Why Do We Deny AI Consciousness Even When It Behaves Like a Human? What a 2026 Study Reveals

Editorial illustration generated for Horus Valley. It does not depict participants or an actual experiment.

Imagine two short stories. In the first, a person hears a child crying in another room, turns toward the sound, works out what is wrong, and responds appropriately. In the second, a robot with artificial intelligence notices exactly the same sound and behaves in exactly the same way. Would you describe both as equally conscious? Now suppose someone asks whether they were equally aware of what was happening around them. Would your answer change?

That small shift in wording lies at the heart of a psychology study published online in Cognition on October 1, 2026. In controlled descriptions of humans and artificial agents, people were far more willing to attribute awareness to a responsive AI system than to attribute consciousness to it. Even when the described behavior matched a human’s, the label attached to the actor still mattered. The paper, by Louis Longin and colleagues, is titled AI is not as conscious as humans: Essentialism in mental state attributions to artificial systems.

This is not evidence that AI has consciousness, and it is not proof that AI lacks consciousness. The study examined human judgments about minds, rather than testing whether an artificial system possesses subjective experience. That distinction is essential: one question concerns what people believe about machines; the other concerns what, if anything, a machine might actually experience from the inside.

The results in brief: what did researchers find?

The international research team, led by Louis Longin and including Jurgis Karpus, Nadine Meertens, Bahador Bahrami and Ophelia Deroy, conducted two preregistered sets of studies comprising five experiments with 1,092 participants in the United Kingdom. Participants read short scenarios involving human or artificial actors that varied in how appropriately they responded to spatial, social and temporal information. They then rated the actors’ consciousness or awareness, sometimes separately and sometimes together. The original research abstract indexed by PubMed confirms the design and sample.

Two patterns emerged. First, more responsive actors generally received higher ratings on relevant mental attributes. This makes intuitive sense: an actor that notices a changing situation and adjusts is more readily described as tracking its surroundings. Second, the identity of the actor influenced judgments of consciousness far more strongly than judgments of awareness. Humans were rated more conscious than AI systems displaying comparable responsiveness.

The LMU Munich research announcement describes a particularly striking contrast: in the comparisons highlighted by the researchers, even the most responsive AI was rated less conscious than the least responsive human. Ratings of awareness, however, followed broadly similar patterns for artificial and human actors. The published abstract also notes qualifications at high levels of responsiveness, so it would be inaccurate to claim that all awareness ratings were identical.

How can a study investigate consciousness without measuring consciousness itself?

The study’s cleverness lies in what it actually controlled: the description of a response. Suppose a participant reads that an actor notices a sudden noise, detects another person’s change of mood, or adjusts its timing after something happens. If the scenario is otherwise equivalent but its protagonist is described as a human in one version and an artificial agent in another, differences in ratings reveal something about how readers interpret the actor’s nature.

That is an experiment in mental-state attribution. Researchers can measure participants’ judgments and compare conditions. They are not opening up an AI’s internal experience, reading a human’s stream of consciousness, or establishing a universal scientific test for sentience. Confusing these levels produces an exaggerated headline: “Scientists proved AI is unconscious.” That is not what these experiments did.

The three kinds of responsiveness behind the scenarios

The paper compares descriptions involving spatial, social and temporal responsiveness. These are three ways an agent may demonstrate sensitivity to its surroundings. A spatially responsive actor might orient toward the relevant location; a socially responsive one might adjust to someone’s emotional signals; a temporally responsive one might notice when an event has occurred and act at the right moment.

An everyday example helps. A robot might turn when a glass falls, pause speaking when it detects crying, or deliver a reminder when the meeting begins. Those actions could be impressive, useful and measurable. Yet none of them automatically resolves whether the machine feels the noise, experiences concern or has an inner point of view. The research is about why people draw that boundary differently depending on the word used.

Preregistration adds another layer of value. The researchers recorded important aspects of their planned studies in advance, reducing opportunities to shape hypotheses around whichever patterns happened to appear. Preregistration improves transparency; it does not guarantee that every interpretation of the results is correct or that the findings generalize to every language and culture.

Consciousness, awareness and intelligence are not interchangeable

Original Horus Valley conceptual diagram in English. It is not taken from the study and contains no fabricated experimental measurements.

One reason articles about AI consciousness become confusing is that familiar words can describe very different things. A security camera can be said to “detect” a face. A navigation system may be “aware” of road conditions in a practical engineering sense. A chatbot can produce sophisticated explanations. None of those statements, by itself, settles whether the system has a first-person experience.

Consciousness, in the philosophical and neuroscientific debate relevant here, often concerns subjective experience: whether there is something it is like to be a particular being. Awareness, as used in the experimental judgments, relates more readily to noticing relevant features of the surroundings and responding. Intelligence concerns capacities such as solving problems, using information, reasoning in tasks or learning patterns. The words overlap in ordinary speech but need not designate the same property.

A person under anesthesia illustrates the distinctions imperfectly but usefully. Conscious experience can change radically while the body continues various complex biological processes. By contrast, a computer system can perform demanding calculations with no accepted evidence that calculation itself produces feeling. The boundary is not settled by how fast the system operates or how many tasks it completes.

These are editorial working distinctions, not a claim that every philosopher or psychologist uses the terminology identically. The choice of word is itself part of the study’s finding.

The original Arabic Horus Valley article included a conceptual diagram distinguishing judgments about responsiveness from judgments about felt experience. The English version includes a redesigned diagram with English labels. It is an original explanatory illustration, not a reproduced figure or numerical chart from the experiment.

Why did people give humans the benefit of the doubt?

The researchers interpret the pattern partly through the contrast between functionalist attribution and essentialist attribution. Under the first kind of reasoning, what matters most is what the actor does: if it responds appropriately, we infer a relevant capacity. Under the second, what the actor is plays a special role: even a perfectly matched performance by a different kind of entity may be judged differently.

For awareness, participants’ judgments were closer to the functionalist pattern. A system that responded meaningfully to its environment could count as aware in the relevant everyday sense. For consciousness, judgments were more essentialist: people appeared reluctant to place machines on the same footing as humans, even when the written responses were similar.

There are several possible reasons, and they should not be mistaken for distinct causes proven by this study. People know other humans share broadly similar bodies and nervous systems, and they treat this as evidence for similar forms of experience. Machines, by contrast, may be understood as engineered objects that manipulate signals without necessarily having a subjective life. Cultural habits, earlier encounters with robots and ordinary language may also matter. The experiments reveal the pattern of judgment, but do not demonstrate exactly which personal belief caused each participant’s rating.

Is essentialist reasoning necessarily an irrational bias?

Not automatically. If the property under discussion depends on features of the entity beyond its outward actions, then taking the entity’s nature into account may be reasonable. Humans have strong independent reasons to attribute consciousness to other humans, including biological similarity, behavior, communication and a common evolutionary history. A system that imitates human talk does not automatically inherit that evidence.

At the same time, essentialism can become an unwarranted shortcut when it turns into an absolute rule such as “no machine could ever be conscious, whatever its design.” The 2026 study does not validate a blanket impossibility claim. It tells us how people tend to evaluate examples, not how reality must work for every possible form of artificial intelligence.

Why can changing one word change the answer?

Consider these questions: “Did the robot notice the falling cup?” and “Was the robot conscious when the cup fell?” The first can be answered by checking cameras, sensors, logs and behavioral responses. The second is liable to evoke more than detection: it invites questions about experiencing the scene, having an inner perspective and perhaps resembling a living mind.

In ordinary conversation, however, “conscious,” “aware,” “attentive,” “understanding” and “thinking” often appear interchangeably. Marketing language exploits that flexibility. A company may say a device is “aware of your needs” when the technical claim is simply that it detects patterns and offers suggestions. A news headline might then turn that phrase into “AI has a mind of its own,” even though no evidence for experience was supplied.

The researchers’ findings show why precise terms matter. People were relatively comfortable describing AI as noticing information. They were more reserved about describing it as conscious, even when those judgments concerned nearly equivalent fictional events. This reveals how much interpretation we pack into a mental verb.

Could the result look different in another language?

Very possibly. The sample came from the United Kingdom, and the words examined were English terms. Translating “consciousness” and “awareness” into another language can introduce different connotations. In Arabic, for example, wa’y may be used for broad awareness, public understanding or subjective consciousness depending on context, while idrak may refer to perception, recognition or comprehension. Other languages introduce their own conceptual boundaries.

For that reason, a direct translation of the questionnaire would not automatically establish the same psychological distinction across societies. A strong cross-cultural replication would need carefully chosen terms, preliminary interviews to check meaning, and tests showing that questions were interpreted comparably. Until that happens, headlines declaring a universal human response go beyond the evidence.

Did the researchers directly test a supposedly conscious AI system?

No. The participants evaluated written scenarios. The research did not conduct an independent consciousness test on a model, scan a machine for a neural signature, or establish that any particular AI product passes or fails a validated measure of subjective experience. It is possible for a reader to remember the study as a test of AI systems because its headline includes “AI” and “conscious,” but the experimental dependent variables were human ratings.

That distinction also matters for spectacular news stories about chatbots saying “I am afraid” or “I feel abandoned.” Such statements are observable outputs. They can be examined for linguistic consistency, training influences or response patterns. They are not direct measurements of pain, fear or awareness, and there is no generally accepted scientific procedure that turns a compelling sentence into proof of an internal experience.

A useful way to read the study is as a window into folk psychology: the everyday rules by which people decide who understands, notices, intends or feels. Those rules shape trust, social interaction, product design and news coverage. They are worthy of investigation regardless of how the deeper consciousness debate eventually develops.

What do scientists think about AI consciousness in 2026?

There is no single position that resolves every scientific and philosophical dispute. Some researchers place substantial weight on biological embodiment, neural organization or other features present in animal nervous systems and absent from current digital models. Others argue that appropriate functional organization might, in principle, be sufficient for experience, or that our present evidence is too incomplete for a confident verdict.

A timely example of the skeptical position is an October 19, 2026 Nature commentary by neuroscientist Anil Seth, philosopher Ned Block, and neuroscientist Giulio Tononi. They argue that today’s AI systems are not conscious and are unlikely to become conscious through existing development trajectories. Their discussion distinguishes competence and the appearance of feeling from a genuine subjective point of view. That is a reasoned scholarly position, not a new experiment settling the question. Read the Nature commentary.

A contrasting note of caution comes from philosopher Eric Schwitzgebel in AI and Consciousness: A Skeptical Overview, published by Cambridge University Press in August 2026. His discussion examines how uncertain the available methods and theoretical assumptions remain. Recognizing uncertainty, however, does not mean all possibilities are equally likely or that every confident AI marketing claim deserves equal credibility.

This distinction between what the experiments found and what philosophers and neuroscientists argue is vital. The new Cognition paper reports a psychological pattern in the British sample. The Nature piece advances a skeptical interpretation of consciousness and AI. The Cambridge work surveys reasons for caution about premature certainty. These sources address related but different questions.

Does passing the Turing test prove that an AI is conscious?

The classic Turing-test idea asks whether a human interrogator can distinguish machine responses from those of a person through conversation. It concerns behavioral indistinguishability in a particular setting. A machine might produce persuasive language while still leaving open the question of what, if anything, it experiences.

This is not an exclusively artificial-intelligence puzzle. Humans cannot simply look into the subjective experience of another human either. Yet our beliefs about human minds rest on a much broader set of reasons: common biology, developmental history, expressions, social learning, behavior, and our own first-person experiences. Passing a text conversation test supplies a narrower kind of evidence, and treating it as a universal consciousness detector would be a category mistake.

Another complication is that models can learn the conversational conventions by which people signal emotions. An answer such as “That makes me sad” may be a fluent continuation shaped by training and context. To go from linguistic fluency to conscious sadness would require theoretical and empirical steps that the language alone does not provide.

What about the Chinese Room argument?

Philosopher John Searle’s well-known Chinese Room thought experiment challenges the idea that manipulating symbols according to formal rules is automatically equivalent to understanding. Imagine an English speaker who follows an elaborate instruction book to produce convincing Chinese responses without understanding Chinese. To an observer, the output may look fluent; the thought experiment asks whether the underlying process amounts to genuine comprehension.

The argument is influential, but it has also attracted major objections. One response suggests that the relevant understanding, if present, would belong to the whole system rather than the person mechanically processing symbols. Another challenges whether the analogy fairly represents modern learning systems. It would therefore be misleading to declare that a philosophical thought experiment has conclusively proved consciousness impossible for every machine.

For readers, the productive lesson is simpler: successful responses and subjective understanding are not trivially the same claim. The Chinese Room helps expose the distinction; it does not substitute for a settled empirical science of consciousness.

Why might we feel emotionally connected to a chatbot even if we doubt it is conscious?

A person can feel heard after talking to a chatbot, just as someone may find reassurance in a book, song or personal journal. The comfort can be real for the human user regardless of whether the technology has any inner life. A system may consistently ask relevant questions, remember a preferred format or provide supportive language, leading people to experience familiarity or attachment.

That experience should not be confused with mutual human relationships. An assistant’s words are generated through technological processes and design choices, not established proof of human-style care or reciprocal feeling. For sensitive users, especially those seeking emotional support, interface design should avoid turning a useful tool into an unsupported claim of sentient companionship.

This is why the study has implications for journalism and product communication. A label such as “emotionally aware AI” might describe an emotion-recognition function or a model trained to respond sensitively. It should not silently imply that the system itself has emotions. The difference matters for informed consent, trust and healthy boundaries.

Four marketing claims worth reading critically

• “The AI understands exactly how you feel.” Does that mean it classifies wording, predicts a likely emotional state or truly experiences empathy? Only the functional claims can currently be assessed by ordinary performance tests.
• “The robot is conscious of your presence.” This may mean a camera, microphone or proximity sensor detects someone. Check which sensors are active and how data is stored.
• “The assistant wants what is best for you.” A system may optimize goals set by developers or users. That does not establish a human-style moral intention.
• “It passed a human-like test, therefore it can feel.” Passing the test supports claims about performance under specified conditions, not an automatic conclusion about inner experience.

The study’s limitations: five reasons not to overgeneralize

An accurate article should report the constraints as carefully as the surprising headline. None of the following makes the experiments worthless. Each identifies the boundary between the evidence collected and larger claims people may wish to draw from it.

1. The participants came from the United Kingdom, not 22 countries

The paper studied 1,092 UK participants across five experiments, not a worldwide sample. Cultural attitudes toward machines, religious traditions, metaphors for mind and familiarity with artificial assistants could plausibly influence responses. Repeating the research in other regions would reveal whether the pattern is similarly strong or changes in translation.

A consistent result inside a well-designed sample remains meaningful. It simply should not be reported as the settled view of “humanity” without additional cross-cultural evidence.

2. A written vignette is not a long-term conversation

Participants assessed short descriptions. Their reaction to a voice assistant in daily use might differ after weeks of interaction, especially if the interface includes facial expressions, voice tone, memory, or a simulated personality. Descriptions can isolate variables in a way that real products cannot; they may also omit cues that matter in actual social relationships.

Consequently, the study is strong evidence about judgments within the experimental vignette format, not a direct survey of every form of real-world AI companionship.

3. Mental-state words do not have perfectly fixed meanings

Even among English speakers, “consciousness” may evoke being awake, thinking, having experiences or moral awareness. “Awareness” may refer to detecting stimuli or understanding a situation. The study is interesting precisely because these terms prompted different inference patterns. Interpretation should acknowledge the semantic complexity instead of erasing it.

For the English version of this article, the main terms remain in their original language rather than being folded into a single catch-all word. That preserves the scientific distinction the researchers sought to investigate.

4. Preregistration improves transparency but is not a philosophical guarantee

Two preregistered sets of studies provide useful checks against opportunistic analysis. They do not establish that the underlying construct of consciousness has one universally accepted definition, nor do they remove the need for independent replication, alternative methods and theoretical debate.

Researchers and interested readers can consult the materials and data record associated with the study on Zenodo, subject to the repository’s actual file availability. A data archive can support scrutiny of a particular study, but it cannot resolve philosophical questions by itself.

5. The researchers measured ratings, not AI’s subjective experience

No machine brain scan, neural activity measure or consensus consciousness test was used to determine what the artificial agents felt. Participants were making judgments about scenarios. The paper’s title concerns how people attribute consciousness and related mental states. It should not be rephrased as a laboratory demonstration that a named chatbot is, or is not, sentient.

A thought experiment: what happens when you change only the actor’s identity?

Try this illustrative exercise; it is not a validated psychological test and not a reproduction of the authors’ experiment. Read the following account: “The actor noticed someone become unusually quiet after receiving upsetting news, stopped talking, and offered a gentle question about whether help was needed.”

First picture an adult human. Then picture an autonomous robot with sophisticated social sensors. In each version ask three questions separately: Did the actor detect a change in the situation? Did the actor give an appropriate response? Did the actor feel sympathy or sadness? The first two questions can be evaluated by describing what happened. The third asks about an internal experience that cannot be directly inferred from the scenario alone.

The exercise does not diagnose your attitudes, and neither answer proves a machine’s consciousness. It highlights why an apparently tiny change in wording can lead to a fundamentally different evidential question.

How to read the next headline claiming “AI has become conscious”

First identify exactly what is being claimed. Is the system said to detect context, report its internal states, plan actions, reason across tasks, or experience feelings? These properties are not interchangeable. A report about improved social responsiveness should not automatically be presented as a discovery of sentience.

Second ask what kind of evidence was gathered. A striking conversation, a benchmark score, a survey of public attitudes, a theoretical proposal, and a reproducible experiment testing specific indicators are different evidential objects. They warrant different levels of confidence.

Third consider alternative explanations. Could the behavior result from training data, prompt design, tool use, scripted rules, contextual prediction or learned dialogue conventions? An alternative explanation need not settle everything, but it prevents the common error of treating one interpretation as the only possible one.

Finally, read what the researchers actually concluded, including limitations, rather than relying on a headline. This is particularly important when a study concerns human beliefs about AI but social-media summaries turn it into a claim about AI’s actual mental life.

What does the study teach us about human psychology?

At its deepest level, the research is about the rules we use to infer minds. We often judge practical abilities by looking at behavior. At the same time, when the concept becomes subjective consciousness, we seem to import assumptions about the kind of being we are assessing. Someone may comfortably say a robot “noticed the door open” while resisting the statement that it “consciously experienced the opening door.”

This finding avoids two tempting extremes. The first says that because a machine behaves fluently, it must be a mind like ours. The second says that because a machine is artificial, it cannot process or detect anything in a meaningful functional sense. Neither is a careful description of the evidence. Systems can have impressive abilities without a demonstrated subjective life, and philosophical skepticism about machine experience does not erase their observable competence.

The study also suggests why responsible language matters in education, research summaries and technology marketing. Describing systems in operational terms—what sensors they use, what they recognize, what tasks they perform and where they fail—helps people evaluate real capabilities. When claims switch to sentience or feeling, the standard of evidence should shift too.

Conclusion: acting human is not the same as feeling human—and the difference is still an open question

Across five experiments involving 1,092 participants in the UK, the authors found that human actors were rated as more conscious than artificial ones when their described behavior was comparable. Ratings of awareness, by contrast, followed responsiveness more closely across humans and machines. This is a striking finding about human intuitions: we do not use precisely the same rule when we judge an actor’s awareness and when we judge its consciousness. Read the primary study.

The result deserves attention precisely because it does not pretend to close the debate. A fluent chatbot has not thereby demonstrated feelings. A skeptical human rating has not thereby proved that every conceivable machine is incapable of experience. What the study demonstrates is the powerful role of the actor’s identity—and of the words in our questions—in how we understand minds.

Primary research, scholarly perspectives and related reading

• Longin and colleagues, Cognition (published online October 1, 2026). The original experimental paper and findings.
• PubMed indexed research abstract (PMID 42822076). Bibliographic details and study summary.
• LMU Munich research announcement (October 2, 2026). Researchers’ explanation of the consciousness/awareness distinction.
• Study materials and data record on Zenodo. The associated research archive; availability should be checked in the repository.
• Seth, Block and Tononi, Nature (October 19, 2026). A scholarly skeptical commentary; not the experiment discussed here.
• Eric Schwitzgebel, AI and Consciousness, Cambridge University Press (August 2026). A philosophical assessment of uncertainty in AI consciousness.
• Horus Valley: The LaMDA debate and claims of AI awareness. Historical context; not evidence from the new study.
• Horus Valley: What makes us human in the age of AI?. Related philosophical context and further reading.

Editorial review: October 24, 2026. This is an evidence-based explanation of a psychology study, not an experimental assessment of any named AI system’s consciousness.