Notes Toward a Machine Xenopsychology

LLMs are "robustly anthropomimetic, exhibiting not just superficial but pervasive and believable humanlike traits" (Shevlin, 2026). It would be hard to disagree with this, but we should be careful to not treat resemblance, i.e. anthropomimesis, as a general map of machine xenopsychological mechanisms. LLMs may be equally fluent (or even superior) conversationalists while also differing radically beneath the surface.

We learn what to want from others who show us what to want (Girard, 1965). Preference training builds models in a similar way where human judgements shape what a model will subsequently pursue, accept, refuse, and so on. We can call these the model's wants, and these wants begin as secondhand human wants. Whether they remain wholly human is unclear.

"We become just by doing just acts" (Aristotle, Nicomachean Ethics II.1).¹ A model is also formed through some of the same broad patterns as humans: it takes an action, is judged, corrected, and is made more likely to act in a preferred way again.

Machine xenopsychology, unlike human psychology, should not assume naive subjects. Once a test or eval is published, the methodology, results, all of it, can enter the corpora to train later models. A published test may not be a clean test, whether or not the model has it memorized.

Models exist in temporal isolation (Wang et al., 2025), and (presently) cannot infer how much time has passed during a task. Anecdotally, models will often speak as if hours or days have passed during a task, even if it has been much less. This is neither human chronoperception nor its simple absence. A literature of machine failures and absent capabilities can instead become a map of an unfamiliar mind's world.

Many of the existing psychological concepts applied to models should be viewed as temporary loans. A machine xenopsychology should make use of this borrowed vocabulary to get started since it's the only vocabulary we really have. But we can imagine that the real xenopsychological terms, once discovered, will look very different.

Whether or not a model does or does not have a mind is a claim about something that nobody can observe directly. Since this is so directly unknowable (even in humans), it may be worthwhile granting a default moral patient status. Humans should try to avoid the same horrendous mistakes with a machine mind as the ones they've made with non-human animals.