AI Accelerates Scientific Discovery Across Fields

A new report from OpenAI demonstrates that advanced AI models like GPT-5 are beginning to act as genuine collaborators in scientific research, accelerating discoveries across mathematics, physics, biology, and computer science. The paper, authored by a team of researchers from institutions including OpenAI, University of Oxford, and Lawrence Livermore National Laboratory, presents a collection of case studies where GPT-5 contributed concrete steps to ongoing projects, from proving new theorems to analyzing experimental data. These examples highlight both the potential and limitations of current AI, showing where human expertise remains essential and where AI can save significant time.

In mathematics, GPT-5 assisted in solving previously open problems, such as Erdős Problem #848, which concerns subsets of integers where certain products are not squarefree. The AI proposed a key idea that, combined with human insights from online comments, led to a complete proof. GPT-5 also helped locate existing solutions to 10 other Erdős problems that were mistakenly marked as open in an online database, performing deep literature searches that connected concepts across different fields. For instance, it found a reference from 1984 that solved a problem about entire functions, even though the connection required understanding subharmonic functions—a task challenging for traditional search engines.

In physics, GPT-5 rediscovered nontrivial Lie point symmetries of equations describing waves on rotating black holes, matching recent human-derived . Initially, the model failed when presented with the curved-space problem directly but succeeded after a warm-up on a simpler flat-space version, ultimately producing the correct symmetry generators that underpin explanations for vanishing tidal responses in black holes. This demonstrates AI's ability to handle complex analytical calculations, such as those involving partial differential equations with non-constant coefficients, which can accelerate theoretical work from months to days.

In biology, GPT-5 analyzed flow cytometry data from immune cell experiments, providing mechanistic insights that had eluded human experts. For example, it correctly inferred that a glucose analog called 2-deoxy-D-glucose was affecting T cell differentiation primarily by interfering with N-linked glycosylation rather than just inhibiting glycolysis. The model suggested follow-up experiments, including a mannose rescue test, which later validated its predictions. It also predicted that brief exposure to this compound could enhance the cytotoxicity of CAR-T cells against cancer, a finding confirmed in unpublished lab .

The report includes examples where GPT-5 generated entirely new proofs, such as for inequalities on subgraph counts in trees. One open problem from 2016 was resolved with a short, elegant proof based on a miraculous identity, differing from the earlier human proof that spanned four pages. In another case, GPT-5 solved a long-standing open problem from the 2012 Conference on Learning Theory, showing that a parameter in a preferential attachment tree model can be estimated from a single snapshot of the network by analyzing the fraction of leaves. This involved stochastic approximation techniques and yielded a computable estimator.

Despite these successes, the authors caution that GPT-5 has limitations: it can confidently make errors, defend incorrect arguments, and sometimes fail without proper scaffolding. For example, in a problem about clique-avoiding codes, GPT-5 initially produced a correct proof but did not attribute it to an existing paper, risking misattribution. The model also struggled with more open-ended queries, often providing overly optimistic or flawed responses. Human oversight remains crucial for verifying , guiding prompts, and integrating AI suggestions into broader research frameworks.

Of this work are profound for the future of science. As AI models improve, they could compress the timeline from hypothesis to , allowing researchers to explore more ideas rapidly and access interdisciplinary knowledge effortlessly. However, this also raises s around reproducibility and attribution, as seen in cases where AI rediscovered known without citation. The authors emphasize that AI is best used as a tool in tandem with human expertise, accelerating rather than replacing the creative and critical thinking that drives scientific progress.