Ask a docket database how many state court cases mention artificial intelligence and you get a precise, confident, wrong answer. On August 12, 2026, I ran keyword searches for terms including “artificial intelligence,” “machine learning,” “ChatGPT,” “generative AI,” and “OpenAI” across a commercial index of state trial court dockets. The query matched 2,189 cases. Seventy-four percent of all matches on record, 1,635 of 2,205 by filing year, were filed in 2026 through August 12. The single largest concentration sits in Georgia, at 977 cases. The second most common practice area is family law.
None of those facts mean what they appear to mean. When I audited the 40 most recent matches one by one, 27 of them turned out to be opaque, sequentially numbered filings from a single Georgia courthouse with no named parties, no complaint summaries, and no case classification. Five more were routine debt collection suits. Roughly four were disputes actually about an AI system.
Here is the uncomfortable version of that finding: if raw docket counts are feeding the AI litigation explosion narrative, most of that explosion is courts filing paperwork about paperwork. The count did not surge because Americans started suing over AI. It surged because the measurement instrument changed underneath everyone counting.
The honest answer: somewhere between 100 and 2,000, depending on the ruler
Start with the counts that exist. The curated trackers, which read filings and apply editorial judgment, produce numbers between roughly one hundred and just over two hundred. One widely used tracker reports about 200 AI lawsuits across 68 defendants, pulled from PACER, state e-filing systems, and primary court records. Its own April 2026 methodology piece counted 189 verified cases and made an argument I want to extend rather than repeat: “AI lawsuit” is not a legal category, it is a topic, and every public count draws the perimeter differently.
Law professor Edward Lee’s Chat GPT Is Eating the World reported 87 United States copyright suits and 112 worldwide as of March 5, 2026. A copyright-focused tracker put the figure above 125 active or recently resolved cases as of July 2026. Narrow the perimeter to suits against the major model developers and the number drops toward a few dozen. Widen it to any complaint that names an AI product, however thin the connection, and it climbs past two hundred.
That definitional spread is real, and the trackers are honest about it. What the trackers do not capture, because they curate rather than count, is what happens when you drop editorial judgment entirely and just ask the raw docket text. That number is 2,189, and it is broken in a more interesting way than any perimeter dispute.
What raw state court dockets actually say
The methodology is simple to state. I searched a commercial index of state trial court dockets for a fixed set of AI terms across case names, party names, law firm names, and docket entry descriptions. The run date was August 12, 2026. This measures something specific: cases where AI vocabulary appears anywhere in the machine-readable court record, which is a different quantity from cases about AI.
The yearly series tells the story. The terms appear in 42 cases filed in 2023, then 194 in 2024, then 281 in 2025, then 1,635 in 2026 through August 12. That last number is 74 percent of every match ever recorded, generated in the roughly 32 weeks through August 12. A phenomenon that quadruples in 2024, grows 45 percent in 2025, and then multiplies nearly sixfold in seven months of 2026 is not describing a wave of new disputes. Litigation does not move like that. Instrumentation does.
The geographic distribution confirms the suspicion. Georgia leads with 977 matches, Texas follows with 439, then Pennsylvania at 138, California at 136, Kansas at 107, Missouri at 100. California, the state with the largest population, the deepest technology industry, and the most aggressive privacy statutes, ranks fourth. Georgia, with a quarter of California’s tech sector, holds 45 percent of the national total. Either Georgians sue over AI at twenty times the base rate, or something in Georgia’s court records generates matches that have nothing to do with disputes.
The practice area breakdown settles it. The largest bucket, at 1,009 cases, is unclassified. The second largest, at 434, is family law. Divorce and custody dockets are not where product liability suits against model developers live. They are, however, exactly where you would expect to find court-generated text about AI use in filings.
The audit: 27 of the 40 newest matches come from one Georgia courthouse
Aggregates hide mechanisms, so I pulled the 40 most recently filed matches and read what the index held on each one. The composition was decisive.
Twenty-seven of the 40 were filings from Henry County, Georgia, carrying sequential case numbers, filed in batches on the same dates, with no party names, no complaint summaries, no practice area, and no accessible detail. These are not 27 independent AI disputes. A pattern of consecutive docket numbers matching an AI keyword across an entire county’s intake is the signature of something systematic: either the court attaches a standard notice referencing artificial intelligence to every new civil filing, or an indexing pipeline stamps AI-related text across the county’s records. I could not determine which from the aggregate data alone, and I flag that gap explicitly in the limitations below. What the pattern rules out is the naive reading. Whatever mechanism produced these records appears to operate systematically across the sampled Henry County batch, which means the keyword match carries no information about the underlying disputes.
Five more of the 40 were debt collection cases: credit card defaults and auto finance suits from Texas and Florida, several filed by the same plaintiff-side attorney. Nothing in a collections complaint concerns machine learning. The plausible source is boilerplate, either in the filings themselves or in attached certifications, and collections practices generate filings at industrial volume, so a single template change propagates instantly across hundreds of dockets.
That left roughly four cases in the sample that a reasonable person would call AI litigation. Four out of forty. Applied naively to the 1,635 matches filed in 2026, a ten percent subject-matter rate would suggest the true count of new AI disputes in state courts this year sits in the low hundreds, not the thousands. I want to be precise about the epistemics here: 40 recent cases is a recency sample, not a stratified one, and the Henry County batches cluster in time, so the sample overweights them. The ten percent figure is an observation about the newest matches, not a population estimate. But the direction of the correction is not in doubt. The raw count overstates AI disputes by a large multiple, and the overstatement is growing.
The instrument changed: courts started writing about AI in 2024
Why would court records suddenly fill with AI vocabulary starting in 2024 and exploding in 2026? The leading candidate mechanism: the courts themselves started producing it, on a schedule that matches the curve almost exactly.
The precipitating event is well documented. In June 2023, Judge P. Kevin Castel of the Southern District of New York fined two lawyers and their firm 5,000 dollars in Mata v. Avianca after they filed a brief citing six cases that did not exist, invented by ChatGPT. Days before that sanction landed, Judge Brantley Starr in the Northern District of Texas issued the first widely reported standing order requiring filers to certify that no generative AI drafted their filings, or that a human verified anything that was AI drafted. Judges across the country copied the pattern within months.
Then the state court systems moved from individual orders to institutional policy. The Delaware Supreme Court adopted an interim policy on generative AI use by judges and court personnel on October 22, 2024. The Illinois Supreme Court announced its AI policy in December 2024, effective January 1, 2025, and notably concluded that disclosure of AI use should not be required in a pleading. And Georgia, the state carrying 45 percent of the keyword matches, established the Judicial Council of Georgia Ad Hoc Committee on Artificial Intelligence and the Courts in 2024, chaired by Justice Andrew A. Pinson in partnership with the National Center for State Courts. The committee submitted its report, Artificial Intelligence and Georgia’s Courts, on July 3, 2025, recommending statewide governance, education, and new court rules on a three-year rollout.
Every one of these interventions injects AI vocabulary into court workflows. Standing orders get docketed. Certifications get attached to filings. Notices get stamped onto intake paperwork. Clerks add entries. None of it requires a single litigant to have a dispute about AI, and all of it matches a keyword search for artificial intelligence. The heterogeneity matters too: Illinois rejected mandatory disclosure while individual judges elsewhere require certification on every filing, which is exactly why contamination varies so wildly by jurisdiction. A county under a certification regime lights up the index. A county under the Illinois approach stays dark. The keyword map of AI litigation is, to a first approximation, a map of local court administration choices.
Epidemiologists know this failure mode by heart. When diagnostic codes change, disease incidence appears to jump, and every naive trend analysis breaks until someone separates the coding change from the biology. Court dockets in 2026 are living through their coding change.
There is also a genuine phenomenon feeding the docket text, and it deserves its own number. Legal researcher Damien Charlotin of HEC Paris maintains the AI Hallucination Cases database, which tracks court decisions where a party relied on AI-fabricated citations or content. Reported at 719 entries in January 2026, it stands at 1,870 identified cases as of this writing, more than doubling in seven months, with sanctions climbing from Castel’s 5,000 dollars to five-figure penalties and the first bar suspensions. Every one of those decisions puts AI language into a docket, in cases that are about car accidents and contract breaches and custody, not about AI. AI as a litigation tool, and courts reacting to it, now generates an order of magnitude more court text than AI as a litigation subject.
What the real cases look like
Strip out the boilerplate and the sanctions traffic, and the residue is where state AI jurisprudence is actually forming. Four suits from the last two weeks of the sample show the shape of it.
In Thompson v. SoundHound AI, filed July 30, 2026 in Alameda County Superior Court, the complaint alleges that an AI voice ordering system intercepted and recorded consumers’ restaurant phone orders without consent, pleading violations of California’s wiretap statutes. Whether a speech recognition pipeline “intercepts” a call within the meaning of a 1967 statute is a question no appellate court has answered, and it governs every AI phone agent deployed in California.
In Maier and Howard v. Omni Hotels Management Corporation, filed August 4, 2026 in Los Angeles County Superior Court, two models allege that a photoshoot they licensed for limited use was altered with AI and deployed in holiday advertising without consent, pleading statutory and common law misappropriation of likeness. The open question is what consent to a photograph means once generative editing can produce images the subject never posed for.
In Beltran v. Conservative Solutions for America, case number 26-CA-008578, filed August 7, 2026 in Hillsborough County, Florida, a former state legislator’s congressional campaign seeks declaratory and injunctive relief over campaign ads he alleges were fabricated with AI to depict him holding a protest sign, claiming statutory disclosure violations. It is among the first civil tests of synthetic media disclosure law in a political campaign.
And in Leon County, Florida, a filing docketed August 11, 2026 in Mall v. OpenAI touches products liability claims connected to the April 2025 shooting at Florida State University, a tragedy that gets exactly one sentence here because the legal question, whether a language model’s output can constitute a defective product, is the part that will outlive the caption. The docket entry I sampled concerned a fee arrangement motion, so the underlying claims deserve a careful read before anyone characterizes them, and they will get one.
These four cases will each receive a full technical examination in this publication’s new Legal section, because they share a property the boilerplate matches lack: their outcomes turn on technical facts about how the systems work, facts that the complaints assert and that discovery will test.
Why a broken count matters
Numbers like “AI lawsuits up 600 percent” are load bearing in places that matter. Insurance underwriters price AI liability coverage against litigation frequency. Policy staff cite filing trends in testimony about whether new statutes are needed. Compliance vendors sell monitoring products against the fear the trend line generates. Boards allocate legal budgets against it. If the trend line is mostly an artifact of court administration, every one of those decisions is calibrated against noise.
This is the same failure mode this publication has documented in AI evaluation itself. A model’s score can swing twenty points based on the evaluation setup alone, with no change in the model. A flagship system’s hallucination rate can fall 38 points without the model learning anything, because the measurement moved. Litigation counting in 2026 has the identical structure: the instrument and the phenomenon changed at the same time, and the headline number silently conflates them. Anyone who has cleaned a dataset knows the rule. Before you trust a trend, ask what changed about the collection.
What this count cannot tell you
Honest limits, stated plainly. First, the 40-case audit is a recency sample. The Henry County batches cluster in the newest filings, so the sample overrepresents them, and the ten percent subject-matter rate should not be projected onto all 2,189 matches. A stratified sample across years and states would produce a defensible population estimate. This piece does not have one yet.
Second, the keyword count is simultaneously an overcount and an undercount. It overcounts because of the boilerplate mechanisms described above. It undercounts because a lawsuit about a discriminatory screening algorithm that never uses the words “artificial intelligence” in its docket text is invisible to the query. Both errors are large and they do not cancel.
Third, I could not confirm the specific Henry County mechanism from index data alone. Distinguishing a court-attached AI notice from an indexing artifact requires reading the underlying documents for a sample of those stub dockets. Until that read happens, the boilerplate explanation is the leading hypothesis consistent with the observed pattern, not an established fact.
Fourth, scope. This measures state trial courts in covered jurisdictions. Federal courts, where most of the marquee AI copyright litigation lives, are outside the instrument entirely, and coverage depth varies by state and county. A small bookkeeping note for anyone reproducing the numbers: 2,189 dockets matched the query, while the filing-year series totals 2,205 because a handful of records carry clerically misdated filing dates.
What happens next
The contamination will get worse before it gets better. Georgia’s committee report recommends new rules and statewide governance through 2028. Other states are drafting policies now. Each adoption adds AI vocabulary to another jurisdiction’s docket stream, which means naive keyword counts will keep accelerating regardless of what litigants do. Expect the gap between curated counts (about two hundred) and raw counts (thousands) to widen every quarter, and expect the raw numbers to keep showing up in marketing material precisely because they are bigger.
There is a constructive inversion available. If keyword matches track court AI policy adoption more than disputes, then the keyword series is a usable sensor for institutional adoption: which counties adopted AI orders, when, and at what pace, measured passively from public dockets. Nobody publishes that curve today. The contamination is itself a dataset.
For counting actual AI litigation, the method that survives contact with this data is the one the careful trackers already use, plus one addition: anchor on doctrine rather than keywords, read the filings, and audit the technical claims inside them against how the systems actually work. That last step is where this publication’s Legal coverage will live. Each case that passes a strict selection test, a first-impression legal question, a generalizable answer, a question that outlives the case, and an outcome that turns on a technical fact, gets a full examination: every claim the complaint makes about the AI system graded against documented system behavior, and the specific evidence that would settle each one, named in advance.
Frequently asked questions
How many AI lawsuits are there in 2026?
Between roughly 100 and 200 by curated counts that read filings and apply editorial judgment, with copyright-focused counts near 125 and the widest tracker near 200. Raw keyword searches of court dockets return far larger numbers, 2,189 in state trial courts alone, but most of those matches reflect court boilerplate and AI-related administrative text rather than disputes about AI.
Why do AI lawsuit counts disagree so much?
Two reasons stack. Different trackers draw different perimeters: copyright only, versus any suit naming an AI company, versus any complaint with an AI angle. And keyword-based counts are contaminated by the courts themselves, because standing orders, certifications, and policies about AI use in filings inject AI vocabulary into dockets of cases that have nothing to do with AI.
Are most court cases that mention AI actually about AI?
No. In an audit of the 40 most recent state docket matches, 27 were administrative stub filings from a single Georgia county, five were routine debt collection suits, and roughly four were genuine AI disputes. The most common way AI appears in a courtroom today is as a drafting tool and a sanctions problem, not as the subject of the suit. A database tracking AI-fabricated citations in filings has identified 1,870 such decisions worldwide, roughly nine times the size of the largest curated count of suits about AI.
What is a judicial standing order on AI?
A rule issued by a judge or court system governing generative AI use in that forum. The first widely reported example came from the Northern District of Texas in 2023, requiring filers to certify that AI-drafted content was human-verified. State systems followed with institutional policies: Delaware adopted interim guidance for judges and staff in October 2024, Illinois adopted a statewide policy effective January 2025 that declines to require disclosure, and Georgia’s judiciary published a full governance framework in July 2025.
Methodology: figures marked as docket counts come from keyword searches of a commercial index of state trial court dockets, run on August 12, 2026, using the query terms listed above. Case descriptions summarize allegations in complaints. Defendants had not answered as of the retrieval date, and allegations are exactly that. Nothing here is legal advice.