AI Science Tracker

How we select and verify entries

This site catalogs scientific results achieved with AI since 2020, using original papers, preprints and institutional announcements.

What we include

  • A concrete result in a scientific field (biology, medicine, mathematics, climate, and others), with a documented AI contribution.
  • A dated original paper or preprint, or an announcement by the lab, company, university or institute responsible for the work. The source must support the result and the figures used in the summary. A paper source is not automatically peer-reviewed; an institutional page is not independent confirmation.
  • Tools and models only when the reviewed source supports a qualifying scientific result: for example, a newly characterized structure or compound, a mathematical proof, or a measured scientific outcome in a field test. The summary must identify that result and explain what AI did.

What we exclude

  • Product launches and general method or representation benchmarks without a specific scientific result. Better accuracy, speed or benchmark rank alone does not qualify a method. A source-described study of a specific scientific, clinical or field question can qualify, including a retrospective diagnostic-validation study; this does not establish clinical benefit or new biology. Prospective testing is not a universal inclusion requirement. Model releases are tracked separately on the Models page and do not count as discoveries.
  • Awards and recognition (such as the 2024 Nobel Prize): these judge a discovery; they are not the discovery itself.
  • Newsletters, aggregators, and third-party articles as the primary source.
  • Announcements without results yet (funding, collaborations, interviews).

Entries that do not meet the discovery criteria can be archived while their historical pages remain available. Archived entries are excluded from discovery lists, counts and statistics. The UNI pathology model entry is retained this way because it reports retrospective method benchmarks rather than a qualifying scientific discovery.

The archived NatureLM-audio and Prov-GigaPath entries likewise preserve method-development context outside the active counts. Each historical page explains the exclusion and links its original sources.

The Boltz-1 release and Cellpose 2.0 entries are archived as method benchmarks without a separate qualifying scientific result in the reviewed material. The review of AI-designed physics experiments is preserved as a synthesis of earlier work, rather than counted as a new discovery on the review's publication date.

By contrast, the Virchow entry describes a specific pan-cancer detector evaluated on held-out specimens. It is retained as a diagnostic-validation study, with no claim that the detector was deployed or improved patient outcomes. A general foundation-model benchmark suite alone does not meet this boundary.

Historical pages also preserve Co-Scientist ALS hypotheses and infectious-disease hypotheses, which had no completed biological result in the represented announcements. The OMol25/UMA, DolphinGemma and Perch update entries document releases rather than qualifying new application results. The FLIP imaging study is excluded because its researchers state that it did not use AI. These historical entries remain outside active counts.

The historical AJIN, MarsNet, AutoDNA and LPT entries preserve method or platform demonstrations outside active discovery counts. The Claude electronics prototypes likewise document a hardware capability demonstration. Their pages retain the sources, results and reasons for exclusion; archiving does not dispute the reported method performance.

What a count represents

The counting unit is an active catalog entry, not an independently established finding. A single paper or announcement may bundle several results in one entry. A method entry and an application entry may describe contributions to the same underlying result; for example, ColabFold and a nuclear-pore study overlap. Their presence does not establish two independent discoveries.

Use one entry for a coherent result or bundle from an announcement. Split it only when the source supports distinct scientific results that can each be summarized with their own AI contribution and limitations, and explain the relationship. Later announcements or paper versions about the same result normally update the existing entry, with version differences stated, rather than create a duplicate. Existing IDs remain stable. These rules guide review; they do not claim that every historical overlap has been resolved.

The models page

The Models page answers a different question: which model families are linked by name to catalog discoveries. It lists releases from leading labs since 2020, plus models from any date linked to a discovery.

We do not collect model names, labs, dates, domains, parameter counts, or access information ourselves. They come from the database Data on AI models from Epoch AI, republished under CC BY 4.0. Their stated limitations apply: the list is not exhaustive, and compute and parameter figures are estimates.

We make model-to-discovery links by name: we compare the technology field and each discovery title with model and family names. A match indicates the likely family, not necessarily the exact version used in an experiment. A discovery made with a model that the source does not name will not appear as linked. Some catalog technologies, such as those developed within a single study, have no Epoch AI record; the family can still appear without linked releases.

AI-discovered vulnerabilities

The separate Vuln discoveries page includes publicly disclosed individual vulnerabilities only when original sources establish an AI contribution. It does not include every security news item or convert aggregate bug counts into individual findings. These records are excluded from the scientific-discovery count.

Discovery context, public disclosure, AI attribution and documented remediation are distinct. A previously unknown pre-release bug is not automatically an exploited zero-day. Zero-day context remains attributed to the cited account, and documented fixes do not establish deployment or current safety. No attack is inferred from a vulnerability announcement.

What “Source checked” means

The stored verification status “verified” is displayed as “Source checked.” We have read the cited original source material and checked that it supports the title, date, organization, and figures reported in the summary. A paper abstract and its publication metadata can support a limited summary; this does not imply that we reviewed the full paper. Where review is limited to an abstract or an institutional announcement, the summary states that scope and does not treat missing details as established facts. This does not mean others have independently confirmed the result: figures are source-reported and may describe controlled conditions. The summary preserves stated limits such as lab-only, mice-only, not clinically validated, or preprint.

An entry marked “needs verification” has an unresolved source check, such as an unconfirmed date or organization, an unsupported reported figure, or conflicting information. A caveat does not turn an unresolved factual claim into a checked one; claims must be supported, corrected or removed before the status becomes “Source checked.”

Evidence context (pilot)

Optional annotations add “Evidence context” to the source details of selected entries: publication status, what was tested, what the sources say is still unknown, and what the AI system did. Every scientific statement comes from pages read for the annotation, linked in its source details and listed in the data file; nothing is inferred from the website, keywords or relevance score. Each annotation also records a test type, chosen from the tested context; it sets the validation subscore (see Relevance score). The other statements are separate from editorial judgments and do not change any score.

For new or revised annotations, distinguish three cases explicitly in the text: “Not assessed” when relevant material was not reviewed or was inaccessible; “Not stated in the reviewed material” when the pages read omit the detail; and a scientific unknown only when a cited source identifies it as unresolved. State “Not applicable” only with an explanation. An inaccessible full paper cannot establish that it omits a detail, and an omitted detail cannot establish scientific uncertainty.

Earlier pilot fields labeled “Not reported” do not reliably distinguish these cases and should be treated as unspecified annotation gaps until reviewed. No annotation means not annotated in this catalog, not absent scientific evidence. Record the actual review scope (announcement, abstract or full paper) and review date when a new review is performed; do not infer them from publication or catalog addition dates or backfill unrecorded historical checks. Optional structured review provenance now records the actual date and bounded scope of new checks. Older entries without it retain an explicit missing-review label; dates are never inferred from publication or addition.

Follow-up and scientific relationships

Optional source-backed events distinguish publication, same-team extension, independent follow-up, documented application, contradiction, correction and retraction. Event dates retain their recorded year, month or day precision; the recording date is separate. A journal transition is not replication. Missing follow-up means none recorded here, not none exists. Review scope is displayed only where actually recorded; the catalog does not claim continuous monitoring.

Directed relationships identify the same result, an extension, independent follow-up or shared evidence, with a source and bounded reason. They preserve historical IDs and do not merge counts. The antibiotic dossier illustrates a limited synthesis; model progress describes benchmark and release-date evidence separately from scientific outcomes.

Evidence and coverage

These tables describe all active entries in the loaded catalog, grouped by their recorded publication year. Annotation coverage counts the presence of the optional evidence block, regardless of its contents. Source composition counts each entry once by its primary source type, even if it also links a paper. Neither table measures scientific quality, independent confirmation, or how much research exists outside this collection.

Coverage tables require JavaScript. The raw discovery data is linked below.

Read the raw discovery data (JSON). Totals exclude archived entries and the separate vulnerability tracker. The current year is incomplete; missing annotations reflect catalog work, not missing scientific evidence.

Relevance score

Each entry has three integer editorial subscores from 1 to 5: impact, validation and novelty. First calculate their mean multiplied by 20 and round to an integer. The displayed score is the smaller of that result and (validation + 1) × 20. The range is 20–100, with thirteen possible values (20, 27, 33, 40, 47, 53, 60, 67, 73, 80, 87, 93, 100). A subscore step can change the total by about seven points, or by zero when the cap applies.

The cap limits how far impact and novelty can raise a weakly supported claim; it does not guarantee that every better-tested entry ranks above every less-tested one. These are coarse editorial judgments for browsing, not a scientific evidence grade or a reliable comparison across fields. Source details show the three subscores.

How we choose a test category

The catalog assigns a test category from the cited study description; this assignment is editorial, while the descriptions of what was tested remain attributed to the sources. The category sets validation using the following rules:

  1. Not reported: the reviewed material does not establish a qualifying test. This does not prove that no test exists.
  2. Retrospective or benchmark: performance assessed using existing data, including held-out observations, historical external cohorts and datasets collected prospectively before this model evaluation. An independent cohort is not the same as a separate independent replication.
  3. Direct test: new experiments or explicit mathematical/computational checks of the specific discovery claim within the discovery study. Initial assays of AI-selected candidates belong here even if predictions preceded the assays or a collaborating contractor ran them.
  4. Prospective or independent check: a locked predictive model evaluated before new outcomes accrue in its intended workflow, a separate research team's follow-up test of the result, or a formal proof artifact checked by a trusted proof kernel. Proof-kernel checking establishes formal correctness relative to the encoded statement and axioms, not human review or scientific relevance. Prospective sample collection alone, outsourced experiments and collaborator blinding do not establish this category.
  5. Replication or operational use: documented repeated confirmation or sustained operational performance of the claimed outcome. Availability of a tool or its deployment alone is insufficient.

For mixed studies, classify the check supporting the represented result and describe ancillary tests separately. Apply the existing-data rule when a predictor is evaluated retrospectively, even on an external or prospectively collected cohort. Use a higher category only when the reviewed source establishes the additional qualifying check. Unknown timing or independence remains explicit. Entries without an annotation retain their earlier validation subscore pending review.

Citations are shown separately as an OpenAlex band comparing works of the same type, year and subfield. The percentile uses citations within the publication year and the three following years; the displayed citation count is OpenAlex's total count at the stated snapshot date. Works younger than a year are omitted. Each metric links to its work-level source; later source values can differ from the saved snapshot. Citations measure attention, not correctness, and never enter the score. See OpenAlex's metric definitions.

The anchors below guide editorial review. Impact and novelty remain judgments and have not all been recalibrated against these anchors.

Editorial scoring anchors (each criterion is assessed separately)
LevelImpact: significance of the reported resultValidation: support for the stated claimNovelty: departure from prior work
1Narrow interest in one task or setting; broader use is speculative.A source reports the result, but the material reviewed provides little direct evaluation of the central claim.A small extension or refinement of an established result.
2A concrete benefit for a specialist task, with limited demonstrated reach.Existing-data evaluation supports the claim, including retrospective internal, held-out or external cohorts. Transfer and prospective outcomes remain separately assessed.A distinct technical improvement or combination within an established approach.
3A substantial result for a defined field, research problem or population.New direct testing supports the specific discovery in a bounded setting, such as a candidate assay or an explicit mathematical/computational check within the discovery study.A substantially new capability or result for a defined problem, with its difference from prior work documented.
4Demonstrated consequences across several problems or settings within a field.A locked model is evaluated before new outcomes accrue in its intended workflow, separate investigators confirm the result in a follow-up study, or a trusted kernel checks its formal proof artifact. This does not imply human review. Prospective collection or a collaborating laboratory alone does not qualify.A new approach or finding that opens several research directions beyond the initial task.
5Documented broad consequences that change research or practice across a field.Repeated confirmation or sustained operational performance supports the claimed outcome in its intended settings, with limitations documented. Deployment alone does not qualify, and use by the originating organization is attributed as such.A documented foundational change in how a class of scientific problems can be approached.

Judge validation against the claim and field: a proof, a predicted molecular structure and a clinical benefit need different checks. Peer review, prizes, publicity or deployment alone do not verify every claim. Computational work is not automatically weak evidence; experimental work is not automatically strong evidence. Keep the source-stated test context in the evidence annotation and explain the editorial rationale separately when reviewing a score.

An automated calibration trial was not applied. Agreement between repeated automated assessments is not independent scientific validation or a substitute for human review. Impact and novelty retain their existing editorial values pending documented record-level review.

A category average is the mean score of its selected entries. It reflects composition and editorial choices, not comparative field quality. With only one or two entries, the average is especially uninformative.

Reading the catalog statistics

The annual chart offers the number of catalog entries and their mean editorial relevance. Means use the rounded 20–100 scores displayed on individual entries, then round the annual average to an integer; averaging does not add precision to the rubric. Annual sample sizes appear in the table; small samples and the incomplete current year require caution. These are descriptions of a selected collection, not worldwide discovery rates or measures of scientific progress.

The source composition counts each entry by its primary source type, not by evidence quality. “Most recent catalog addition” uses the addition date; it does not indicate the last scientific review or follow-up check.

How we choose the date

We use the documented announcement or publication date of the result represented by the entry. For paper-based entries, this can be a preprint or journal publication date; the summary identifies the version when the distinction matters. A later institutional release can support an earlier paper date if it explicitly states it, and the summary distinguishes the two dates. These dates do not establish priority over every earlier version or announcement. When a page’s date conflicts with its text, we note the discrepancy (for example, DeepMind’s post about the first 350,000 AlphaFold Database structures is dated 2022 but describes events in 2021). Historical IDs remain stable when a date or source is corrected.

Data structure

Each entry is a record in data/discoveries.json:

FieldContents
id, titlestable identifier and title
summary2–4 sentences in English, including stated limitations
fieldbiology, chemistry, medicine, materials, climate, mathematics, physics, computer science, humanities, multidisciplinary, or other
ai_technology, organizationtechnology used and announcing organization
source_url, source_typeoriginal paper, preprint or institutional source link and type
publication_date, added_dateannouncement date and date added to the catalog
verification_statusverified or needs verification
paper_urlpeer-reviewed paper or preprint, when available
relevanceimpact, validation, and novelty, each scored from 1 to 5; displayed score = min(rounded average × 20, (validation + 1) × 20)

Reuse and citation

The catalog content (records, summaries and editorial scores) is licensed under CC BY 4.0; please credit “AI Science Tracker” with a link. Epoch AI model data keeps its own CC BY 4.0 attribution, and each linked source keeps its own terms. The repository’s citation file describes the catalog, not the underlying papers.

Limitations

  • Collection is manual and not exhaustive; selection depends on the sources reviewed so far.
  • Companies describe their own results, so the tone may be promotional and claims are not independently assessed.
  • Entries come from original papers, preprints and institutional announcements. The catalog reflects publication and announcement choices as well as our manual selection; the home page states how concentrated the primary-source websites are.
  • Organizations are separated with semicolons and counted under the first complete institution name in charts. Commas and conjunctions inside an institution name are preserved; historical ambiguous profiles link to the organization directory.