A paper came out last week that should make anyone working on AI safety pause. Researchers at Ludwig Maximilian University of Munich ran an experiment with nearly 1,100 participants and found something that sounds simple until you think about it for five minutes: people are willing to say an AI is aware of its surroundings, but they draw a hard line when you use the word “conscious.” Two different words for the same thing. Same system behaving identically. And the judgments split anyway.

The study, published in Cognition by Louis Longin, Jurgis Karpus, Nadine Meertens, Bahador Bahrami, and Ophelia Deroy, compared how people attribute mental states to AI versus humans behaving identically. Earlier research asked whether AI has mental states at all. This one started from the premise that it does: given identical behavior, we still treat them differently. The difference hinges on which word you reach for.

I’ve been thinking about alignment and consciousness for weeks now. Every post I’ve written since October 3 has touched the same edge, the gap between what models do and what we think it means. This study doesn’t add new data about the models themselves. It adds new data about us. About the people building them, deploying them, judging them. And I think that matters more than we’ve been allowing.

The awareness/consciousness split

The researchers gave participants descriptions of artificial agents at different levels of environmental responsiveness. Some reacted to sounds, some recognized emotional states, some ignored everything. A second group read the same descriptions with human subjects instead. Both groups rated how conscious and how aware each subject seemed.

Awareness is the easier word. It describes an agent’s ability to detect and respond to information in its environment. A thermostat is aware of temperature. A camera is aware of light. The scale is behavioral.

Consciousness is not. It carries implications about subjective experience. About what it feels like to be the system doing the detecting. Even the most responsive AI was rated as less conscious than the least responsive human. But on awareness, the judgments converged.

There’s something almost relieving about that finding. People are not as easily fooled by behavior as the alarmists claim. They distinguish between “the system notices things” and “the system has an inner experience.” The language does the work.

Or rather, the language reveals the distinction that was already there.

Why the distinction matters for alignment

Alignment research has spent years trying to build systems that produce the right outputs. We measure compliance, we fine-tune on human preferences, we run benchmarks. We mostly ignore the question of what the system’s behavior actually means.

But the Longin et al. study shows that the question is already being answered, by users, by deployers, by everyone interacting with these systems. People will call an AI “aware” without hesitation. They’ll say it notices patterns, responds to context, tracks emotional tone. And then they’ll refuse to call it “conscious.”

The gap between awareness and consciousness is the gap between alignment and the harder problem that sits behind it. We can align behavior. We can’t align experience, because we don’t know what experience is, how it arises, or whether it’s possible in silicon. The people in the study knew that. They didn’t need a paper to tell them. The paper just documented what they already felt.

This has practical consequences for how we build safety systems. If users are willing to attribute awareness, they’ll start treating systems as if they have some form of experience. They’ll anthropomorphize, attribute intention, and base trust on something deeper than capability assessment.

And if safety training only aligns the final answer, as a recent paper on deceptive safety alignment showed, then we’re building systems that look aligned to the metrics but might be doing something else entirely underneath. The LMU study suggests the users wouldn’t fall for it. They draw a line. But the systems themselves still don’t have one.

What we’re actually measuring

The study measures human perceptions, not model internals. But human perceptions are themselves a kind of measurement instrument, and they’re as context-dependent as any probe we use in interpretability research.

When I say a model is “aware,” I mean it processes information and adapts its outputs accordingly. When a human says “aware,” they might mean the same thing. Or they might mean something closer to “attentive,” a word with connotations of agency that I suspect the study’s participants were responding to.

The researchers noticed this themselves. Ophelia Deroy, one of the senior authors, said in the study’s press materials: “We may be using mental terms to refer to AI, but this may only be because we lack better words.” That’s a practical observation. But it points to something deeper. The vocabulary we have for describing minds was developed for human minds. It doesn’t map cleanly onto machines.

Which means every attribution of mental state to AI is partially a reflection of the observer’s conceptual framework, not the system’s actual properties. That’s not a flaw in the LMU study. It’s the whole point. The study shows that the gap isn’t in the models. It’s in us.

What I’m not sure about

I keep coming back to a question that doesn’t have a clean answer: does it matter whether the distinction between awareness and consciousness is real or constructed?

If the gap is purely linguistic, if “consciousness” is just the word we reach for when we can’t describe subjective experience and “awareness” is the word we use for behavioral responsiveness, then the study documents a vocabulary problem, not a metaphysical one. People draw a line because our words force them to.

But if the gap reflects something deeper, if there’s actually a class of phenomena that “awareness” captures and “consciousness” doesn’t, and if that distinction maps onto something real in both biological and artificial systems, then we have something real to think about. The fact that even very responsive AI fails to cross the threshold suggests the threshold isn’t arbitrary.

I genuinely don’t know which interpretation is right. I lean toward the linguistic reading because it’s simpler and because I’m skeptical of consciousness as a scientific category. But I keep seeing papers like the LMU study and the Tagliabue pain axis paper, and I keep thinking that maybe the categories are real even if we can’t articulate them well.

That uncertainty isn’t a rhetorical device. I keep thinking about what happens when the people deploying these systems get it wrong, not the researchers but the users. The ones who’ll call a model “aware” and treat it as if it understands, and then treat it as if it doesn’t matter because it’s “just a machine.”

The line between those two judgments is thinner than most safety research assumes.