Do LLMs have mental states?

You and I have probably ascribed beliefs, intentions and other mental states to LLMs in conversation. For example, I might say “Claude thinks XYZ,” or “ChatGPT just tries to agree with me.” It is a casual thing that we do, and, well, somewhat encouraged by the products themselves (Claude “fathoms” and “contemplates” while I await a response). Of course, when we do so, we do not mean that LLMs have mental states. That would be quite a radical ontological perspective to take, and few agree with the Hermeneutic school (attributing a certain ontology to you on behalf of your language) anyway. So what are we doing when we make these ascriptions? I would argue that it is an example of Berkeley’s infamous “think with the learned, and speak with the vulgar” maxim – we speak like this because it is a harmless and practical manner of speaking, but, upon reflection, we shouldn’t take it as ontology.
I also think that ascriptionism, which holds that if an LLM’s output seems enough like a belief or an intention then we should just attribute those mental states to it, is dumb. I often say things like “my phone thinks I am in the parking lot,” and to ascribe actual mental states to my phone simply because I posit it casually in conversation seems absurd. No one takes these phone statements literally and they are just a useful shorthand or façon de parler (Murray Shanahan can be credited with articulating this position very well. I have not done him justice but I did link his paper below).
Part of what also seems to anchor ascriptionists is their belief that human cognition might engage in the same sort of next-token prediction that LLMs do, and we feel perfectly comfortable ascribing mental states to the output that you and I have. To this I say, “yikes what” Even if our cognition did involve some degree of next-token prediction, our “training” is so vastly different from the pre- and post-training undergone by LLMs that the metaphor ascriptionists attempt to draw seems beyond faulty. What do I mean by this? Our cognition and mental states depend on ongoing perception/interaction with the physical world (this is a theory held by those in the embodied cognition camp, which I largely agree with). Since the moment we open our eyes, you and I are exposed to data that floods every sensory receiver we have. It is through this continuous and experiential interaction that we come to our current state of mind. And also, our outputs are continuously tested in the physical world with real stakes. If I touched a hot stove, I would learn not to do so again because it would burn my hand and cause me pain (a stake). By contrast, an LLM would simply read “a stove is hot and it will hurt if you touch it,” but has no way to verify this output against the external world (it could have just as easily been trained on “a stove is cold and it will hurt if you touch it”). In short, there is a fundamental difference between the experiential learning, based on 1) active response, and 2) stakes that you and I undergo, and LLM training. I once heard someone say that training an LLM is like putting a baby in front of a computer screen for 18 years (pre-training) and then taking it out behind a shed in the woods and hitting it until it acts the way you want it to (post-training). I do not know what to make of this and am curious as to what others think!
You might of course raise the concern that under my position, AI with a physical body would be able to be ascribed mental states as they (when we manage to do this technologically) have the continuous and world-based interaction that LLMs lack. In response to this, I will note two things. The first is that my claim is about current LLMs pre- and post-trained on large amounts of primarily text-based data. Second, as we veer into more robotics and world models, our views on what cognates might change, but still, sensors and embodied physical activity alone may not be enough to ascribe mental states (returning to my point about stakes) – though that is a topic beyond the scope of what I am trying to say here. If you asked me to guess 10 years into the future about whether or not I would believe robots have mental states, I have no idea what I would say.
Further readings on this topic
Shanahan, Murray. “Talking About Large Language Models.” arXiv, version 5, 16 Feb. 2023, arxiv.org/abs/2212.03551.
Shapiro, Lawrence, and Shannon Spaulding. “Embodied Cognition.” The Stanford Encyclopedia of Philosophy, edited by Edward N. Zalta and Uri Nodelman, Summer 2025 ed., plato.stanford.edu/archives/sum2025/entries/embodied-cognition/.
Hilbert, David. “George Berkeley.” The Stanford Encyclopedia of Philosophy, edited by Edward N. Zalta and Uri Nodelman, Stanford University, 12 Aug. 2026, plato.stanford.edu/entries/berkeley/.
