People call AI aware more readily than conscious, even when behaviour matches

The agent label changed the answer

If an AI and a person are described as reacting in exactly the same way, do people assign them the same mental state? In a study reported by LMU Munich, the answer depended strongly on which word people were asked to judge. Nearly 1,100 participants read scenarios involving more or less responsive protagonists and rated them as conscious or aware. Some saw artificial agents; others saw human protagonists in the same situations.

Responsiveness raised ratings for both kinds of agent. Yet the university reports a striking gap on consciousness: even its most responsive AI was rated less conscious than its least responsive human. Awareness ratings, by contrast, followed much more similar patterns for AI and humans. The study is about human attribution. It neither proves that AI has subjective experience nor proves that such experience is impossible.

Two everyday words invite different kinds of judgement

In ordinary product language, 'aware of its surroundings' can describe a functional ability: a system detects an input and changes its output. 'Conscious' often asks a different question, whether there is something it is like for the system to have an experience. The words are not perfectly defined or used identically by every speaker. That ambiguity is exactly why the distinction in the LMU report is newsworthy. A respondent may accept that an artificial agent notices a sound yet resist saying that it experiences the sound.

The experiment tests how people apply these words. It does not define consciousness scientifically or solve whether any artificial system could possess it. It also does not claim that all AI systems have the same abilities. The vignettes manipulate descriptions of protagonists, so the observed difference is a difference in attribution under those descriptions. Keeping this distinction visible prevents two opposite overstatements: that participants proved AI is conscious because they used the word 'aware', or that their reluctance to use 'conscious' proves a machine could never have experience.

Terms in the story and the narrower claim each supports.

Term
Responsive
In this experiment
The protagonist reacts more or less to its environment
What cannot be inferred
That the protagonist understands or feels the event
Term
Aware
In this experiment
A participant rates the protagonist as noticing its surroundings
What cannot be inferred
A direct measurement of inner experience
Term
Conscious
In this experiment
A participant applies a more demanding mental-state label
What cannot be inferred
A laboratory test that detects consciousness itself

Term	In this experiment	What cannot be inferred
Responsive	The protagonist reacts more or less to its environment	That the protagonist understands or feels the event
Aware	A participant rates the protagonist as noticing its surroundings	A direct measurement of inner experience
Conscious	A participant applies a more demanding mental-state label	A laboratory test that detects consciousness itself

What participants actually judged

The researchers gave participants short written situations. An agent might respond to sounds or another person's emotional state, or be less responsive to its surroundings. One group saw an artificial protagonist, while another saw a human in matched scenarios. Participants then rated how conscious or aware that protagonist seemed. This is a clean way to ask whether the agent category changes a judgement when the described behaviour is held constant.

It also sets the study's limit. A vignette rating is not a neurological measurement, a test of a particular deployed language model, or a direct measure of experience. Participants could draw on beliefs about biology, machines and the meanings of the English terms. The paper's title refers to essentialism in mental-state attribution: the impression that being human matters beyond observable response.

The comparison built into the study.

Factor
Protagonist
Conditions
Human or artificial agent
Question answered
Does identity change the judgement?
Factor
Responsiveness
Conditions
More or less reaction to the surroundings
Question answered
Does observable behaviour matter?
Factor
Rating word
Conditions
Conscious or aware
Question answered
Does vocabulary change the boundary?
Factor
Outcome
Conditions
Participants’ ratings
Question answered
What people attribute, not what the agent experiences

Factor	Conditions	Question answered
Protagonist	Human or artificial agent	Does identity change the judgement?
Responsiveness	More or less reaction to the surroundings	Does observable behaviour matter?
Rating word	Conscious or aware	Does vocabulary change the boundary?
Outcome	Participants’ ratings	What people attribute, not what the agent experiences

What holding behaviour constant actually buys

The researchers' key comparison changes the category of the protagonist while retaining the same described situation. Imagine a short scenario about an agent responding to a noise. If one participant reads it with an AI protagonist and another with a human protagonist, any systematic rating difference cannot be explained simply by one version performing a more impressive action on the page. LMU also says scenarios varied in how strongly the protagonist responded to surrounding sounds or other people's emotions. That lets the study ask two questions: does more responsive behaviour move ratings upward, and does the human-versus-AI label still matter when that behaviour is matched?

The release reports nearly 1,100 participants across experiments, but does not give a condition-by-condition sample table, recruitment method, demographic breakdown, rating-scale anchors or uncertainty intervals. We therefore cannot independently calculate the size of a subgroup effect or draw a percentage chart from the announcement. The study design is stronger than a general poll asking 'Can AI be conscious?' because it specifies a behavioural comparison. It remains a vignette study, not an observation of how respondents behave during a sustained interaction with an actual system.

Consciousness and awareness did different work

LMU reports that people credited both human and artificial protagonists with more awareness and consciousness when they acted more responsively. On the consciousness scale, however, the AI category remained below the human category so strongly that the most responsive AI still trailed the least responsive human. On the awareness scale, the gap narrowed substantially and the response patterns were similar. That makes a one-line claim such as 'people think AI is conscious' inaccurate; it also makes 'people never use mental words for AI' inaccurate.

The public university release gives the directional comparison but no full distribution of scores or effect-size table. The linked data and code record was embargoed until 7 October 2026 when checked on 2 October. A numerical chart would therefore overstate what is independently inspectable today. The qualitative table below preserves the comparison without inventing bars or percentages.

Results reported by LMU; qualitative because numerical distributions are not publicly inspectable yet.

Rating
Consciousness
Effect of responsiveness
Higher when the protagonist responds more
AI versus human comparison
Clear gap; most responsive AI below least responsive human
Rating
Awareness
Effect of responsiveness
Higher when the protagonist responds more
AI versus human comparison
Broadly similar pattern across AI and human protagonists

Rating	Effect of responsiveness	AI versus human comparison
Consciousness	Higher when the protagonist responds more	Clear gap; most responsive AI below least responsive human
Awareness	Higher when the protagonist responds more	Broadly similar pattern across AI and human protagonists

The key result is an interaction between behaviour and vocabulary

Both human and artificial protagonists received higher mental-state ratings when they were described as more responsive. That first result matters: participants did not ignore the described behaviour or automatically assign every AI a zero. The second result is the separation between the two rating words. On consciousness, the category boundary was so pronounced in LMU's report that the most responsive AI was still judged less conscious than the least responsive human. On awareness, the AI and human ratings followed much more similar patterns. The same kind of behavioural cue was processed differently depending on the concept respondents were asked to apply.

The university's public wording does not supply means, distributions, confidence intervals or exact contrasts for us to plot. A proportional bar chart would therefore fabricate precision. The safe visual is the qualitative comparison already shown in this article. The Zenodo record for the authors' data and scripts says its files are embargoed until 7 October 2026. That is a useful future audit point, not a reason to invent numbers today.

How far the evidence allows a reader to go.

Claim
Responsiveness affected participants’ ratings
Status
Supported by LMU release
Reason
Both agent categories were rated higher when more responsive
Claim
The word changed the AI-human gap
Status
Supported by LMU release
Reason
Consciousness separated the categories more than awareness
Claim
A specific AI model is conscious or not
Status
Not tested
Reason
Participants rated written scenarios, not subjective experience
Claim
A numerical effect-size ranking
Status
Not available from release
Reason
Full rating distributions and data are not public on 2 October

Claim	Status	Reason
Responsiveness affected participants’ ratings	Supported by LMU release	Both agent categories were rated higher when more responsive
The word changed the AI-human gap	Supported by LMU release	Consciousness separated the categories more than awareness
A specific AI model is conscious or not	Not tested	Participants rated written scenarios, not subjective experience
A numerical effect-size ranking	Not available from release	Full rating distributions and data are not public on 2 October

Why the paper uses the word essentialism

The paper's title, AI is not as conscious as humans: Essentialism in mental state attributions to artificial systems, points to the possibility that people treat being human as relevant to consciousness beyond an observed response. That is an interpretation of their judgement pattern, not proof that a particular biological ingredient is necessary for experience. The reported consciousness gap is compatible with a belief about the kind of entity being described, but the public release does not isolate exactly which beliefs or prior experiences drove each participant's rating.

It is also possible for a person to speak functionally in one setting and philosophically in another. Someone may say a camera is 'aware' of movement when its software triggers an alert, while reserving 'conscious' for a being with subjective life. The study supplies controlled evidence that this verbal distinction appears in group judgements. It does not license a universal claim about every language, profession, culture or future AI design without further testing.

The practical consequence is language discipline

A product description that says a system is 'aware of a sound' can mean it detects an input and changes its output. 'Conscious' often implies something stronger: subjective experience. The study suggests readers do not treat those terms as interchangeable even when the agent's behaviour is identical. For journalists and companies, the accurate move is to name the observable operation: the model detected speech, classified an emotion cue, or responded to a user message. That says more than an unsupported claim that it felt, understood or experienced anything.

A follow-up worth watching is whether the final paper and released data show the same separation across different scenarios, participant groups and rating wording. The present conclusion is narrower and useful already: behaviour influences attribution, but the category 'AI' and the precise mental-state term still shape the judgement. The study measures that public distinction, not the true metaphysics of minds.

How to describe the capability without smuggling in a mind

The reporting implication is practical. Describe the input, operation and output before choosing a mental-state verb. If a system transcribes speech and raises an alert, say that. If it classifies an image and changes a workflow, name the classification and action. This gives readers a testable account of the product. Saying it 'understood fear' or 'felt concern' implies more than those operations establish. Conversely, avoiding every word such as 'aware' is unnecessary if the sentence makes clear that it means detection or monitoring.

For AI product teams, this is also a copy-review problem. A claim such as 'the assistant knows when you are upset' can be rewritten as 'the assistant uses text cues to select a response'. The second sentence tells a buyer what the system actually does and creates a more useful question: how often does it misclassify those cues? The study does not measure trust, attachment or purchasing behaviour, so those downstream effects should not be attributed to it. It demonstrates that the vocabulary chosen in a description can alter the judgement readers are being asked to make.

Examples of testable capability language; these are editorial rewrites, not quotes from the study.

Loose description
The AI knows you are distressed
More precise description
It detects text cues associated with distress and offers a response
Question to test
What is its error rate on those cues?
Loose description
The camera is conscious of movement
More precise description
Motion detection triggers recording or an alert
Question to test
When does it miss or falsely flag motion?
Loose description
The assistant understands your goal
More precise description
It infers a likely goal from the prompt and proposes steps
Question to test
Does the proposed plan match the user’s stated constraints?

Loose description	More precise description	Question to test
The AI knows you are distressed	It detects text cues associated with distress and offers a response	What is its error rate on those cues?
The camera is conscious of movement	Motion detection triggers recording or an alert	When does it miss or falsely flag motion?
The assistant understands your goal	It infers a likely goal from the prompt and proposes steps	Does the proposed plan match the user’s stated constraints?

The next evidence check is the paper and released data

When the authors' data and scripts become public, the first checks should be the sample sizes in each condition, exact vignette wording, rating scales, exclusions and uncertainty around the AI-human contrasts. The result may also differ by scenario: reacting to a sound is not the same cue as responding to an emotional state. None of these unknowns invalidates the reported qualitative pattern, but they set the boundary of what can be concluded from a press release on 2 October. Until those materials are inspected, the article should keep its numbers to the sample size and the directional comparisons LMU actually supplies.