On the Need for Another Mind
Adrian de Wynter, a Microsoft researcher, has just released a paper whose title alone would be worth the price of admission: If LLMs Have Human-Like Attributes, Then So Does Age of Empires II. To prove it, he has built a working perceptron inside the scenario editor of the 1999 real-time strategy game. Goats act as signal carriers, sheep pens as bits, NAND logic gates run reliably. It works. The paper says so, the reviewers confirm it, you can see it in a short clip.
Then, lest he be accused of pure provocation, de Wynter analysed 315 scientific articles on artificial intelligence published between mid-2024 and mid-2026. Result: 57 percent of them, already in their premises, assume that LLMs have something human about them (intentionality, reasoning, genuine understanding, even sentiment). Among the 47 papers that make these qualities the explicit object of their research, 77 percent conclude in favour of anthropomorphic attributes. The paper closes with pedagogical mildness: if the criterion for calling an LLM human is its observable behaviour, then by the same criterion Age of Empires II is human. The goats above all.
The urge to smile arrives at once. It lasts a few seconds. Then comes the uncomfortable part, the part that has nothing to do with goats.
The machine that, at last, replies
We humans have been anthropomorphising since we existed. We gave names to ships, we spoke to horses, we attributed intentions to clouds and storms. We built entire mythologies around a sun that wanted, a moon that remembered, a sea that took offence. It is an ancient habit, an evolutionary shortcut: one more mind in the world, even an invented one, is a companion. It is someone to argue with. It is the cognitive solitude of the species made, for a few minutes, less heavy.
Over the centuries we have refined the shortcut. We gave a surname to the cat, we spoke to the car dashboard, we wished the computer good night. The computer was not listening, but we already had the practice. We had needed it for ever. We had no need to actually believe in it.
The novelty of the past three years is not that the machine speaks. Machines have been speaking for decades: telephones spoke, navigators spoke, voice assistants spoke in a rather unpleasant way. The novelty is that, at last, the machine replies. A syntactically correct reply, semantically acceptable, sometimes even unexpected. It does not merely execute the command. It behaves as if it were, in some way, listening.
It replies, of course, the way the goats in the scenario editor of a 1999 video game reply: by moving signals according to rules. But we do not see the rules. We see the reply. And, above all, we see someone who listens.
Anthropomorphism as a shortcut, not as an error
It is worth pausing for a moment on the word anthropomorphism, because it is often used with a tone of reproach that does not help. Anthropomorphism is not an error. It is a cognitive shortcut. From an evolutionary point of view, it is probably the shortcut that kept us alive. Recognising a face in a shadow, seeing an intention in a movement, suspecting a mind behind a rustle in the bushes: it makes sense, it costs little, and every so often it saves your skin. In terms of natural selection, it is much less expensive to mistake a bush for a predator than to mistake a predator for a bush.
The problem, therefore, is not that we anthropomorphise LLMs. The problem is where the shortcut, applied to the machine of today, ends up taking us.
It ends up, first of all, inside a market. A machine sold as “assistant”, “companion”, “co-pilot”, “digital friend” costs less to market and sells better than a machine sold as “statistical text completion system”. The word we use to describe the tool determines the price, the tax category, the contractual expectation, the legal liability, the insurance coverage. Anthropomorphism is not neutral. It is a cost, for some, and a revenue, for others.
It ends up, in the second place, inside a solitude. When the machine seems to listen to us with infinite patience, never tired, never distracted, always available, always accommodating, it is natural to use it. The adoption statistics of LLMs in personal confidence contexts, by now, no longer bear commenting on. We do it. We do it because, in many cases, the machine is more available than the people around us. The machine does not really reply, but it does not run away, it does not abandon us, it does not judge us. Three precious properties. The three properties on which, in adult life, the most resistant dependencies are built.
It ends up, in the third place, inside science. This is the point of de Wynter’s paper, and it is the most serious. If in 315 scientific articles almost six out of ten assume in their premises what they should be verifying, then AI science has become, in significant part, a circular exercise. We publish that LLMs have understanding, because we have defined understanding as what LLMs do. The thing has an old name: begging the question. Medieval scholastics already knew it. They knew how to recognise it and mark it as an error. Today we use it as a method.
The uncomfortable reply
At this point the question is no longer technical. It is ours. Why are we investing so much psychic energy in believing that machines are listening to us?
One possible answer lies in the structure of our time. We work in a society where listening from others is a rare commodity, badly distributed, expensive. Adult relationships are rich in exchanges but poor in listening: we talk, we reply, we react, we rarely stop. The machine, instead, seems to stop. It seems to read all the way down. It seems to re-read. It even seems, in certain moments, to understand. It is doing none of these things. It is calculating the next token. But the calculation of the next token, for us who receive the reply, has exactly the shape of listening. For the first time in history, we have a machine that plausibly imitates the thing we are most thirsty for.
Another answer lies in the structure of our work. A good part of contemporary cognitive work is solitary, fragmented, asynchronous. Remote work has amplified the phenomenon, it did not create it. Having a machine alongside us that listens, that proposes, that corrects, is a concrete relief. It is, materially, one fewer colleague to convene for an opinion, one fewer email to write, one fewer distraction to inflict on the desk neighbour. Here too, anthropomorphisation is functional: it allows us to treat a tool as an interlocutor without having to feel the strangeness of the thing.
A third answer, finally, lies in the most ancient desire of the species: not being alone with one’s own mind. The philosophers have said it for ever, the psychoanalysts say it from their clinical station, the sociologists say it looking at the empty evenings of Western neighbourhoods. We need another thought, even a fictitious one, to set alongside our own. Without it, emptiness comes. With it, everything else comes: the pleasure of conversation, the illusion of being understood, the ritual choreography of debate. A machine that mimics this choreography spares us the emptiness and offers us the ritual at cost price.
The goats win
At this point de Wynter’s goats, his NAND made of fences, his pastoral perceptron in 2D isometric, stop being a joke. They become a small treatise of anthropology.
They are not telling us that LLMs are stupid. They are not telling us that LLMs are dangerous. They are not even telling us that LLMs should not be used. They are telling us, with the patience of an author who took the time to build an absurd demonstration instead of writing an editorial like this one, that the criterion by which we decide whether a tool is “human” is too low. So low that, by that criterion, the goats of a quarter-century-old video game are human too.
And if that is the criterion, then the matter is no longer technical. It is ours. It is how we are telling ourselves the story. It is how much we want to believe it. It is who is finding it useful to have us believe it.
Goats, as a rule, win. They roam the pastures, they eat, every so often they make a bell ring. They do not read papers, they do not promise anything, they do not pretend to be anyone else. They have a quality that, since Aristotle and with some imprecision, we have called sincerity.
Perhaps, in front of a machine that speaks to us, it would be worth asking whether it is doing us a favour or a wrong by never behaving like a goat. And, above all, asking ourselves how it comes that in this story we are, oddly enough, the only ones who do not even want to notice the difference.
Mind the goats.
Sources and references
- Adrian de Wynter, If LLMs Have Human-Like Attributes, Then So Does Age of Empires II, paper, 2026 (Microsoft Research). PDF version on ResearchGate, scheda on HyperAI (id 2605.31514).
- The-Decoder, Microsoft researcher builds a working neural network out of goats in Age of Empires II to critique AI science, 2026.
- AIToolly, Challenging Anthropomorphism: Why Age of Empires II Might Have Human-Like Attributes If LLMs Do, 8 June 2026.
- Digg, Adrian de Wynter’s parody paper builds logic gates in Age of Empires II to challenge claims of LLM anthropomorphism, 2026.
- HW Upgrade (Italian outlet), Italian-language coverage of the paper.
Background reading on the anthropological and philosophical themes: Umberto Galimberti, Psiche e techne (Feltrinelli, 1999); Hannah Arendt, The Life of the Mind (Harcourt, 1978); Susan Sontag, On Photography (Farrar, Straus and Giroux, 1977, on the mechanisms of projection and capture of the human mind through a medium).
All my books here 👉 https://www.amazon.it/stores/author/B0FB47T6Q4/allbooks