How Fast Is AI Research Actually Growing?

Published September 18, 2026 · paper counts are current, not as-published

Everyone agrees that computer-science and AI research is growing. Almost nobody puts a number on it, and the numbers that do circulate usually measure something else: a database adding sources, a publisher widening its index, a preprint server that did not exist a decade ago. Growth in what you can see is not the same as growth in what is there.

Conftrace indexes 224,410 papers from 36 leading CS/AI conferences and journals. That total has the same problem: it grew partly because the corpus itself grew. So this piece asks a narrower question that the data can actually answer - did the same venues, compared with themselves, publish more?

The Same Conferences, Twice the Papers

21 of the 36 venues indexed here published in every year from 2019 to 2025, with indexing complete enough in each of those years to be comparable. That holds out two distortions at once: the corpus adding venues over time, and a venue whose proceedings were only partly available when the corpus was built. 1 otherwise-qualifying venue failed the second test and was dropped.

The same 21 venues went from 9,989 papers in 2019 to 26,637 in 2025 - a rise of 167%. Averaged across the set, a venue went from about 475 papers a year to roughly 1,268. That is an average, not a typical venue: a few large ones pull it well above the middle, and the conferences directory carries each venue's own volume rather than leaving the spread asserted here.

The Growth Is Not Steady

The year-over-year column is the more interesting one. The slowest year in this window added 8.1% in 2022; the fastest added 28.8% in 2025. This corpus records what was published, not why, so the shape is an observation rather than an explanation.

One caution about reading any single year, and it is the reason incompletely indexed venue years are dropped at all. Proceedings reach this corpus as they become available, so a year can look flat merely because part of it has not arrived yet. Before that filter existed, a single under-indexed venue was enough to turn the final year from the fastest in the window into the slowest.

Why the Raw Count Misleads

It is worth seeing how wrong the naive version would be. Read end to end, this corpus starts in the late 1990s with a handful of papers from a single venue and ends with tens of thousands from more than thirty. Plotted without care, that curve looks like an explosion in research. Some unknown part of that early slope is simply Conftrace covering more ground, and nothing in the raw count separates the two.

This is the ordinary failure mode of "research is growing" claims. A corpus that adds venues, years or document types over time will show growth whether or not the underlying field moved at all. The fix is not a better chart. It is refusing to compare two populations that are not the same population - which is why the table above spans 7 years rather than the full history, and why it drops every venue that missed even one of them.

The cost of that discipline is real: holding the set fixed throws away venues that joined recently, including some of the most active ones. So the figure describes these 21 venues and nothing wider. It is not a lower bound on the field either, because the venues left out are not a random sample of it - they are disproportionately the new and fast-growing ones.

Where the Growth Shows Up

Volume is only the outer layer. The same corpus records which topics that volume went into, and the composition moved at least as much as the size did - terms that barely existed a few years ago now carry hundreds of papers a year, while once-standard vocabulary thins out. The research trends page ranks those year by year, with a floor on the prior-year count so a term going from one paper to twenty does not outrank a genuine shift across a thousand.

You can also come at it from the venue side. Side-by-side comparison puts two venues against each other, and if you want the underlying records rather than the aggregates, every number on this page resolves to papers you can search directly.

What This Does Not Show

• Not the whole of research. This is a sample of 36 leading CS/AI venues, not a census. Work published in smaller venues, in journals outside this set, or only as preprints is not counted, and there is no reason to assume it moved the same way.
• Not submissions, and not acceptance rates. These are published papers. Conftrace holds no submission or acceptance data, so this page cannot tell you whether venues grew because more work was sent to them, because they accepted a larger share, or both.
• Not quality, impact or citations. None of that is in this corpus. More papers is a statement about volume and nothing else.
• Not a complete view of indexing gaps. A venue year is dropped when it falls below a quarter of that venue's own median. A subtler shortfall, one missing track say, clears that bar and stays in the count.
• Not the current year. The window stops at 2025 on purpose. Venues publish throughout the year, so the year in progress is always partially ingested and would read as a collapse.

The numbers on this page are read from the database when you load it, so they move as the corpus does. The date at the top is the vintage of the writing, not of the figures.