Can AI Make Scientific Discoveries? What It Has Found

Yes, in specific and still mostly early cases. In the science stories we have reported and checked, AI has helped prove a record result in number theory, picked out cosmic lenses that telescopes then confirmed, designed working enzyme parts, flagged a strange pattern in virus DNA, found hidden human proteins by their shape and turned 27 years of radio data into a movie of a black hole jet. In every case people still had to check the work, and none of it has reached everyday life yet.

That is how science should work: the machine proposes, and a telescope, a lab bench or a mathematician decides. The first to gain are researchers. Patients and the rest of us come later, if the results hold.

Checking the hardest math

For 80 years mathematicians have tried to show that more of the Riemann zeta function’s zeros sit where the Riemann hypothesis says they must. This summer an unreleased research version of Anthropic’s Claude produced a proof that more than 67.25 percent of those zeros are both simple and on the critical line, up from a record of about 41.7 percent. Two Anthropic mathematicians checked it and posted it, and Youness Lamzouri, a mathematician in Nancy, France, then proved the same bounds by a clearer route of his own. Both are preprints. The hypothesis itself is still open. Read how a machine’s record proof got a version people can follow.

Finding needles in the sky

Gravitational lenses are rare alignments where a massive galaxy bends the light of something far behind it, working like a free telescope. A neural network searched the DESI Legacy Imaging Surveys and produced thousands of candidates. A team led by Emerald Lin, an undergraduate at UC Berkeley, and Ivonne Toro Bertolla of NSF NOIRLab took spectra of 76 of them: 55 were fully confirmed, 15 need more observations and six were not lenses at all. Lenses can help study dark matter and how fast the universe is expanding. See how astronomers gained 55 natural telescopes.

AI can also sharpen what telescopes already saw. Astronomers at IAA-CSIC, Caltech and the University of Toronto built kine, which uses a neural network to fit 116 radio observations of the blazar 3C 345, taken from 1995 to 2022, into one continuous video. It reached about four times the nominal resolution, and the motion it revealed puts the usual shock wave explanation for the jet’s bright blobs into question, at least for this one source. The authors suggest the method could one day help image moving organs in MRI, but that has not been tested. Watch 27 years of a black hole jet.

Designing and finding molecules for medicine

Some antibiotics, immunosuppressants and cancer drugs are built by microbes on enzyme assembly lines that are hard to rewire. A team at the Helmholtz Institute for Pharmaceutical Research Saarland used generative protein models to design 76 new carrier parts, built them into 578 assembly lines in living E. coli, and found the best designs raised output of a model peptide by up to about three times. No new drug was made and nothing was tested against bacteria. This study is peer reviewed. Read about rebuilding the microbe factories behind some medicines.

Most medicines work by latching onto a protein, and many human proteins are still barely understood. University of Miami scientists compared 214,528,851 AlphaFold2 structure predictions against the shape of a classic drug target and found TM184C, a protein that looks like a G protein coupled receptor but lives inside the cell and helps build bridges between neighboring cells. Lab work in cells and yeast showed what it does. There is no drug and no animal or patient data. Learn how hidden human proteins were found by shape.

Spotting odd patterns in DNA

CRISPR was first noticed as a strange repeat in bacterial DNA. Software agents running Claude, searching 1.9 billion protein clusters, flagged something with a similar look: arrays of repeated DNA sitting beside a reverse transcriptase and an unknown partner gene in jumbo phages, the oversized viruses that attack bacteria. Nobody has shown the enzyme is active or what the system does. The work is a preprint, all authors are Anthropic employees, a provisional patent has been filed, and a University of Copenhagen researcher disputes who found it first. In additional runs, the agents missed the array entirely. Read about the odd repeats in virus DNA.

How close is any of this to everyday life?

Already in scientists’ hands: the 55 confirmed lenses are a shared resource for astronomers, and the black hole movie is public for anyone to watch.
Getting closer, still years off: the designed enzyme parts. The next real test is whether they work in assembly lines further from the ones they were designed for, and then whether new peptides made this way can fight bacteria.
Years away: the zeta proof (still preprints, and the hypothesis is open), the hidden protein TM184C (cells and yeast only), and the DNA repeats (function unknown).

These six stories are the cases we have reported and checked so far, not the whole field. What they share is a pattern worth trusting: AI widens the search, and people decide what is true.

Questions people ask

Can AI do math proofs?

In at least one case, yes. A research version of Anthropic’s Claude produced a proof that more than 67.25 percent of the zeros of the Riemann zeta function are simple and on the critical line, a record bound. Two mathematicians checked it and posted it, and Youness Lamzouri then re-proved it his own way. Both papers are preprints, and the Riemann hypothesis itself is still open.

How does AI help astronomers?

In the cases we have reported, it helps in two ways. It searches huge sky surveys for rare objects, such as the gravitational lenses a neural network flagged before telescopes confirmed 55 of them. And it rebuilds blurry, scattered observations into something clearer, like a continuous movie of a black hole jet made from 116 radio observations taken over 27 years.

Can AI find new drug targets?

It can help point to candidates, though nobody has a drug from this yet. University of Miami scientists searched AlphaFold2 structure predictions by shape and found TM184C, a little understood human protein that resembles a major family of drug targets. Lab work in cells and yeast showed what it does, but there is no known molecule that switches it on, and no animal or patient data.

Did AI discover anything new in DNA?

Possibly, but it is early and contested. Software agents searching 1.9 billion protein clusters flagged a repeat array and a partner gene beside a reverse transcriptase in the DNA of viruses that attack bacteria. Nobody yet knows what the system does, the work is a preprint by Anthropic employees, and a University of Copenhagen researcher has publicly questioned who found it first.

Can AI make scientific discoveries on its own?

Not in any case we have reported. Even the most independent example, the zeta proof, needed mathematicians to check it, and a human then rewrote it in a form people can follow. In the DNA search, agents missed the repeat array entirely in additional runs. Every result here became science only after people tested it with telescopes, lab work or careful reading.

The next result could come from a student pushing past two thirds on the zeta zeros, a telescope working through the thousands of lens candidates still waiting, or a lab finally learning what that phage enzyme copies. When it does, you will find it on our science page, checked and explained in plain words.