DeepMind Co-Founder Proposes 'Einstein Test' to Assess AI's Scientific Creativity

Demis Hassabis, Nobel laureate and co-founder of Google DeepMind, has proposed a new benchmark for AI scientific reasoning called the 'Einstein Test'. The concept would restrict an AI's knowledge to the year 1911, then evaluate whether it can independently develop a novel explanatory framework and perform reasoning beyond its training data, rather than simply repeating known answers.

Hassabis suggested that passing such a test could serve as a strong indicator of artificial general intelligence. The idea, reviewed in a recent Nature article, leverages known scientific history to compare model outputs against actual discoveries, allowing researchers to control knowledge boundaries and observe whether models acquire reasoning abilities that exceed their training input.

The article also highlights an attempt by independent researcher Michael Hla. In March 2026, Hla trained GPT-1900, a historical language model built from books and newspapers published before 1900, totaling about 22 billion tokens. The 3.3-billion-parameter transformer model was supplemented with 290 million tokens of historical physics literature. Hla aimed to see if the model could rediscover theories such as light quanta and special relativity. During the experiment, GPT-1900 produced a passage strikingly similar to Einstein's 1905 paper, but it failed most physics tasks. Hla acknowledged that modern AI models, including Claude, were used to generate instruction data and reinforcement learning feedback, meaning the model's isolation from modern knowledge was incomplete.

The work has fueled discussion on whether AI can generate genuinely new ideas. Google DeepMind researcher Tom Zahavy, in an opinion paper, categorizes scientific reasoning into induction, deduction, and abduction, arguing that current models excel at induction but lack Einstein-style abductive leaps. MIT computer scientist Jacob Andreas noted that producing a relativistic statement may not be the hardest part; the real challenge is determining which theories are correct or worth testing.