: the apparent anthropomorphising of AI in everyday language may be largely a linguistic convenience, a shorthand for sophisticated behaviour, rather than a genuine attribution of an inner life. The new data give this intuition empirical teeth, showing that the loose talk and the deep belief come apart in a measurable way.

The findings also carry a practical lesson for how AI is communicated about in public. Bahrami, a senior author and expert in social neuroscience, observes that companies and journalists often reach for hyped descriptions of AI models, crediting them with intentions, plans, or even moral conscience and hesitation. What the study shows, he argues, is that everyday judgements are much more discriminating than such rhetoric assumes: people do not simply place AI and humans on the same mental scale. At the same time, the boundary depends on which terms are used. For more neutral descriptors such as awareness, the distinction between machines and humans can largely disappear. That makes the choice of language consequential, because different words invite different degrees of mental ascription and, potentially, different expectations about responsibility, trust, and moral standing.

Published on 1 October 2026 in Cognition under DOI 10.1016/j.cognition.2026.106733, the study arrives at a moment when questions about machine consciousness have moved from philosophy seminars into headlines, regulatory debates, and everyday interactions with conversational agents. Its central contribution is to separate the layers of public intuition that hype and fear tend to blur. People, it turns out, are neither indiscriminate anthropomorphisers nor stubborn deniers of machine capacity. They grant AI a functional, behavioural kind of mentality, an ability to notice and respond, while withholding the experiential core that makes consciousness matter to us. Whether that line reflects a deep truth about machines, a parochial bias in favour of our own kind, or simply the limits of our vocabulary remains an open question. What the LMU team has demonstrated is that the line is real, robust, and drawn with surprising precision by ordinary minds judging the minds of others, human or otherwise.

Subject of Research: Public attribution of consciousness and awareness to artificial intelligence compared with humans

Article Title: When AI behaves like us, we still don’t think it is conscious

Article References: When AI behaves like us, we still don’t think it is conscious. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, consciousness, awareness, mind perception, philosophy of mind, social neuroscience, LMU Munich, Cognition, mental state attribution, anthropomorphism, AI ethics, experimental psychology