How Many Authors Does an AI Paper Have Now?
Published · paper counts are current, not as-published
A research paper used to be something one or two people wrote. At the venues measured here that is no longer the ordinary case, and the change is recent enough to watch happen year by year.
The figures cover 21 venues - those that published in every year from 2019 to 2025 with indexing complete enough in each to be comparable. That leaves out 15 venues of the 36 indexed here, and not only recent additions: any venue with a gap or a thinly indexed year drops out, including large and long-established ones. The restriction matters more for this measure than for a simple paper count, because team-size norms differ sharply between subfields.
The Middle Paper Gained an Author
| Year | Papers | Middle paper | Mean | Solo | 10+ authors |
|---|---|---|---|---|---|
| 2019 | 9,989 | 4 | 4.04 | 3.3% | 1.2% |
| 2020 | 11,740 | 4 | 4.22 | 2.8% | 1.7% |
| 2021 | 12,764 | 4 | 4.35 | 2.7% | 2.1% |
| 2022 | 13,793 | 4 | 4.58 | 2.4% | 2.9% |
| 2023 | 17,698 | 4 | 4.75 | 2.4% | 3.3% |
| 2024 | 20,676 | 5 | 5.08 | 2.1% | 5.2% |
| 2025 | 26,637 | 5 | 5.42 | 1.9% | 7.4% |
The middle paper went from 4 authors to 5. That column is a median - the count an actual paper in the middle of the year carries, not an interpolation - so it is not dragged by the handful of very large collaborations further down this page.
It Is Not Just the Big Venues Growing
Fixing the venue set stops venues entering and leaving, but it does not fix their weights: the venues that grew fastest carry more of the pooled figure every year, so a pooled number can move without any venue changing. The test is to recompute inside each venue separately.
Of the 21 venues, 12 ended the window with a higher middle paper than they started, 8 finished unchanged and 1 ended lower. On the large-team share below, 17 rose, 2 held and 2 fell. Both movements are happening within venues, not only between them - which is the claim the pooled table on its own could not make.
Solo Work Is Thinning Out
Single-author papers fell from 3.3% of the corpus to 1.9%. The share did not fall in every single year of the window, but the direction across it is not in doubt.
And Large Collaborations Are Not Rare
At the other end, papers with 10 or more authors went from 1.2% of the corpus to 7.4%. The largest single paper in this window carries 125 authors.
The two ends moved in opposite directions, which is the part a mean would have hidden. Work that one person finishes alone is becoming less common at the same time as work needing a large group is becoming more so, and a single average sitting between them describes neither.
What This Does Not Show
- Not who did what. These are author counts. The corpus holds no contribution statements and no authorship-order semantics, so a paper with twelve names could be twelve collaborators or two authors and a dataset team.
- Not people, exactly. The count is of author slots on a paper, which is the reliable quantity. Deciding whether two similar names are one researcher is a separate and imperfect problem - see the authors directory - and nothing here depends on getting it right.
- The completeness filter is a heuristic. A venue year below a quarter of that venue's own median is treated as incompletely indexed and its venue is dropped. A genuine collapse in output is indistinguishable from an indexing gap in this data, so the rule can remove a venue that really did shrink.
- Not causes. Larger compute budgets, dataset papers, industry labs and changing credit norms would all produce this shape. The corpus records the shape, not the reason.
- Not the whole of research. A sample of top CS/AI venues, not a census - and a further 15 venues indexed here did not qualify for the comparison at all.
For the volume side of the same window, see how fast the field is growing. The numbers here are read from the database when you load the page, so they move as the corpus does.