Goated: Large language models and chatbots can understand natural language and generate human-like responses, but there is no actual humanity behind them. According to a Microsoft researcher working on AI, the idea that LLMs possess human-like attributes is fundamentally flawed. And he has goats to prove his point.

Adrian de Wynter is an AI scientist at Microsoft and a researcher at the University of York. In addition to studying the fundamental challenges of AI and related technologies, he is also a longtime gamer who has been playing Age of Empires II since 1999. In a recent study, de Wynter combined both passions to demonstrate that modern AI systems are far less intelligent and capable of reasoning than many people assume.

The study's title is self-explanatory: "If LLMs Have Human-Like Attributes, Then So Does Age of Empires II." According to de Wynter, many researchers working on LLMs and LLM-powered agentic systems assume that these technologies possess some form of anthropomorphic qualities. Some even argue that chatbots can exhibit moral reasoning or have a conscious understanding of the meaning behind natural-language communication.

De Wynter said he is not trying to argue whether these qualities exist. Rather, his goal is to show that such assumptions are fundamentally flawed. To make his point, the researcher used the scenario editor in Age of Empires II to replicate the basic functions of a simple neural network.

In the custom scenario, grass tiles represent a value of "0," while bridges represent "1." Goats act as bits, carrying digital signals between the two states. The study's GitHub page provides a detailed explanation of how the goat-powered neural network works.

The researcher said that Age of Empires II contains all the tools needed to replicate the most basic functions required by ChatGPT and other complex AI models to perform semantic tasks. Using Microsoft's classic RTS is central to the study because it highlights the absurdity of treating neural networks as human-like entities.

The study argues that the supposedly anthropomorphic features of LLMs are "empirically non-unique" because they can be replicated with digital goats in a strategy game from the 90s. According to de Wynter, many studies interpret the behavior exhibited by chatbots as evidence of uniquely human traits when they should instead assume that other neural network systems can display the same characteristics.

Speaking with 404 Media, de Wynter said people should stop assuming that LLMs exhibit human-like behavior simply because their computational models have been trained on natural-language data. While many AI companies are actively working to develop some form of artificial general intelligence, de Wynter argues that people should not become emotionally attached to prompt-based expert systems any more than they would to a toaster.