Do AI models like ChatGPT really think like humans? A bold experiment says no

A new study challenges the common belief that large-language models (LLMs) like ChatGPT possess human-like qualities. Microsoft AI researcher Adrian de Wynter argues that people too often assume these systems have consciousness or human traits without evidence. His research highlights a growing trend in computer science papers that make such assumptions without proof. De Wynter’s paper examines recent academic work on LLMs. Over half of the papers he reviewed in the last two years assumed these models had human-like capabilities. To demonstrate the flaw in this thinking, he built an LLM inside the game Age of Empires 2 using goats, grass, and bridges.

The system worked like a simple computer. Goats and other in-game objects represented binary logic, forming a NOT AND gate and a 1-bit perceptron. The processes were similar to those in advanced LLMs, but the unusual setup made it clear the system lacked human-like understanding.

He argues that the idea of sentient LLMs persists because humans naturally project human qualities onto non-human things. AI companies also contribute by avoiding clear statements on the issue. De Wynter calls for more experiments to understand LLMs as they are, rather than assuming they mimic human behaviour. The study urges researchers to stop attributing human traits to LLMs without proof. It shows that even simple, non-human systems can perform tasks resembling AI. The hope is to shift the focus toward understanding these models based on their actual mechanisms.