AI for science has produced a small number of results that survive scrutiny and a much larger number that do not. The verified wins are concentrated in structure prediction, algorithm and code search, literature synthesis, and now formal mathematics: on August 1, 2026, OpenAI published ten solutions to long-open math and theoretical computer science problems with machine-checkable Lean 4 proofs. The failures cluster around a single pattern, which is a claim of novel discovery that turns out to be retrieval, misinterpretation, or an unreplicated result. Telling the two apart is now the core skill for anyone reading this field.

The October 2025 Incident That Set the Standard for Skepticism
Start with the cautionary tale, because it explains why careful people now ask a specific question first.
In October 2025, an OpenAI executive posted that GPT-5 had “found solutions to 10 (!) previously unsolved Erdos problems and made progress on 11 others.” Thomas Bloom, the mathematician who maintains the Erdos problems database, called the claim a dramatic misrepresentation. Problems marked “open” on his site meant only that he was personally unaware of a paper solving them. GPT-5 had located existing published references.
OpenAI researcher Sebastien Bubeck conceded that only solutions already in the literature had been found. The post was deleted. Google DeepMind’s Demis Hassabis said publicly, “This is embarrassing.”
The useful part is that this is not a nothing result. Exhaustive literature search across a century of mathematics is genuinely hard and genuinely valuable, and the same retrieval strength is what makes these models useful for ai in research generally. It just is not discovery, and the two got conflated.
Where AI Is Solving Math Problems for Real
By mid-2026 the picture is materially better than it was, and the improvement came from a specific change: verification moved from human review to machine-checkable formal proof.
Three results anchor the current state of ai solving math problems.
The Lean formalization is what changes the epistemics. A proof either compiles against Lean’s trusted kernel or it does not, and anyone can check without being a specialist. Bloom, the same mathematician who debunked the 2025 claim, called the Astra results big news.
Two caveats belong beside that. A successful compile confirms the formal statement is proved; it does not confirm the formal statement faithfully captures the problem as the community understood it. And OpenAI has not confirmed Astra’s public release date, whether it ships as GPT-6 or as a separate family, or its availability timeline.
Competition Benchmarks Are Saturating, Which Tells You Less Than It Seems
In 2025, Google DeepMind’s Gemini Deep Think scored 35 out of 42 at the International Mathematical Olympiad and became the first AI system graded and certified by official IMO coordinators. An OpenAI experimental model reached the same score but was graded by three former IMO medalists rather than by the official process.
At IMO 2026 in Shanghai in July, reporting indicates two systems, from Huawei and from Xiaohongshu, received official coordinator grading with perfect 42 out of 42 scores. Separately, an independently run and self-administered evaluation reported four frontier models at 42 out of 42, but that test used AI agents as graders rather than human medalists, and its organizers described the scores as strong but not authoritative.
Olympiad problems have known-good answers and bounded search. Research does not. The gap is visible in the numbers: one 2026 analysis noted a frontier model scoring in the high seventies on competition-level mathematics and around 25 percent on open-ended research tasks. Benchmark saturation is real and it is not the same thing as research capability.
Where AI Genuinely Accelerates Science Today
Outside mathematics, four areas have produced results that hold up.
Structure Prediction
AlphaFold remains the strongest case in the field, now five years past its breakthrough and embedded in ordinary practice through the AlphaFold Protein Structure Database. The 2026 development worth noting is a governance one: Nature reported in February 2026 that Isomorphic Labs, DeepMind’s drug-discovery spin-off, has a successor model informally described as an AlphaFold 4 that it is not releasing, leaving open-source developers guessing at the method. The science advanced and access narrowed at the same time.
Algorithm and Code Discovery
AlphaEvolve, Google’s Gemini-based agent for evolving and optimizing code, reached general availability on Google Cloud on July 9, 2026. Its published customer results are unglamorous and checkable, which is what makes them credible: PacBio reported a 30 percent reduction in variant-detection errors in genomics sequencing, Schrodinger a fourfold speedup in a molecular discovery workflow, FM Logistic a 10.4 percent routing improvement, and Google internally reported a 20 percent reduction in write amplification in Spanner. Writing better code for simulations and analysis pipelines is a large, real, and underrated contribution.
Hypothesis Generation With a Human in the Loop
Google’s AI co-scientist, a multi-agent system for ai hypothesis generation, was published in Nature in May 2026. In acute myeloid leukemia, several suggested drug-repurposing candidates inhibited tumor viability in multiple cell lines at clinically relevant concentrations. In liver fibrosis, Stanford’s Gary Peltz used the system to surface vorinostat, an approved anticancer drug, which reduced a key TGF-beta-induced chromatin change by 91 percent in hepatic organoids, subsequently published in Advanced Science. Everything is preclinical and in vitro, and one 2026 evaluation reported that agents ignored available evidence in a majority of their reasoning traces.
Literature Synthesis
Edison Scientific, spun out of the nonprofit FutureHouse in November 2025, launched Kosmos, an ai scientist system that reads roughly 1,500 papers and executes about 42,000 lines of analysis code in a single run. The company claims 79.4 percent accuracy in its conclusions and beta users estimated a day of Kosmos equals six months of their work. Retrieval and synthesis at that scale is a genuine capability, and it depends on the same indexing and ranking dynamics behind how search engines work.
Where It Is Oversold
Capital has moved much faster than published results. Periodic Labs raised a $300 million seed round in September 2025 and was reported at roughly a $7.5 billion valuation by early 2026, with no peer-reviewed discoveries announced. Lila Sciences launched out of Flagship Pioneering with $200 million and has raised toward $550 million total, passing a $1.3 billion valuation with Nvidia participation. CuspAI has raised more than $130 million.
None of that is evidence of fraud. It is evidence that valuations are pricing a capability that has not yet been demonstrated in the literature, and that is a normal early-stage condition that gets misreported as achievement.
The autonomous-laboratory story is the sharpest illustration. A 2023 Nature paper from Berkeley’s A-Lab reported 40 compounds synthesized autonomously. After external re-analysis of the diffraction data, an author correction published on January 19, 2026 clarified that novelty meant new to the prediction platform rather than new to science, reclassified four of the 40 as inconclusive, and removed one compound that had mistakenly been in the training data. The corrected count is 35. The correction took more than two years.
The Verification Problem Is the Actual Bottleneck
Every one of these cases turns on the same question: who checks the output, and how long does checking take? Three regimes exist right now, and they are not equally trustworthy.
That last point is the structural problem. If a system produces a hundred plausible hypotheses and 6.5 percent are meaningfully correct, the bottleneck moves entirely onto expert attention, which is the scarcest input in science and the one AI does not create.
AI Slop in Journals Is Measurable Now
The downstream cost of cheap generation is showing up in the publication record. In April 2026, the journal Organization Science published an editorial analysis of 6,957 submissions and 10,389 reviews. Submissions had risen 42 percent since ChatGPT’s launch. Manuscripts with more than 70 percent detected AI content received a revise-and-resubmit decision only 3.2 percent of the time, against 11.9 percent for low-AI papers. More than 30 percent of peer reviews showed detectable AI use, and AI-influenced reviews did not correlate with editorial decisions the way human reviews did.
The editors’ conclusion is the right frame for the whole subject: AI is an agnostic tool, and you can point it at quality or at volume. Their response was to nearly double their volunteer editorial ranks to filter submissions, which is a tax paid in unpaid human labor.
How to Read an AI for Science Claim
A short checklist handles most of it.
- Was the result verified formally, replicated experimentally, or only asserted? If the answer is “a blog post,” wait.
- Was the problem actually open, or open on one curated list? This distinction sank the October 2025 claim.
- How much human work went into the final artifact? “One shot” claims usually precede a paper describing extensive human reorganization.
- Is the model available for independent testing, or is the claim tied to an internal system nobody outside can probe?
- Did a named domain expert who does not work for the vendor go on record?
Applied honestly, that filter leaves a real and narrower conclusion. AI is now a working instrument in mathematics, structure prediction, code and algorithm search, and literature synthesis, and it is a promising but unproven instrument for autonomous discovery in the physical sciences. The gap between marketed capability and verified capability follows the same shape you see in Physical AI and Humanoid Robots, and the efficiency pressures now pushing labs toward cheaper task-specific systems are the same ones described in Small Language Models. Anyone whose job involves being cited accurately by these systems, which is increasingly what answer engine optimization addresses, has a direct interest in that gap closing.
Frequently Asked Questions
Has AI actually solved any previously unsolved math problems?
Yes, a small number with real verification behind them. In May 2026 an OpenAI model produced the counterexample that disproved the Erdos unit distance conjecture, and on August 1, 2026 OpenAI published ten solutions to decade-old open problems with machine-checkable Lean 4 proofs. Earlier claims from October 2025 turned out to be literature retrieval, not discovery.
What is an AI scientist?
An AI scientist, sometimes called an artificial scientist, is a multi-agent system that generates hypotheses, plans experiments or analyses, and writes up results. Google’s AI co-scientist and Edison Scientific’s Kosmos are the two best-documented examples, and both currently operate with a human researcher in the loop.
Is Astra a real OpenAI model?
Yes. OpenAI publicly confirmed Astra on August 1, 2026 as its next major model family alongside the ten formalized math results. What OpenAI has not confirmed is the release date, public availability timeline, or whether it ships under the GPT-6 name.
Can AI replace peer review?
No, and current evidence suggests it is straining peer review instead. An April 2026 analysis in Organization Science found more than 30 percent of submitted reviews showed detectable AI use, and that those reviews did not correlate with editorial decisions the way human reviews did.
What is the biggest limitation of AI for science right now?
Verification capacity. Models generate candidate results far faster than experts can evaluate them, and outside mathematics there is no fast automated check, so expert attention and wet-lab replication remain the binding constraints.
Make Your Technical Content Accurate and Citable
WD Morgan Solutions is a veteran-owned digital marketing and development firm in St. Petersburg, Florida, helping businesses across Tampa Bay and nationwide with answer engine optimization and application development, including making technical content accurate and citable for AI search. Visit our homepage or contact us for a free consultation.

























