AI Spotlight Exclusive: Machines Are Now Doing Real Science. The Hard Part Is Proving It

Something shifted in science between late September and mid-October 2026. In the space of four weeks, artificial intelligence systems from China, the United States and Sweden were credited with work that, until recently, only trained researchers could do: settling an open problem in pure mathematics, discovering a new biological system, writing hundreds of research papers and running laboratory experiments on their own. Look closely at each claim, though, and a second story appears. The machines are getting faster at producing results. Checking those results is now the bottleneck.

The cleanest case came from Shanghai. On October 14, Shanghai Artificial Intelligence Laboratory announced that its Shusheng-S2 model (also called Intern-S2) had helped resolve the three-dimensional case of the Ghys conjecture, a problem in complex dynamical systems first posed by the French mathematician Étienne Ghys in 1995. The paper, posted on arXiv as preprint 2609.37695, has a single author: Jiesong Zhang, a postdoctoral researcher at KTH Royal Institute of Technology in Stockholm, working with his former doctoral supervisor Xu Disheng of Great Bay University. According to TechTimes, the work ran over seven rounds starting in August, with the AI exploring proof branches and the mathematicians choosing which ones to pursue; the proof graph grew from 15 nodes to 389. What makes this result unusually strong is what happened next. A separate team at Peking University, led by Dong Bin and Academician Liu Ruochuan, checked the entire proof in Lean, a programming language in which a computer verifies every logical step. Using up to 100 AI agents over roughly five days, they produced more than 550,000 lines of Lean code confirming all four main theorems. The code is reported to be public, so anyone with Lean installed can check it. Two caveats remain: the lab's claims about the model's benchmark scores have not been independently tested, and the model itself is not available to outside researchers.

In biology, the speed is the headline. On September 24, Anthropic unveiled the first discovery from its molecular biology lab: a previously unknown enzyme system, named ART, found in viruses that infect bacteria. According to the company, about 950 agents powered by its Claude model searched a DNA database for 21 hours, found more than 200,000 genes for one type of enzyme, and narrowed them to 20 worth studying. Anthropic says that search would take an expert weeks or months. Early experiments suggest the system might be programmable in the way CRISPR is, which is why it attracted attention. But the finding is a preprint that other scientists have not yet reviewed, chief executive Dario Amodei himself called it preliminary, human scientists carried out all of the laboratory experiments, and its function is still unknown. Feng Zhang, the CRISPR pioneer at MIT and the Broad Institute, described the repeated DNA fragments as genuinely intriguing and worth further investigation.

The biggest and most contested release came from OpenAI. On October 6, the company published 722 maths manuscripts, grouped into 372 families of related results, produced by an internal model it has not released. Some of them touch on famous problems such as the Riemann hypothesis, a special case of the Hodge conjecture and the four-dimensional Kakeya conjecture. Touching on is not solving: OpenAI has not claimed to have proved the Riemann hypothesis. Within a day, three manuscripts were withdrawn because a sign error broke one argument and two papers that depended on it, and 14 others were revised, leaving 719. According to TechCrunch, about 300 of the top-line results, roughly 42 percent, were formalised in Lean; the rest still need human checking. The release followed a disputed September claim about the Navier-Stokes equations, one of the Clay Millennium Prize Problems; the Clay Mathematics Institute still lists that problem as open.

The reaction from mathematicians has been sharp, and it is less about correctness than about what the discipline is for. Fields Medalist Terence Tao, in comments reported by TechCrunch, warned that problems are being solved by AI prompters who do not understand the output well enough to answer questions about it, give talks or engage with the rest of the field. The Association for Human Mathematics went further, calling the release not a demonstration of scholarship, but a demonstration of power, and urging mathematicians not to collaborate on this kind of work.

In Sweden, the direction is the same but quieter. Researchers at Chalmers University of Technology described a closed-loop AI laboratory, published in the Journal of the Royal Society Interface, that studies brewer's yeast by combining language models, automated reasoning and laboratory robots. It can pose questions, propose experiments, assess the results and update its understanding. Senior author Ross King said that, for now, AI will augment rather than replace scientists; humans still set priorities, interpret the meaning of the results and provide ethical oversight.

Why this matters, including for India's students and researchers: these tools are compressing weeks of searching into hours, and in mathematics, formal verification in Lean is emerging as the standard that separates a result from a press release. But Lean is not magic either. It checks a proof against the statement as typed into it, so a mismatch with the written claim can slip through, and earlier in 2026 a widely shared Lean disproof of the Collatz conjecture collapsed when it turned out to rest on a bug in Lean itself. The lesson of these four weeks is simple. Treat AI results as strong leads, not settled facts, until independent experts or verifiable proofs have checked them. The Ghys proof shows what that looks like when it is done well.

Sources: TechTimes on the Ghys conjecture proof; Phys.org (AFP) on Anthropic's ART enzyme discovery; Phys.org on the Chalmers AI scientist; Online Tech Tips, Tech AI Magazine and Shattered on OpenAI's 722 manuscripts and the reaction; FatherOfAI on AI for science in 2026. OpenAI's own figures are as reported by these outlets and TechCrunch.