Scientific discoveries often start when someone notices something unexpected. Can artificial intelligence (AI) systems, specifically large language models (LLMs) that generate text, do this on their own without being told what to investigate?
Researchers built simulated worlds with altered physics, like heavier objects falling more slowly. They let twelve LLMs explore freely, ask questions, design experiments, and build tools. Other AI agents automatically checked if the LLMs correctly found the hidden laws.
Only one LLM, GPT-6-Astra, succeeded, and only 16.7% of the time. The other eleven never discovered any law. This shows that independent scientific curiosity remains extremely hard for current AI, even in simple simulated worlds.