Adrian de Wynter, an AI researcher at Microsoft, got tired of an argument and decided to win it with goats. Inside the scenario editor of Age of Empires II, the 1999 strategy game, he built a working neural network out of game objects. Grass stands in for a binary zero, bridges for a one, and goats play the part of the bits moving through the system. The result is a functioning NOT-AND gate and a one-bit perceptron, the simplest unit of a neural network, running entirely on medieval livestock. He wrote it up in a paper, reported by 404 Media, with the deadpan title "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II."
The title is the whole argument compressed into a joke. "I have this tendency to dial up things to 11 when I really think I need to make a point," de Wynter said, noting that "absurdism is pretty standard in philosophy and theoretical computer science." The processes shuffling his goats around are, at the most basic level, the same kind of operations that power tools like ChatGPT and Claude. Because they are dressed in grass and goats rather than fluent English, nobody watching is tempted to call the output thoughtful, conscious or human. That is exactly the point he wants to land.
Behind the stunt is a serious complaint about how the field talks to itself. De Wynter says he has peer-reviewed more than 300 computer science papers over the past two years, and that over half of them opened by simply assuming large language models possess human-like traits. When an unproven assumption becomes the starting line for that much research, the conclusions built on top of it inherit the wobble. A model that produces fluent sentences is easy to mistake for a model that understands them, and the mistake quietly compounds.
Why the illusion is so sticky is not really a mystery. Humans are primed to read minds into almost anything, from weather to pets to cartoon shapes, and a system trained on natural language pushes hard on that reflex. AI companies have not always helped, equivocating over whether their products might be edging toward sentience rather than firmly saying no. Stripping the language away, and replacing it with goats, removes the thing our brains keep tripping over.
De Wynter's prescription is modest and sensible. He wants researchers to separate what a model actually is, a set of relationships between numerical weights manipulated by some operation, from what it is perceived to be. "We should perform experiments that allow us to see LLMs as how they are, not how we believe they should be," he said. In a news cycle thick with claims about what AI wants, fears or intends, a herd of digital goats turns out to be a surprisingly clarifying way to ask everyone to slow down.