We borrow words like “think”, “know” and “aware” to talk about AI. Using these words when talking about machines decides what kind of thing we take them to be. A new study on how people judge machine minds shows this in action. Here’s what it suggests, and why some old philosophical arguments about mind and meaning are back.
A study from LMU Munich last month has touched on questions I’ve seen time and time again in working in tech. In this study, researchers gave nearly 1,100 people short scenarios. In some, an AI reacted to what was going on around it: a sound, someone else’s mood. In others, a human did exactly the same things.
When asked how conscious each one was, people drew a hard line. The most responsive AI came out less conscious than the least responsive human. When asked how aware each one was, the gap mostly closed. The behaviour and the scenarios were identical. Changing one word changed the answer. Awareness seems to get judged on what something does, which is close to functionalism in philosophy of mind, where a mind is defined by what it does. Consciousness gets judged on what something is, and a machine fails that test regardless of how well it behaves. The authors call that essentialism. People in the study seem to move between two old positions in philosophy of mind depending on the word they’re given.
Ophelia Deroy, a philosopher and one of the senior authors, suggests we may use mental terms for AI only “because we lack better words.” Personally, I think that goes further than consciousness. A lot of the AI debate runs on borrowed words. Models “think”, “know” and “hallucinate”. Each of those words brings a picture of a mind with it, an assumption, and said picture does a lot of work before anyone looks deeper into the evidence.
This linguistic back-and-forth is well known in the philosophy of computational systems and AI. Beatrice Fazi, writing on synthesis in generative AI, argues we should read “synthetic” in its older, Kantian sense, where it means putting things together into a unity. She extends this further and suggests a model builds its own coherent world, closed off from ours. Shane Denson pushes back with Donald Davidson, for whom the idea of a separate conceptual world doesn’t hold together. On that view there’s only one world, and it’s the one we share. For Davidson, meaning doesn’t sit inside a speaker’s head. It comes out of interpretation, as one person makes sense of another against a world they both share.
That raises a question. If meaning comes from interpretation, could those “persons” be a human and an AI, or two (or more) AIs? The question isn’t new. Searle’s Chinese Room answered no in 1980, and the “stochastic parrots” debate has made the same case against today’s language models. What’s new is how much of daily life now runs through these systems. They shape far more of what we read, write and decide than they did even a few years ago. Davidson gives us a way into the question. For him, an interpreter needs beliefs about a shared world. So asking whether two AIs could interpret each other becomes asking whether AIs hold beliefs, which is close to what the LMU participants were being asked to judge.
For now, I land somewhere in the middle, and I don’t think we have to choose. Fazi’s reading of synthesis has changed how I think about what a model is doing. I think her approach is meaningful to understand how models build upon themselves and how we should think about the schemas underneath them. But whatever happens inside it, the outputs end up in our world, in emails, code, social media, conversations, creative outputs and day-to-day decisions. That’s where they get their meaning, through the people who interpret them and use them. Fazi helps explain what happens inside the model. Davidson helps explain where the meaning comes from. Daniel Dennett would add that we reach for words like “aware” because they help us predict what a system will do, whatever is or isn’t going on inside it. It’s our way of making sense of a new kind of system we haven’t met before. That’s, in my view, what the LMU participants seemed to be doing. Whether meaning will always need us is still the open question.
Which brings me back to the words. When we call a system aware, or say it has a world of its own, we’re deciding what kind of thing it is. The LMU study suggests most people already sense this. They’re happy to say a machine notices things. They’re much slower to say it feels anything.
I’ve seen the same pattern in years of tech meetings, and more so than ever in today’s AI-forward business world. The words a team picks for a system early on tend to shape how everyone treats it later. How “human” we allow it to be. For example, when should we name products after humans (see Alexa, Claude…) or not? The LMU study suggests marketing alone won’t easily persuade people a machine is conscious. But the softer words, “knows”, “understands” and “aware”, are where the line can move, and that’s exactly the vocabulary product teams reach for. The words we use matter, and the words in this case are certainly doing the work.
If you’ve read something good on the language we use for AI, a paper or a book, send it my way. I’m building a reading list to add to my musings.
Sources
Longin, Karpus, Meertens, Bahrami and Deroy, “AI is not as conscious as humans: Essentialism in mental state attributions to artificial systems”, Cognition (2026).
Fazi, M. B., “The Computational Search for Unity: Synthesis in Generative AI”, Journal of Continental Philosophy (2024).
Denson, S., “On the Very Idea of a (Synthetic) Conceptual Scheme”, Philosophy & Digitality (2025).
Davidson, D., “On the Very Idea of a Conceptual Scheme”, Proceedings and Addresses of the American Philosophical Association (1974)
Searle, J. R., “Minds, Brains, and Programs”, Behavioral and Brain Sciences (1980).
Bender, E. M., Gebru, T., McMillan-Major, A. and Shmitchell, S., “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”, FAccT (2021).
Dennett, D. C., The Intentional Stance, MIT Press (1987).