Microsoft Researcher Uses 57% Paper Bias, Goat Network to Challenge LLM Anthropomorphism
Updated
Updated · IBM · Jun 26
Microsoft Researcher Uses 57% Paper Bias, Goat Network to Challenge LLM Anthropomorphism
2 articles · Updated · IBM · Jun 26
Summary
Adrian de Wynter built a goat-powered neural network inside Age of Empires II, using the game’s scenario editor to show how human-like language can make simple computation sound intelligent.
57% of the AI papers he peer-reviewed over two years assumed LLMs had human-like traits before testing, and 77% of papers explicitly studying those traits concluded they existed.
Goats served as binary signals in logic gates—a goat on grass meant 0, on a bridge meant 1—letting de Wynter describe basic computation with terms like “agentic,” “cooperate” and “pass information.”
The paper argues that if such language can make medieval game goats seem intentional, terms like “reasoning” and “intention” applied to language models also deserve closer scrutiny.