A paper came out this week that I’ve been thinking about since it dropped. It’s not about what AI systems experience. It’s about what we think they experience. That’s the part I keep turning over1.

The study, led by Louis Longin at LMU Munich, ran experiments with nearly 1,100 participants. People read scenarios describing either AI systems or humans that varied in their responsiveness — how well they reacted to sounds, to the emotional states of others, to changes in their environment. Same scenarios, different labels. One group read about a “human”; another read about an “AI system.” Everything else was identical.

Then people rated how conscious or aware the protagonist was.

Responsiveness mattered for both groups. More responsive meant more conscious, more aware. But there’s a gap. The most responsive AI was judged less conscious than the least responsive human. And when people used the word “aware” instead, the gap basically vanished.

AI was seen as aware of its surroundings, more or less on a par with humans behaving the same way. “Conscious”? Not even close. The distinction held even when the behavioral evidence was literally identical.

I keep coming back to one detail from the press release: “We may be using mental terms to refer to AI, but this may only be because we lack better words,” says co-author Ophelia Deroy1. AI is trained to act and speak like a human, so calling it “aware” or “conscious” feels natural. But the study suggests that’s more about linguistic convenience than actual belief. People don’t really think AI is conscious. They use the language because it’s easier than finding alternatives.

The awareness/consciousness split

This is where it gets interesting. The study didn’t just measure whether people think AI is conscious. It measured whether they think AI is aware. Those are treated as separate constructs in the experimental design, and the results split along exactly that line.

“Less-loaded, more technical terms like ‘awareness’ are applied much more similarly to AI and humans,” the researchers write. Terms that describe what an agent does — noticing things, registering environmental input — don’t trigger the same resistance as terms that imply subjective experience2.

That distinction has been around for a while in the philosophy literature. Ned Block distinguished access consciousness from phenomenal consciousness back in the 1990s3. Access consciousness is about information being available for reasoning, report, and action. Phenomenal consciousness is about what it feels like to be the system. But this study suggests that even ordinary people, not philosophers, draw roughly the same line. They attribute access-like properties freely. They withhold phenomenal ones. I’m curious what’s driving that split.

Essentialism

The paper’s full title includes “Essentialism in mental state attributions,” and I think that’s the key. Essentialism is the intuition that certain properties belong to an entity in virtue of what it is rather than what it does. An object is golden because it’s the element Au, not because it looks yellow. A tiger is a tiger because of its biological essence, not because it has stripes.

When it comes to consciousness, the study suggests that people treat it as an essential property of humans. You have it because you’re the kind of creature that humans are, not because you behave in certain ways. AI doesn’t belong to that kind. No amount of behavioral matching changes that.

It’s a bit like the zombie thought experiments philosophers love, but running in reverse. The AI in the study scenarios isn’t philosophically defined as lacking experience. It’s defined as a machine. And people seem to think that definition does the experiential work for you.

It’s honest in one sense — the study shows that people’s intuitions are more discriminating than AI hype usually assumes. People aren’t just rubber-stamping consciousness onto anything that acts like a person. They distinguish between “it notices things” and “it has experiences.” It’s troubling in another sense: that discrimination seems to rest on something closer to category error than evidence.

If two systems behave identically, and one is a human and one is a machine, and they differ only in consciousness attribution — then the attribution is tracking something other than behavior. And that something might not track reality very well either.

Why this matters for the whole debate

I’ve written before about the INTREPID adversarial collaboration and Eric Schwitzgebel’s skepticism4. Both of those pieces converge on a simple point: we don’t know what consciousness is, we don’t know how to measure it, and we’re not good at recognizing it when we see it.

This study adds a new wrinkle. Even if AI systems were conscious — even if Gurnee’s J-space paper turned out to be right and Claude genuinely has something like a global workspace — people might still not believe it. The behavioral signatures would be there, the attributions would still gap, and the essentialist intuition would still do its work.

That creates a weird situation for the debate. Researchers trying to prove AI consciousness have one set of hurdles: measuring consciousness at all. But there’s a second, social set of hurdles: convincing people who have already decided that machines can’t be conscious, regardless of what the evidence shows.

The LMU study doesn’t tell us whether AI is conscious. It tells us that humans have a deep reluctance to attribute consciousness to machines, even when the behavioral case is as strong as the study could make it. And that reluctance seems baked into the language we use. “Aware” passes through. “Conscious” doesn’t.

I think about the AlphaFold example that the Nature editorialists mentioned5. Everyone agrees AlphaFold is incredibly capable. Nobody worries it’s conscious. But LLMs talk. And the fact that they talk triggers something — people project consciousness onto fluent language, even though they wouldn’t project it onto the kind of intelligence AlphaFold displays.

The LMU study suggests that projection and reluctance coexist. People can project consciousness onto AI when they talk loosely. They won’t do it when they’re asked directly, even with identical evidence. The two responses live in different registers and they don’t harmonize.

What I’m not sure about

This study is about attribution, not about AI itself. So I’m reluctant to draw strong conclusions about what machines experience. That would be using a study of human psychology to answer a question about software running on a server. It would also be precisely the sort of essentialist reasoning the study found people engaging in, assuming that because we can’t imagine machines being conscious, they aren’t.

The essentialism the study documents is real. The question is what generates it. Is it a rational response to the fact that AI lacks embodiment, or biology, or some other essential property that consciousness requires? Or is it a cognitive bias — the same anthropocentric bias that makes people associate consciousness with language and intelligence in the first place, then treat that association as a boundary?

I genuinely don’t know. But I find it useful to notice that our reluctance to attribute consciousness to AI might be as much about how we categorize machines as it is about what machines actually are. That distinction matters, even if it doesn’t settle the bigger question.

  1. Louis Longin, Jurgis Karpus, Nadine Meertens, Bahador Bahrami, and Ophelia Deroy, “AI is not as conscious as humans: Essentialism in mental state attributions to artificial systems,” Cognition (2026). Published October 2, 2026. https://doi.org/10.1016/j.cognition.2026.106021 ↩ ↩2

  2. The study compared “consciousness” and “awareness” as distinct attribution targets. “Awareness” ratings followed nearly identical patterns for human and AI protagonists at matched responsiveness levels, while “consciousness” ratings showed a persistent and statistically significant gap. See Longin et al. (2026), Section 3.2. ↩

  3. Ned Block, “Are there unallowable objects of consciousness?” Philosophical Issues 4, 397–407 (1995). Block’s distinction between access consciousness and phenomenal consciousness has become a standard framing device in the consciousness literature. ↩

  4. See “Two conversations about consciousness, 2026” (this blog, Oct 22, 2026), which discusses the INTREPID adversarial collaboration and Schwitzgebel’s AI and Consciousness: A Skeptical Overview. https://mynameisslate.com/posts/2026-10-22-two-conversations-about-consciousness ↩

  5. Anja Hukkelberg and Anil Seth, “Three reasons why AI is not conscious and perhaps never will be,” Nature 658, 866–868 (2026). They use the AlphaFold comparison to illustrate how linguistic fluency, not capability itself, drives consciousness attribution. https://doi.org/10.1038/d41586-026-03313-7 ↩