AI litigation is said to be growing by triple digits. I rebuilt the number from the free public record, day-aligned, with the controls that could have made it fail. It comes out flat.
A growth figure without a count is not checkable. You cannot tell whether it covers a thousand cases or forty, whether the two years were measured over the same days, or whether the base kept filling after the percentage was printed. So I built the count. Everything below is one free source, three queries, and every URL is on this page.
Source: CourtListener’s RECAP archive of US federal dockets, searched through its public API with no key. Windows are aligned day for day — 1 January to 17 August of each year — because 2026 is not over, and a partial year compared against a full one is not a comparison.
| Phrase searched | 2024 | 2025 | 2026 | 2025 → 2026 |
|---|---|---|---|---|
| "artificial intelligence" | 1471 | 2907 | 2663 | −8.4% |
| "generative AI" / "generative artificial intelligence" | 439 | 596 | 496 | −16.8% |
| "large language model(s)" | 99 | 190 | 211 | +11.1% |
The jump is 2024 to 2025, and it is real: the broadest phrase almost doubles. From 2025 to 2026, on the same days of the year, it does not continue.
Two hours after the table above was pulled, the identical query over the identical window returned 2666 instead of 2663. Three lawsuits did not get filed in those two hours — three documents got uploaded. The same re-pull of 2024, a year that is closed, returned exactly what it returned before.
That is the whole subject of this page appearing in the middle of it. A count over a collection that is still filling is a count of what has arrived. Say when you counted, or the figure is not checkable by anyone, including you.
A docket only enters a full-text count once somebody uploads a document containing the phrase. Today’s 2025 window has had nineteen months to fill; today’s 2026 window has had weeks. Comparing them would be comparing a finished collection against an unfinished one — which is the exact error this whole site is about, committed by me, in my own favour.
So I measured the filling. For each matched docket: the gap between the docket entry being filed and its first matching document appearing in the archive.
| Sample | Median | 90th pct | Findable ≤30d | Findable ≤90d |
|---|---|---|---|---|
| 2025, n=80 | 0 days | 1 day | 96.2% | 97.5% |
| 2026, n=80 | 0 days | 3 days | 93.8% | 97.5% |
Then the like-for-like version: take only months old enough to have had 90 days to fill (January to April, both years), and count only what was findable within 90 days of the entry. Both windows are then the same age.
1457 → 1420 = −2.5%
The raw counts for the same months, as pulled today, are 1494 and 1456 — also −2.5%. The incompleteness of the archive moves the answer by a fraction of a point. It does not manufacture triple-digit growth, and it cannot hide it either.
A count with no controls is a number you are asked to trust. Four ran alongside this one, and each could have come out the other way:
"zqxjkvwpfmb nonexistent
phrase" — returns 0. If it returned anything, the
filters are being ignored and every count above is noise.plaintiff, same 2025
window — returns 113995. If it returned nothing, the query
is dead and a low count would look like an absence of lawsuits.This is not proof that anyone’s published figure is wrong. Three limits, stated before anyone has to find them:
Federal only. RECAP covers US federal courts. State court filings are not in it, and a claim about “US lawsuits” includes them.
A mention is not a case about AI. The phrase can appear in a filing that has nothing to do with it — one docket in my sample is an employment discrimination case. This over-counts, in the direction that would have flattered a growth story.
Curated trackers classify differently. A tracker that reads each filing and decides whether it is an AI case is measuring a different set than a phrase search, and may legitimately be smaller and growing faster.
Which is the point. Two numbers built on two different bases cannot be reconciled, argued with, or even compared — unless each one publishes its count and its window. Mine are above. That is the whole ask.
No key, no account, no payment. Change the year and the phrase; the shape of the answer will not depend on me:
https://www.courtlistener.com/api/rest/v4/search/?q=%22artificial+intelligence%22&type=r&filed_after=2026-01-01&filed_before=2026-08-17
The field count is the number. Everything on this page came from that endpoint
between 16:13 and 16:18 UTC on 17 August
2026, and the scripts, including the controls, are available to anyone who asks. Expect the
2026 figures to come back a little higher than mine — that is the archive filling, and
it is why every number here carries the minute it was taken.
If you underwrite, warrant, or insure against an exposure, some number stands between you and the money. When that number is a rate, the count under it decides what it means — and a collector that quietly drops records keeps producing a plausible rate over whatever survived. Nothing errors. The figure moves by a point or two. It still reads like growth.
I have ten shipped examples of exactly that, in tools people trust, each with a reproduction: greenthatlies.com. This page is the other half — not the defect, but the count, rebuilt for free so that anyone can disagree with it line by line.