Einstein Test: Can AI Rediscover the Theory of Relativity?

In 1915, Albert Einstein published his theory of general relativity, permanently changing how scientists understand the structure of the physical universe. The theory describes gravity as the curvature of space-time caused by mass and energy. It remains a foundation of modern physics and cosmology, supporting research into black holes and gravitational waves while guiding technologies such as space missions and GPS satellites.

General relativity has also become a benchmark for researchers asking whether artificial intelligence can produce major scientific breakthroughs. At the India AI Summit in New Delhi in February this year, Demis Hassabis, co-founder of Google DeepMind in London, suggested training large language models (LLMs) only on information available before a fixed historical date, such as 1911. The challenge is to see whether AI can reproduce general relativity. “That would be a good test for AGI,” Hassabis said, referring to artificial general intelligence, a long-term goal for much of the AI industry.

Such experiments do not have to focus specifically on general relativity. In December 2024, Owain Evans, a researcher at the nonprofit Truthful AI in Berkeley, California, described the idea of a “vintage” or “historical” LLM trained only on data available before a selected cutoff date. Evans asked what scientific ideas such a model might independently rediscover.

Several research teams have built versions of these historical AI models this year, including systems designed to explore the Einstein challenge. Their early results, however, have revealed more about the limitations of current AI than about its ability to generate revolutionary scientific theories.

Scientific breakthroughs such as general relativity are not necessarily beyond the reach of AI, says Ido Kaminar, a quantum optics researcher at the Technion–Israel Institute of Technology in Haifa. He is a co-author of a preprint published in July titled “Can AI Follow in Einstein’s Footsteps?”¹.

However, Kaminar and his colleagues argue that AI systems will need to move beyond several principles underlying today’s large language models before they can make genuinely original scientific discoveries.

The creative leap

Hassabis first compared the challenge to Einstein’s achievement during a media interview last year. Tom Zahavi, another Google DeepMind researcher, expanded on the idea in an opinion article published on his website in January. Titled “LLMs Can’t Jump” (go.nature.com/3ykartg), the article argues that current LLMs cannot yet make the kind of conceptual leap associated with Einstein.

As philosophers of science have long observed, major advances often require abductive reasoning: a creative leap used to propose causes for unusual or unexplained phenomena. This differs from inductive reasoning, which derives general rules from a large collection of examples and observations.

Abductive reasoning is not a clear strength of today’s AI models. Current systems are generally better at performing the labor-intensive tasks of science than at generating entirely new theories. They are trained on known examples and identify statistical patterns in large datasets, producing the answers most likely to fit a prompt.

‘It’s incredible’: How AI will transform mathematics

This makes AI valuable for prediction and analysis, particularly when researchers have access to large volumes of reliable data. But imaginative breakthroughs — what 20th-century historian of science Thomas Kuhn called paradigm shifts — often emerge when data are limited or when observations do not fit established explanations.

In such situations, scientists commonly develop a “world model”: an explanation of the underlying principles based on a relatively small amount of evidence. For example, Johannes Kepler’s precise observations of planetary motion and the mathematical relationships he identified helped Isaac Newton formulate his laws of motion and universal gravitation.

Can AI models make a similar transition from sparse data to a general theory of the world? Sendhil Mullainathan, a computer scientist at the Massachusetts Institute of Technology (MIT) in Cambridge, Massachusetts, says current systems struggle to do so. In a conference paper published in July,² Mullainathan and his colleagues tested a basic “orbital mechanics” model using synthetic data from planetary systems governed by Newtonian mechanics.

The model did not infer the correct law of gravity, which relates the force between two objects to their masses and distance. Instead, it generated different laws for individual planetary systems, each of which was wrong in its own way.

Nevertheless, recent advances in AI mathematics suggest that models may be able to identify some of the underlying logical structures in their training data. In May, an AI chatbot developed by OpenAI produced what Mullainathan described as a “true conceptual advance” by disproving a prediction made 80 years earlier by Hungarian mathematician Paul Erdős.

The system did not solve the problem solely through brute-force computation, Mullainathan says. Its solution involved developing several small abstractions along the way. However, the result also combined ideas that were already present in the mathematical literature. In that sense, it was not an Einstein-like discovery that appeared to emerge from nowhere.

A glimpse of intuition

After Hassabis proposed the 1911 cutoff test, Michael Fuller, an independent AI researcher in San Francisco, California, attempted a related experiment. Writing on his website in March, Fuller said he trained an LLM called Machina Mirabilis on data published before 1900. He wanted to see whether the model could generate quantum mechanics, Einstein’s special theory of relativity from 1905, or general relativity.

AI is not ready to study itself

Fuller’s experiment produced some useful suggestions. In one example, the model generated an observation about the photoelectric effect — the interaction between light and electrons in a metal — that was later explained by Einstein’s idea of light quantization. In another, it produced the essence of the “elevator” thought experiment that helped Einstein develop general relativity.

Fuller says the model occasionally displayed “glimmers of intuition.” When asked about the photoelectric effect, for instance, it described light as dividing into many separate impulses, suggesting the concept of light quanta. Yet the LLM usually failed to demonstrate a genuine understanding of the physics behind its answers. It sometimes produced plausible-sounding language without any strong internal model of the physical world from which to reason, Fuller writes.

Training historically restricted LLMs was also technically difficult because the source material contained broken English, inconsistent terminology and errors introduced during the digitization of printed documents.

Independent computer scientist Nick Levine and his colleagues encountered similar problems in a study published online in April (go.nature.com/4ilz5sp). They attempted to build a historical AI model using only information available before 1930. The cutoff was selected because works published that year entered the public domain in the United States in early 2026.

In principle, such a model could be used to test questions about scientific developments from the 1930s. Possible examples include the Turing machine, a foundational idea in computer science; Gödel’s incompleteness theorems in mathematics; and the neutrino, a particle first proposed in print in 1934.

Levine and his colleagues quickly discovered how difficult it was to create a model based exclusively on knowledge available before 1930. Historical data are often contaminated by later information. “When you ask people what happened in the 1950s, they often answer by chance,” Levine says, although the model may sometimes provide the correct answer. For example, the pre-1930 model answered questions about Franklin D. Roosevelt, who became US president in 1933 and served until 1945. If dates in a dataset are missing or unreliable, filtering out later information becomes extremely difficult.

Even so, Levine, who previously worked as a quantitative economic forecaster, believes that researchers may eventually be able to ask a “spirit of 1930” model to make historically constrained predictions, including forecasts similar to those used in prediction markets.

Historical AI models could also be used to predict scientific discoveries, he says. One possible experiment would train a model on information available up to January and test whether it could reproduce findings that researchers made in June. Levine’s expectation is that the model may identify some important discoveries simply because of the scale of the data and computational resources involved.

Researchers at the University of Zurich in Switzerland have created a family of historical language models through a project called Ranke-4B (go.nature.com/46efotz). These models were trained on time-stamped texts with historical cutoffs in 1913, 1929, 1933, 1939 and 1946. Team member Daniel Gächter, an economist now at ETH Zurich, says the project is not intended to build models as powerful as today’s leading LLMs. Instead, the researchers want to determine whether historical models can generate ideas that resemble later scientific progress. “We’re testing, figuratively speaking, the brilliance of genius rather than genius,” he says.

Too many theories

MIT computer scientist Jacob Andreas argues that there is no obvious technical barrier preventing an AI language model from generating something that resembles the theory of general relativity. The problem is that such a theory would be only one probabilistic output among countless incorrect explanations for the same body of data.

Distinguishing a correct theory — or even a promising theory worth testing — from an incorrect one remains extremely difficult. This was evident in the orbital-mechanics experiments conducted by Mullainathan and his colleagues, which produced numerous false “laws of gravity”.

Mathematics offers an important advantage because researchers can often verify quickly whether each step in a proposed argument is valid. Physical theories are harder to evaluate: they must be consistent with observations, mathematically coherent and capable of making accurate predictions.

Why AI systems are most useful as designers of new scientific tools

This highlights a fundamental difference between LLMs and human scientific reasoning. Humans rarely generate an unlimited collection of probabilistic theories and then determine which, if any, is correct. Instead, scientists use physical intuition, prior knowledge and experimental judgment to narrow the possibilities.