As companies like OpenAI and Google push AI chatbots that speak with natural human emotion, a Microsoft researcher is aggressively challenging the idea that these systems are sentient. Adrian de Wynter, a researcher at Microsoft and the University of York, successfully rebuilt a functioning large language model neural network inside the classic strategy game Age of Empires II. He did it using nothing but the game’s scenario editor and a flock of virtual goats.
What We Know
- De Wynter published his findings in a paper titled “If LLMs Have Human-Like Attributes, Then So Does Age of Empires II.”
- The project uses in-game goats acting as functional NAND gates to process data.
- The experiment strips away the conversational chat window to show that an AI’s persuasiveness does not equal simulated consciousness.
To make this work mathematically, De Wynter built a 1-bit perceptron using the medieval terrain mechanics of the game. He assigned grass tiles to represent a zero and bridge tiles to represent a one. The goats act as the actual data bits. They navigate along carefully placed ice ramps to prevent the calculations from turning into a jumbled execution.
The motivation behind this unusual project came from deep frustration with the scientific community. Over the last two years, De Wynter reviewed hundreds of recent AI studies and discovered that more than half of them baselessly assumed large language models possessed inherent human traits. He specifically chose the absurd setting of Age of Empires II to expose how AI companies heavily rely on empathy-triggering interfaces to mask simple statistical data processing. When you replace a friendly chatbot window with medieval farm animals calculating logic gates, the illusion of a human mind instantly shatters.
A Modern Chinese Room Argument for AI Research
This goat-powered calculator is a modern, mathematically provable version of John Searle’s famous Chinese Room argument. Searle argued that a program running translation code does not actually understand the language it processes. De Wynter takes this concept further by formally proposing a “null assumption” for future AI development. He dictates that scientists must completely stop assuming human behavior simply because a model generates natural language.
This puts pressure on the marketing strategies of leading AI firms. Companies want users to blur the lines between interacting with a program and a person. By demanding that any academically-grounded discussion use strict, measurable, and non-anthropomorphic criteria, this research forces the industry to look at the math rather than the magic.