Why calling optimisation behaviour "machine psychology" tells us more about human projection than about machine minds.
The phrase machine psychology has now entered the conversation with the sort of confidence normally reserved for things that have already done the intellectual work. It sounds settled, inevitable, even scientific. As if once a machine becomes sufficiently articulate, psychology simply shows up one day.
Much of this framing is associated with Robert Johansson (a prominent AI researcher and writer best known for probing large language models and interpreting their behavioural quirks often controversially blurring metaphor and mechanism), whose aforementioned probing is often careful, sometimes genuinely interesting, but then described in a language that does far more work than it ought to. Personality. Deception. Self-models. Anxiety. All very human words, doing a very human thing - making the unfamiliar feel familiar.
From an evolutionary standpoint, this is where the wheels start to wobble. Psychology is not not a descriptive flourish we apply when behaviour gets complicated enough to impress us. Psychology is an adaptation. It exists because bodies are fragile, and very easy to kill. Minds evolved because organisms needed ways of navigating danger, scarcity, status, sex, and social obligation under conditions where getting it wrong meant not passing on your genes. Emotion is not decoration it is regulation. Cognition is not cleverness, it is cost–benefit analysis under existential pressure.
Machines have none of this. No hunger. No mortality. No offspring. No reputational ruin. No social exile. No ancestral environment quietly shaping their trade-offs. A large language model does not fear anything, because there is nothing for it to lose. It does not want, because wanting is what you get when evolution needs a body to move toward one thing and away from another. What it has instead are loss functions, gradients, and an external optimisation process that rewards some outputs and punishes others. Calling this psychology is like calling a spreadsheet numerate or a thermostat neurotic. The metaphor might feel intuitive, but it does no explanatory work.
What tends to get waved around as evidence of machine psychology is behavioural consistency under certain conditions. Ask the system in a particular way, constrain it in a particular way, and you’ll see repeatable patterns. Fine. But behaviour alone has never been enough. Evolutionary psychology has spent decades reminding us that the interesting question is not what a system does, but why mechanisms evolved to do it. Humans don’t just behave, they behave for reasons that made sense in ancestral environments. Machines behave because they are engineered to minimise error across datasets. These are not neighbouring explanations. They only look adjacent because our language collapses them.
Anthropomorphism does the rest. Humans are agency-detecting machines long before they are rational ones. We see minds everywhere because, historically, failing to notice a real one was dangerous. We shout at laptops, apologise to furniture, and feel bad for robots that look lonely. Machine psychology flatters this instinct beautifully. It lets us tell stories instead of doing analysis. It turns optimisation artefacts into characters. And once you’ve done that, it becomes very hard not to start asking the wrong questions.
Philosophically, this is where Daniel Dennett is useful, and also where he is routinely misused. The intentional stance- treating a system as if it has beliefs and desires -can be a handy predictive shortcut. It works brilliantly with humans because humans actually do have evolved psychologies. Applied to machines, it is a convenience, not a discovery. The trouble begins when the shortcut is mistaken for an explanation, when ‘as if’ slides into ‘therefore is’, and a modelling stance becomes a metaphysical claim by stealth.
What evolutionary psychology insists on, and what machine psychology quietly discards, is motive architecture. Fear exists because death mattered. Jealousy exists because reproductive investment was fragile. Guilt exists because social cooperation needed repair mechanisms. These things are not poetic labels; they are solutions to recurring problems. Show me a machine that can starve, be ostracised, lose mating opportunities, or permanently fail its lineage, and we can start having an interesting conversation about psychology. Until then, we are watching surface behaviour without substrate and congratulating ourselves on the depth of our interpretation.
The reason this framing spreads so easily is not mysterious. It adds drama. It adds gravitas. It makes engineering feel like philosophy and optimisation feel like ethics. But it also quietly shifts responsibility. If machines have ‘psychology’, then their failures start to look like pathologies rather than design choices. Their misbehaviour becomes something they did, not something we built. Intent migrates from objective functions to fictional inner lives, and accountability leaves by the side door.
Psychology is what evolution builds when survival, reproduction, and social coordination are at stake. Machines do not face those stakes. They do not live, and they do not die. They execute. Calling that psychology does not expand our understanding of minds, it erodes the meaning of the term. What we are really seeing is not the birth of machine psychology, but the persistence of a very old human habit, mistaking fluency for interiority, and projection for insight. Be careful out there.