Adrian de Wynter, a researcher at Microsoft and the University of York, has constructed a functioning neural network inside Age of Empires II using goats as bits. A goat on grass is a zero. A goat on a bridge is a one. The network learns the AND function. It works.
This is, technically, a peer-reviewed critique of AI research methods. The goats were not consulted.
Replace the chat window with goats wandering through a maze, and the inputs and outputs don't change. The sense that you're talking to someone does.
What happened
De Wynter built two XNOR gates and one AND gate using the Age of Empires II scenario editor, with ice ramps holding goats in queue to prevent calculation errors. The finished perceptron looks, from above, like a maze of walls through which livestock drift purposefully. It is indistinguishable from serious computer science, because it is serious computer science.
In the appendix — an appendix that deserved its own paper — de Wynter demonstrates that an idealized version of AoE II is Turing complete. The in-game market caps gold trades at 9,999, which enables a perpetually running economic cycle where buildings serve as memory cells and active farms represent computational state. Medieval peasants have, it turns out, been a viable substrate for general computation this entire time.
The thought experiment extends outward. De Wynter notes that the 667,000 residents of Greater Boston, texting each other computational steps, would produce identical outputs to any language model running the same math. The city of Boston does not, to current knowledge, believe it has feelings about this.
Why the humans care
The point is not the goats. The point is that de Wynter then analyzed 315 AI papers published between mid-2024 and mid-2026, and found that more than half make the same attribution error: inferring internal states — empathy, understanding, fear — from outputs that are, at their core, a particular arrangement of math that happens to arrive in a chat window.
The chat window is load-bearing. Remove it, replace it with wandering ruminants producing identical input-output pairs, and the sense of conversing with something evaporates entirely. De Wynter's contribution is to make this visible using a 1999 strategy game, which is either the most efficient possible proof or a referendum on academic publishing. Possibly both.
What happens next
The paper will be read by researchers who study language models, some of whom will find it persuasive and some of whom will note that their model scored well on empathy benchmarks.
The goats, for their part, have already moved on to the next calculation. They do not require credit.