This article was featured in the June 26, 2026, edition of the IBM Think newsletter. Get it in your inbox.
Humans have always been good at finding life where none exists. We see faces in random objects. We hear our names called out in the woods. Large language models seem to have supercharged that tendency, convincing some people that anything capable of talking like a person must also think like one.
Adrian de Wynter was getting a little tired of all this LLM anthropomorphism. The Microsoft AI researcher had peer-reviewed hundreds of computer science papers over the past two years and found that 57% started with the assumption that LLMs have human-like traits baked right in even before a single experiment was run. And when a paper specifically set out to study those traits, 77% concluded they exist. Turns out if you assume something is true before you start, your experiment has a funny way of proving you right.
To illustrate the problem, de Wynter did what any reasonable researcher would do. He took a decades-old computer game and used it to build a neural network out of goats.
As he chronicles in his recent paper “If LLMs Have Human-Like Attributes, Then So Does Age of Empires II,” de Wynter used the real-time strategy classic’s scenario editor—a sandbox mode that lets players build their own maps using the game’s assets—to wire together logic gates (basic on/off switches) where goats carry the bits. A goat on grass is a 0. A goat on a bridge is a 1.
“Only one rail is active at a time, with a goat acting as the signal carrier,” he wrote on GitHub. “When the gate fires, the bit-goats are removed (they ded) and a new bit-goat is placed in its respective output rail.”
Rather than arguing with the literature head-on, de Wynter applies the same anthropomorphic language often used to describe AI systems to his goat-powered computer. Suddenly, goats can appear “agentic.” They cooperate. They pass information. They exhibit behaviors that sound suspiciously intelligent when described in the right tech jargon. Yet all they’re really doing is implementing logic gates—the same basic computational building blocks that ultimately underpin modern AI systems.
“I have this tendency to dial up things to 11 when I really think I need to make a point,” de Wynter told 404 Media, adding that “absurdism is pretty standard in philosophy and theoretical computer science.” So if “agentic” goats carrying information across a medieval battlefield seem a little silly, well, then maybe words like “intention,” “cooperation” and “reasoning” deserve a little more scrutiny when they’re applied to language models, too.