About the data

Papers

Publication counts come from fixed OpenAlex searches of titles and abstracts. They include articles, preprints and reviews, and exclude retracted work. Each area shows its query, retrieval date and annual counts for 2019–2025.

A match may describe an application of a method, rather than research that advances it. Areas overlap, and preprint and journal versions can have separate records. Search precision and recall have not been systematically measured. Counts describe the results of each query, not the entire discipline.

Publishing authors

Author calculations use 1,000 randomly sampled papers per area, or 2,000 for pre-training, reasoning, evaluation and tabular ML. Where a query returns fewer papers, every result is included. The samples use seed 42. Each area shows its actual sample size, coverage and missing metadata.

Observed counts are distinct OpenAlex Author IDs across a complete query result. They describe the authors identified in those records. Missing IDs and author disambiguation errors still affect the count.

Scenarios are used when only a sample is available: paper count × average distinct Author IDs per paper ÷ assumed papers per author per year. The base case uses 3 papers; the range uses 1–5. Values cannot fall below the distinct authors already observed in the sample. The publication rate is not measured for each field, so the range is not a confidence interval.

For sampled areas, a 95% bootstrap interval describes variability in the mean number of known authors per paper, using 1,000 resamples. It does not capture search errors, missing metadata or uncertainty in publication rates. No correction is made for sampling a large fraction of a small corpus.

These figures include applied coauthors and exclude people absent from the matched publications. They do not measure employment, full-time research jobs or everyone working in a field.

Private investment

The source is Quid’s 2025 private AI investment data in the Stanford AI Index 2026. Company coverage includes firms that raised more than $1.5 million since 2013. These are investments in companies, not research spending or revenue.

For research areas, broader industry totals are distributed using stated editorial weights. Those allocations are assumptions. The range is 0.5–2× the base scenario; the chart uses the base value. Neither the range nor its midpoint is a measured market size.

Applications use a related industry benchmark where available. For two overlapping segments, the range runs from the larger segment to their sum. Other applications use stated shares of broader segments. Each area lists its source figures and weights.

Organizations

Each area includes examples of companies, research teams and products with documented activity in that field. Names link to official research, product documentation or other primary sources. Selection considers both specialist providers and organizations with documented work across several areas. A short editorial selection is visible first; additional examples can be expanded.

Roles distinguish research, model development, tools, data services and applied products. Research examples can refer to published contributions; they do not imply that an older model is still commercially available. A robotics product can illustrate an application without establishing a contribution to robot-learning research. Acquired businesses and brands are identified where ownership has been checked.

The examples are not ranked and do not form a complete market census. Inclusion does not establish market share, product quality or research leadership. One organization can appear in several areas, and entries can belong to the same corporate group.

Organization sources were reviewed on 26–27 September 2026, separately from the 2025 publication and investment figures. Source availability and company focus can change.

Comparing areas

The groups are a practical guide to AI research, not mutually exclusive segments. A paper, author or company may appear in several areas. Research and application figures must not be added together. Dividing company investment by author estimates does not give researcher compensation or funding per researcher.

Topic tags describe subtopics; they have no separately measured values. Conference scopes inform the grouping, while the exact search in each area determines what its publication figures cover.