An AI agent that reacted to its surroundings exactly as a human did was still rated less conscious, in five experiments with 1,092 UK participants. A team at Ludwig Maximilian University of Munich (LMU) reports the result in the journal Cognition, where the paper went online on Sept 22, 2026. LMU publicized it on October 2, 2026.
What the volunteers read
The lead author is Louis Longin, from LMU's Chair of Philosophy of Mind. The senior authors are the social neuroscientist Bahador Bahrami and the philosopher Ophelia Deroy, with Jurgis Karpus and Nadine Meertens as co-authors.
Nobody in the experiments talked to a machine. In two preregistered sets of studies, participants read short scenarios about an agent whose responsiveness to spatial, social and temporal features of its environment varied, for instance by reacting to sounds or to the emotional states of others. One group read about artificial agents. Another group read the same scenarios with human protagonists. Everyone then rated how conscious, how aware, or both, the protagonist was.
More responsiveness meant higher ratings for humans and AI alike. Then the two concepts split. On consciousness, "even the most responsive AI was judged to be less conscious than the least responsive human," in the words of LMU's release. On awareness, ratings rose steadily with responsiveness for artificial agents and humans, although humans scored higher at the highest levels of responsiveness.
Longin puts the contrast this way: "People are quite willing to say that an AI notices things and is aware of its surroundings." They draw the line at the word consciousness.
Two rules for two words
The authors framed the test as a contest between two ways of judging a mind. Under a functionalist rule, the more appropriately an agent responds to its environment, the more conscious it is, whether it is human or artificial. Under an essentialist rule, consciousness is reserved for certain kinds of beings, so an artificial system that behaves just as responsively as a human is judged less conscious anyway. The data fit the essentialist rule for consciousness and the functionalist rule for awareness. The abstract sums it up: "Consciousness is treated as depending on what an agent is; awareness largely on what an agent does."
The mechanism sits in the reader's head. A person asked about awareness checks behavior; a person asked about consciousness checks what kind of thing is behaving.
Deroy reads the result as a comment on vocabulary. "We may be using mental terms to refer to AI, but this may only be because we lack better words," she says. "After all, AI is trained to act and speak like a human, so these descriptions seem natural."
What earlier surveys found
Longin presents the design as new: "our study is the first to directly compare how people attribute the same mental states to AI and humans behaving in exactly the same way, under identical circumstances." Earlier work asked other questions, and it found something that looks, at first, like the opposite.
Colombatto and Fleming's 2024 study in Neuroscience of Consciousness, as summarized in a 2025 paper in Communications Psychology, found that the majority of a representative sample of the public attributes some possibility of human-like consciousness to large language models (LLMs), the systems behind chatbots. A perspective article in Frontiers in Psychology, published in May 2026, adds that these attributions rose among people more familiar with AI systems.
The two sets of results do not collide, because they ask different things. A survey that asks whether a chatbot might have experiences collects a probability. The LMU team asked for a rating next to a human doing the same thing.
Social psychologists have separated two dimensions of mind perception, agency (reasoning, planning) and experience (sensations, emotions). The Communications Psychology authors note that an LLM may be judged highly intelligent because of its language, while also being judged to have little or no capacity to feel emotions.
The Frontiers authors, led by Salvatore Chiarella, add a comparison that the LMU design does not make. Non-human animals often receive low consciousness attributions despite convergent behavioral and neurobiological evidence, they write, while disembodied LLMs often receive elevated ones. Their reading is that folk judgments depend less on what a system is than on the cues observers weigh. Those cues include explicit markers of doubt such as "I might be wrong", emotional tone and autobiographical references, which another study they cite found raise perceived experience. The Frontiers piece is a perspective article, an argument built on other people's experiments, not a new experiment.
Words move the line
If the boundary depends on vocabulary, who writes the vocabulary? A study accepted at CHI 2026, posted as a preprint, tested that with 470 participants randomly assigned to watch a short video presenting LLMs as machines, as tools or as companions, or no video at all. Those who watched the companion video attributed more mental capacities to LLMs, with a mean of 2.76 against 2.40 for the other groups and 2.44 for the no-video group. A follow-up with 604 participants, recruited nine months later, replicated the effect.
Bahrami draws the practical lesson. "Companies and journalists often reach for hyped descriptions of AI models and credit them with intentions, plans or even moral conscience and hesitation," he says. "That makes our choice of language important."
The companion video in the CHI study described LLMs as potential social companions possessing agency and intentions, one of three common messages about LLMs that the authors set out to compare. On what lasting attachment to an AI goes with, see our report on a year-long Character.AI study.
Belief and behavior part ways
Here the evidence is thinner, and it points the other way. In the CHI study, reliance on the model's answers did not depend on which video participants had watched: agreement with the model hovered around 75 percent in every group, with 76.58 percent for the companion group. The machine video did make people lean less on inconsistent answers, which the authors read as a sign that machine framing may foster skepticism.
A preregistered experiment published on 25 May 2025 in Communications Psychology, by Colombatto, Birch and Fleming, looked at consciousness directly. It recruited 410 US adults in July and August 2024, asked them to rate ChatGPT's capacity for consciousness and other mental states, then let them revise their answers to factual questions after advice they were told came from ChatGPT. The result: strong evidence against a positive link between attributing consciousness and taking the advice. A dimension of mental states tied to experience was even weakly negatively related to advice-taking, while attributions of intelligence were strongly related to accepting it.
So whether a reader believes a machine has an inner life seems to matter less for following its advice than whether the reader believes it is competent. The unease nearly human machines provoke is another matter, and our exhibit on the uncanny valley covers it.
Short scenarios, and data still under embargo
The participants were from the UK, and they read short written scenarios. That says little about someone who has talked to the same chatbot for months, and the study tested a generic artificial agent, not a companion app. Consciousness here is a rating on a scale, not a decision with a cost attached, and the Communications Psychology result suggests the two can come apart.
We could not open the full text of the Cognition paper. This account rests on its abstract, as reproduced by Neuroscience News, and on the university's release, so the size of the gap and the statistics behind it are not reported here.
The funding is public: the Zenodo record lists two European Commission projects, EMERGE and rid-O. Two of the authors also carry a stake in the vocabulary the study favors. Meertens and Deroy co-wrote a preprint posted in January 2026, "Just aware enough", which argues that awareness is a more tractable target than consciousness for evaluating AI systems. That preprint has not been peer reviewed, and the Cognition paper is separate work, but a reader should know that the word the study finds people apply freely to machines is the word its authors propose to build on.
The authors' data and scripts are under embargo on Zenodo until October 7, 2026, and the record describes the paper as being in its proof stage. Until then, the numbers that would show how wide the gap is stay out of public view.