The Einstein Test for AI: Can Language Models Rediscover Relativity, and How Is It Different from AI Math Results?

Demis Hassabis proposed training an AI only on pre-1911 knowledge to see if it could derive general relativity. What experiments with vintage models and world-model probes have found, why AI math results such as GPT-6 Astra's are a different kind of claim, and how to evaluate AI discovery.

PUBLISHED // UPDATED
DURATION // 07 MIN READ
Desk with a printed 1911 physics paper beside a laptop showing a sparse loss curve, cool overhead light, no logos
FIG.01 // VISUAL DOCUMENTATION Desk with a printed 1911 physics paper beside a laptop showing a sparse loss curve, cool overhead light, no logos

Can an AI system make a conceptual leap like Einstein’s? Google DeepMind CEO Demis Hassabis has proposed exactly that as a test for artificial general intelligence: train a model only on knowledge available up to about 1911 and see whether it can come up with general relativity, as Einstein did in 1915. In September 2026, a Nature feature examined what researchers have learned from trying versions of this idea, while AI labs announced results on long-standing math problems. These are two different kinds of evidence about AI and discovery. This article explains the Einstein test, summarises the research behind it, and sets out how to read AI discovery claims.

What the Einstein test is

Hassabis described the idea at the India AI Summit in February 2026, as reported by OfficeChai. He defines AGI as a system that can exhibit all the cognitive capabilities humans can, and said current systems fall short on true creativity, continual learning and long-term planning. He also pointed to “jagged” performance: systems can reach gold-medal level on International Mathematical Olympiad questions yet fail simple problems posed in a certain way.

His proposed test: “training an AI system with a knowledge cutoff of, say, 1911, and then seeing if it could come up with general relativity, like Einstein did in 1915.” He said “today’s systems couldn’t do that,” but that he expects it to become possible eventually.

The test is demanding because general relativity was not an interpolation of existing data. It required new principles, such as the equivalence of gravity and acceleration, and new mathematics applied to physics, guided by a small number of anomalies rather than large datasets.

What the research shows so far

Nature published a feature by Philip Ball in September 2026 (with a correction on September 14) titled “The Einstein test: what happens when AI tries to rediscover relativity?” The full article is behind a paywall, but the works it cites are publicly available and point in a consistent direction.

Vintage models

Researchers have started training “vintage” language models on text from before a cutoff date. Independent researcher Michael Hla described an experiment in March 2026 in which an LLM trained from scratch on text from before 1900 was tested on whether it could arrive at quantum mechanics and relativity. He reports that the model fails at most physics tasks, but shows “glimpses of intuition”: given experimental observations, it could state that light comes in definite quantities of energy and suggest that gravity and acceleration are locally equivalent. These are interesting hints from a small, single-researcher experiment, not a demonstration that a model derived the theory.

Prediction without understanding

A study by Keyon Vafa, Peter Chang, Ashesh Rambachan and Sendhil Mullainathan, presented at ICML 2025, tested whether foundation models trained on sequences actually learn the underlying “world model.” The authors developed an “inductive bias probe” that checks how a model adapts to new tasks generated from a known law. They found that models trained on orbital trajectories can predict well yet “consistently fail to apply Newtonian mechanics when adapted to new physics tasks,” behaving as though they had learned task-specific heuristics instead.

This is directly relevant to the Einstein test. Kepler’s accurate descriptions of planetary motion later led to Newton’s laws; the study suggests that today’s foundation models can match the Kepler step (accurate prediction) without making the Newton step (finding the general law).

Prediction versus theory-building

A July 2026 preprint by Michael Shalyt, Nathan Regev, Marin Soljačić and Ido Kaminer, “Can AI Follow In Einstein’s Footsteps?”, argues that AI’s contributions to physics appear to mirror the historical development of physics in reverse. Human physics moved from pattern prediction, through phenomenological laws such as Kepler’s, to principle-based theories such as relativity. AI went from early equation-discovery methods such as symbolic regression to powerful predictors such as AlphaFold and GraphCast, which are highly accurate but do not provide clear theoretical understanding.

The authors warn that if this trend continues, AI may become extraordinarily good at prediction but “struggle to ever propose its first serious contender to quantum gravity or other paradigm-level theories.” They identify a missing skill: the ability to pose the right questions or invent principles that guide new theories and tests to falsify them.

AI math results are a different kind of claim

In the same week, Axios reported on AI and mathematics:

  • OpenAI said GPT-6 Astra resolved or made “substantial progress” on decades-old problems in areas including high-dimensional geometry, coding theory, group theory, quantum complexity, lattice cryptography and combinatorics.
  • An unreleased research version of Anthropic’s Claude made what Axios called a major improvement on decades-old work related to the Riemann hypothesis, after spending 31 million output tokens and a day and a half coordinating about 60 subagents.
  • Anthropic employees had asked Claude to “take a real stab” at the problem and “believe in yourself” after an initial run failed.
  • MIT mathematician Andrew Sutherland said Claude’s work showed AI is “capable of doing interesting mathematical research,” and Terence Tao predicted mathematics will move “from an era of proof scarcity to an era of proof abundance.”

These results may be genuinely valuable, but they are not the Einstein test. The models were trained on modern mathematics, including the tools developed after 1915, and they worked within established frameworks. Improving a bound or proving a result in a known area is different from inventing a new framework from sparse anomalies. Claims made in company announcements also need independent verification: a published proof checked by mathematicians carries more weight than a description of progress.

Einstein testAI math results
Training dataRestricted to before the discoveryModern, including relevant prior work
GoalInvent a new theoretical frameworkSolve or improve results within existing frameworks
Evidence so farHints, mostly negative resultsCompany-reported progress, some endorsed by mathematicians
How to verifyCheck for data leakage and genuine derivationPeer review and formal proof checking

For background on the model behind OpenAI’s claims, see our coverage of GPT-6 Astra’s safety and hidden reasoning.

How to evaluate AI discovery claims

Whether you build evaluations or read announcements, three questions help:

  1. Was the knowledge cutoff clean? For historical tests, any leakage of later texts invalidates the result. For modern claims, check whether the “new” result already exists in the literature.
  2. Who verified it? A result accepted by independent experts, or formally checked, is stronger than a company blog post.
  3. Did the model find a new representation or search a known space? Searching efficiently within a known framework is useful but different from inventing a framework.

The same discipline applies to other benchmarks. Speed claims, like those for Inception’s Mercury 2.5, can be measured directly. Coding benchmarks, as we discussed for Cognition’s SWE-2, need care about who ran them. Discovery claims are the hardest to measure, which makes clear definitions even more important.

Our analysis

The Einstein test is useful because it separates two abilities that are easy to conflate: predicting well within known patterns, and inventing the principles that explain them. Current evidence suggests AI is advancing quickly on the first and, in mathematics, is starting to contribute real results within established fields. Evidence for the second remains thin, limited to small experiments with intriguing hints. Both findings can be true at once. The practical lesson is to judge each claim by what it actually tested, rather than treating progress on math problems as evidence that a model could have discovered relativity.