What Has AI Actually Achieved in Science?
Recent achievements in mathematics, physics, chemistry, and biology — as of October 27, 2026.
AI is contributing to science in several distinct ways: proposing new proofs, checking existing mathematics, helping physicists control experiments, designing molecules, and finding patterns in biological data. The examples below are concrete results, not predictions. They also show why a laboratory result, a computer-checked proof, and a proposed new discovery are not the same thing.
Many of these achievements were made by scientists working with AI, not by AI alone. Researchers chose problems, developed methods, checked results, and carried out experiments. The newest announcements have had little time for independent scrutiny.
Mathematics
1. Computer-checking the “enormous theorem” of symmetry
Axiom Math · October 2026
The classification of finite simple groups is a map of the basic building blocks of finite symmetries. Its human-written proof grew over generations to roughly 15,000 pages across more than 500 papers.
Axiom reports that AI agents, directed by mathematicians and engineers, translated the first-generation proof into Lean, software that checks formal mathematical reasoning. The team says this process exposed problems in the published arguments, including significant issues that required repair.
What was achieved: A claimed complete, computer-checked formalization of an existing human proof, with corrections and new supporting mathematics—not the original discovery of the classification.
What remains: The announcement is very recent. Experts need to scrutinize whether the formal statement captures the intended theorem and examine the repairs.
Research announcement · X post
2. A proposed proof about the emergence of connected networks
Justin Leder and Claude · 2026
Percolation theory studies what happens as connections form randomly in a network. A long-standing problem asks whether an infinitely connected cluster exists exactly at the critical threshold where such networks begin to form.
Under Leder’s direction, Claude generated a proposed proof and a Lean formalization, using an earlier reduction by mathematicians Gady Kozma and Shahaf Nitzan. The claimed result covers cases that had remained open, including three dimensions.
What was achieved: A claimed new proof with a machine-checked formal derivation.
What remains: Conventional independent mathematical refereeing and scrutiny of whether the formal definitions and the link to the earlier reduction capture the intended problem.
Explanation and review caveat · Formal proof · X post
3. A proof of an Erdős number-theory problem
GPT-5.2 Pro, Harmonic’s Aristotle, and mathematicians · 2026
A research preprint describes a solution to Erdős problem #728, a question involving factorial divisibility and logarithmic gaps. It attributes the Lean-formalized proof to GPT-5.2 Pro and Aristotle, operated by Kevin Barreto. Mathematician Nat Sothanaphan translated the formal argument into a human-readable proof.
What was achieved: A claimed proof of a specific open mathematics problem, with formal checking and human explanation.
What remains: The cited write-up is a preprint; journal peer review is not established here.
4. New mathematical constructions through AI-guided search
AlphaEvolve and collaborators · 2025
Researchers applied AlphaEvolve, a system that uses AI to propose candidate solutions and an evaluation process to improve them, to 67 mathematical problems. The system rediscovered best-known solutions in many cases and improved constructions in several.
What was achieved: New or improved constructions for some problems, alongside recovery of known results.
What remains: This is not a claim that 67 open problems were solved. An improved construction, a rediscovered result, and a complete proof are different outcomes.
5. Four International Mathematical Olympiad problems solved
Google DeepMind · 2024
AlphaProof and AlphaGeometry 2 solved four of six problems from the International Mathematical Olympiad, earning 28 of 42 points—the silver-medal threshold for that competition.
What was achieved: A strong performance on difficult, previously unseen competition problems.
What remains: This was a benchmark, not the discovery of a new theorem. Humans translated the problems into formal language, and some solutions took up to three days—longer than the two 4.5-hour sessions available to contestants.
Physics and quantum information
6. New results on five quantum-information problems
Physical Superintelligence · October 2026
Quantum information theory studies how quantum properties can be used to store, transmit, and process information. PSI researchers report working with an AI research system on five problems involving entanglement, measurements, and causality. They released papers, Lean formalizations, and verification code.
One result concerns when separated regions of a quantum vacuum cannot violate a particular Bell inequality, even though they remain entangled. Another gives a counterexample to a proposed relationship between causal and reversible quantum processes.
What was achieved: Five reported theoretical results, with formalizations and code.
What remains: These are very recent and need expert scrutiny. At least one is a partial result: the question concerning arbitrarily many copies of certain entangled states remains open.
7. Calculating previously uncomputed quantities in particle physics
Matthew Schwartz and Claude · October 2026
Physicists use mathematical quantities called scattering amplitudes to relate theories of particles to the outcomes of collisions. Calculating them can require difficult integrals—mathematical operations that combine quantities across a range of possibilities.
Physicist Matthew Schwartz reports that he and Claude built a toolkit called BootLoops and completed 30 such integrals: 15 reproducing known results and 15 that had not previously been computed.
What was achieved: New reported calculations, plus reproductions of known calculations by the toolkit’s method.
What remains: This is a researcher’s account of work in progress, not the discovery of a new particle or law of physics. Schwartz describes substantial human direction and checking.
8. Anticipating instabilities in fusion plasma
DIII-D tokamak researchers · 2024
A tokamak confines extremely hot plasma with magnetic fields in an effort to study nuclear fusion. The plasma can develop instabilities that disrupt the experiment.
Researchers used reinforcement learning—a method in which a system learns actions through feedback—to anticipate and avoid a type of instability called tearing.
What was achieved: An experimentally tested AI-assisted control technique for a fusion experiment.
What remains: This is not a new law of physics or proof of a commercially viable fusion power plant.
Chemistry
9. Designing an enzyme for a step in HIV-drug synthesis
University of Washington, NVIDIA, and University of Manchester · 2026
An enzyme is a protein that helps a chemical reaction happen. Researchers used the RFD4-Proteina AI design system to propose enzymes for a reaction relevant to making lenacapavir, an HIV drug. Manchester researchers then made and tested the designs.
Four of 67 designs were active. The best converted 79% of the starting material under the reported laboratory conditions.
What was achieved: A physically tested, AI-designed enzyme for a pharmaceutical chemistry step.
What remains: This is a laboratory proof of concept, not evidence that the enzyme is used in commercial drug manufacturing.
University of Manchester · Research preprint
10. Creating new enzymes that perform a chemical reaction
RFdiffusion2 researchers · 2025
Researchers used AI protein design to create new zinc-dependent enzymes and tested them on a model chemical reaction. Four designs were highly active. An experimentally determined structure of the best design closely matched what the researchers had intended to build.
What was achieved: New designed proteins that performed a measured chemical reaction in the lab.
What remains: Success on this laboratory substrate does not establish industrial performance or demonstrate that the enzymes can clean up pollution.
11. Designing antibiotic candidates that can be synthesized
SyntheMol researchers · 2024
SyntheMol generated candidate molecules with an emphasis on designs that chemists could actually make. Researchers synthesized 58 candidates, and six showed antibacterial activity in laboratory tests, including against Acinetobacter baumannii, a bacterium associated with serious hospital infections.
What was achieved: New AI-designed molecules with experimentally observed antibacterial activity.
What remains: Activity in a laboratory test is not evidence of safety or effectiveness in patients. These are candidates, not approved medicines.
Nature Machine Intelligence paper
Biology
12. Designing functional proteins from written descriptions
Pinal researchers · October 2026
Pinal generates protein designs from natural-language descriptions of desired functions. Researchers report observing the intended function in four experimental design tasks, including two involving enzymes.
In one test, a designed protein increased product yield 1.7-fold compared with the natural comparator in a multi-enzyme pathway for fixing carbon dioxide.
What was achieved: Experimentally tested proteins generated from descriptions of their intended roles.
What remains: Results on these specific tasks do not show that an arbitrary requested protein will work or that the carbon-fixation pathway is commercially viable.
13. Finding a genetic signal in billions of mutation pairs
Matthew Schwartz, collaborators, and Claude · October 2026
Schwartz and collaborators report analyzing 5.7 billion pairs of nearby genetic mutations using data from the 1000 Genomes Project. They found evidence for gene conversion, a process that can copy genetic sequence information from one DNA segment to another.
The research direction changed after geneticist Michael Desai suggested studying relationships between pairs of mutations rather than stopping at the AI system’s initial result.
What was achieved: A reported biological finding from analysis of a large public genomic dataset.
What remains: This is ongoing research, not a clinical test or treatment. Its interpretation and replication matter.
14. Measuring what a chance-based model misses in a forest
Matthew Schwartz, James O’Dwyer, and Claude · October 2026
Neutral biodiversity theory asks how much of the rise and fall of species can be explained by chance, without assuming that one species has an inherent advantage over another.
Schwartz reports that Claude helped make a long-standing equation practical to apply to data from Barro Colorado Island in Panama. The analysis found that the mix of tree species changed 4.5 times faster than the neutral model predicted. Ecologist James O’Dwyer then helped redirect the work toward studying the differences the baseline model left unexplained.
What was achieved: A reported large-scale ecological calculation and a follow-on model developed with domain expertise.
What remains: The team is extending the work to other forests. The initial finding alone was not necessarily surprising to ecologists.
15. Proposing biomedical hypotheses for laboratory tests
Google AI co-scientist and research collaborators · 2026
The AI co-scientist system proposed drug-repurposing candidates and combinations for acute myeloid leukemia that researchers tested in cells. It also proposed liver-fibrosis targets tested in human liver organoids—small, lab-grown structures that model some features of an organ.
The system additionally reproduced an explanation of bacterial gene transfer that another team had independently discovered but had not yet published.
What was achieved: AI-generated biomedical hypotheses followed by laboratory testing, and independent recapitulation of a previously unpublished explanation.
What remains: Cell and organoid findings do not establish clinical efficacy. The bacterial explanation was not an exclusive AI discovery.
16. Watermarking functional AI-designed proteins
Google DeepMind · September–October 2026
DeepMind announced SynthID Bio, a family of methods intended to embed identifiable signatures in AI-generated biological designs. DeepMind researcher Pushmeet Kohli reported that the team had synthesized AI-designed proteins that were both functional and watermarked.
What was achieved: A reported proof of concept for marking functional AI-designed proteins.
What remains: This is not evidence that watermarking works for every protein or survives every modification, and it does not solve biological safety by itself.
DeepMind X announcement · Researcher’s X post
The takeaway is not that AI has become an independent scientist. It is that AI can now contribute to several parts of scientific work: generating candidates, carrying calculations across fields, proposing arguments, controlling experiments, and helping make results checkable. The strength of each achievement depends on what happened after the output: formal verification, expert criticism, laboratory testing, replication, or some combination of them.