Claims that artificial intelligence has discovered something now arrive almost every week: new antibiotic candidates, protein structures, materials, mathematical proofs. AI Science Tracker, an open catalogue maintained by Marcello Pichini, collects those claims in one place and says plainly how well each one has been checked.

The catalogue lists 1,362 active entries in 11 fields, attributed to 738 organisations, starting in 2020. Biology, medicine, materials science and physics account for most of them. Mathematics has grown fastest: 35 entries are dated 2026, against eight in 2025. Each record names the AI system involved, the organisation that announced the result, the date, and a link to the original paper, preprint or institutional page. The short summary keeps the source's own limits, such as "mice only", "lab only" or "preprint".

What counts and what does not

The inclusion rules are stricter than most press coverage. An entry needs a concrete scientific result with a documented AI contribution: a newly characterised structure or compound, a proof, or a measured outcome. Product launches, model releases and benchmark gains without a specific finding are left out, and so are prizes, funding rounds and interviews. Several well-known tools survive only as archived pages for that reason, among them the Boltz-1 release and Cellpose 2.0, which the catalogue treats as method benchmarks rather than discoveries.

The site is also explicit about what its numbers mean. A count is a number of catalogue entries, not of independently established findings. One paper can bundle several results, and a method and its later application can describe the same underlying advance.

"Source checked" is not "confirmed"

Every entry carries one of two labels. "Source checked" means the editors read the cited material and found that it supports the title, date, organisation and figures in the summary. It does not mean that anyone has reproduced the result. "Needs source check" marks entries where a date, an affiliation or a figure is still unresolved.

Each entry also gets three editorial scores from 1 to 5, for impact, validation and novelty. Validation follows a five-step ladder: no qualifying test reported; a retrospective or benchmark test on existing data; a direct test within the discovery study, such as an assay of AI-selected candidates; a prospective or independent check, which includes a proof verified by a formal proof kernel; and, at the top, repeated confirmation or sustained use in practice. The overall score is capped by validation, so a striking but weakly tested claim cannot rank high on impact and novelty alone.

The coverage tables on the methodology page show how much checking is still open. In medicine, 17 of the 65 entries dated 2026 await a source check; in materials science the figure is 18 of 57. In mathematics only one of the 35 entries dated 2026 is pending.

Why it is useful

The value for readers lies in the line between what an AI system predicted and what was then tested. A structure prediction, a list of drug candidates and a proven clinical benefit need very different evidence, and the tracker keeps those differences visible instead of folding them into one headline. It states its own weak points too: collection is manual and incomplete, and many entries rest on companies describing their own results.

The records, summaries and scores are published under a CC BY 4.0 licence, and the full dataset can be downloaded as JSON.

27Views

Jenna Mercer

Author

World News Correspondent

Jenna Mercer covers public affairs, politics, business, culture and daily news for Science Official. The role focuses on verification, context, and clear explanations for readers.