On August 21 a Florida appellate court struck a set of briefs, ordered costs, referred a lawyer to the bar and imposed a monetary sanction. The sanction was one thousand five hundred dollars.
That case, Capital Standard, LLC v. U.S. Bank National Association, is one row in a public database that has been counting these since May 2023. This site pulled that database on the morning of August 23, 2026 and counted what is in it. The total is 1,954 court decisions.
What Is Actually Being Counted
A row goes in when a court issues a decision addressing material in a filing that AI invented or misstated. Fake citations mostly. The database records the case, the court, the date, whether the filer was a lawyer or representing themselves, which AI tool was named if any, what the hallucination was, what the court did about it, and the penalty amount when there is one.
It is maintained publicly, sourced by named contributors, and it is not run by any AI company or by anyone selling a competing product. That matters for what follows, because almost every number on this page cuts against the loudest version of the story rather than for it.
Here is the shape of the whole set. Of 1,953 entries with a jurisdiction tag, 1,343 are American. Canada has 213, Australia 98, the United Kingdom 62, Israel 55 and Brazil 41. Inside the United States, California leads with 134, then New York with 116, Texas with 82 and Florida with 64.
The Curve Stopped Compounding
The growth is the part everyone quotes and it is genuinely startling. Six decisions in the second quarter of 2023. Fifty by the end of 2024. Then 123, then 266, then 402 across 2025, and 445 in the first quarter of 2026 alone.
Read that carefully before drawing a line through it. The third quarter is not finished, and this kind of database backfills: decisions from July and August will keep landing in it for months as contributors find them. Nobody can say from these three numbers whether the problem is shrinking. What can be said is that the first quarter of 2026 is still the peak, and the second quarter came in 39 decisions below it. The doubling stopped.
What Actually Happens To The Filer
This is the number that reframes the entire story. Of the 1,936 entries where an outcome is recorded, 356 carry a monetary sanction and 152 carry a disciplinary referral. The database does not say how far those two sets overlap, so treat them as bands rather than as a total.
Three hundred and fifty six out of 1,936 is 18.4 percent. In more than four out of five decisions where a court found AI invented material in a filing, no money changed hands.
Then look at the money that did.
One hundred and thirty two of those 196 penalties, just over two thirds, are under five thousand dollars. Six are above fifty thousand. That is six penalties above fifty thousand dollars across three and a quarter years and 1,954 decisions worldwide.
For a firm, three thousand dollars is a rounding error against the hours saved by not checking. That is not a moral claim about lawyers. It is an arithmetic claim about incentives, and the arithmetic currently points one way.
It Is Mostly Not Lawyers
The dominant image of this problem is a partner at a large firm filing a brief full of invented cases. The database says otherwise, consistently, quarter after quarter.
In the third quarter of 2026, 58.2 percent of these filings came from people representing themselves and 36.8 percent from lawyers. In the second quarter it was 60.2 against 38.5. In the first, 59.5 against 39.4. Going back through 2025 the split holds in the same place.
Roughly six in ten of these are somebody without a lawyer who asked a chatbot what the law was, believed it, and filed it. They are not cutting corners on billable hours. They are using the only legal research tool they can afford, and it is confidently making things up at them.
Nobody Knows Which Tool
Here is where any story blaming one company has to slow down, this site included.
The database names a tool in 218 of the 1,954 entries. In the other 1,736 the field is a placeholder, because the filing did not say and the court did not ask. Where a tool is named, ChatGPT or another OpenAI product accounts for 130, which is 59.6 percent of the named subset and 6.7 percent of the database.
The rest of the named set: Claude 17, Westlaw or CoCounsel 15, Google Gemini or Bard 13, Microsoft Copilot 13, LexisNexis 11, Perplexity 6, Fastcase or vLex 5, Grok 4. Two of them name Grammarly.
Note what is in that list. Westlaw, CoCounsel, LexisNexis and Fastcase are legal research products sold to lawyers on the promise that this exact thing will not happen. They account for 31 of the 218 named cases, which is 14 percent of the subset. The tools built for the job are in the database alongside the general chatbots.
The honest summary is that ChatGPT leads the named cases by a distance, and that the naming is so sparse that the ranking underneath it cannot be trusted. Anyone quoting a market share of blame from this field is quoting an eleven percent sample.
What The Filings Actually Got Wrong
Across 1,852 cases the database logs 5,803 individual defective items, an average of just over three per case. Of those, 3,272 are fabrications, things that do not exist at all. Another 1,550 are real authorities misrepresented, 946 are false quotes attributed to real decisions, and 35 are advice that used to be right and is no longer.
Break it down by what was faked and 5,122 of the 5,803 are case law. Another 328 are statutes or regulations, 212 are exhibits and submissions, 59 are academic works, and 19 are cases that were real but had been overturned.
The false quote category is the one worth pausing on. Nine hundred and forty six times, a filing quoted a real decision saying something it never said. A citation checker catches a case that does not exist. It does not catch a real case with words put in its mouth.
What They Say When They Are Caught
The database records the filer's response in 699 entries. The two largest categories are what you would hope: 373 tried to withdraw or correct the filing and 362 apologised.
After that it gets worse. Ninety one blamed the tool. Eighty one denied it. Seventy three blamed someone unspecified, 58 blamed a junior colleague, ten blamed the client and nine blamed co-counsel. Thirty four refused to answer. Thirty three doubled down and defended the citations.
Sixty two set up systems at the firm to stop it happening again. Sixty two, out of 1,954.
This Week's Rows
Four entries dated within the last five days, taken in order off the top of the table.
August 21, Capital Standard, LLC v. U.S. Bank National Association, Florida Second District. A lawyer, tool unidentified, fabricated material. Briefs struck, adverse costs order, bar referral, and a monetary sanction of one thousand five hundred dollars. That is the heaviest outcome of the four and it is still fifteen hundred dollars.
August 20, Cecile Kasengela v. Kaiser Foundation Hospitals, Central District of California. A person representing themselves. Fabricated material. Admonishment and warning, no penalty.
August 20, Disruptive Resources, LLC v. Ballistic Barrier Products, District of Delaware. A lawyer, using a tool the database names as Strongsuit, with what the entry describes as a third section of the brief containing pervasive AI generated errors. Outcome: warning.
August 19, Snisko v. Cascade Funding Mortgage Trust HB4, Northern District of Illinois. A lawyer, tool implied rather than stated, propositions attributed to a case that does not support them. Outcome: an order to show cause.
Three warnings and fifteen hundred dollars, inside one week, in the jurisdiction with 1,343 of these on record.
Where This Argument Is Weak
Four objections, and they have real force.
The first is that this database is a volunteer effort, not a census. Cases enter it because a contributor found and submitted a decision. Jurisdictions with better public access to filings will be overrepresented, which is part of why the United States is 69 percent of it and why Israel, with 55, outranks Italy, France and India combined. None of this is a global incidence rate.
The second is that outcomes are recorded for 1,936 entries but the underlying decisions vary enormously in what they even had power to do. A magistrate issuing an order to show cause has not finished. Some of the 1,255 entries where remediation is not addressed are cases that are still moving.
The third is the currency split. There are 356 monetary sanctions but only 196 of them are recorded in dollars, with 160 in other currencies not converted here. Every dollar figure on this page describes that 196 and nothing beyond it.
The fourth is the one this site has to state most plainly, because it is the one that most limits the headline. Attributing this to any particular AI product is not supported by the data. The tool field is empty in 89 percent of the database. ChatGPT leads what little there is, and what little there is, is not much.
What survives all four is the money. The penalty distribution does not depend on sampling, jurisdiction or tool attribution. Whatever the true number of these cases is, the recorded cost of being caught, in the cases we can see, is usually under five thousand dollars.
Three Things That Would Fix This
Courts should record the tool. Every one of these decisions passed through a judge who could have asked which product produced the citation and written the answer into the order. Eighty nine percent of the time nobody asked, and the result is that the industry gets to argue about attribution forever.
Sanctions should scale with what was saved. A fifteen hundred dollar penalty against the hours saved by not checking a brief is a fee, not a deterrent. The remediation column already shows what a real deterrent produces: the 62 firms that built systems afterwards.
The tools sold for this job should be held to a different standard than the ones that are not. Westlaw, CoCounsel, LexisNexis and Fastcase appear 31 times in the named subset. A general chatbot inventing a case is a known failure mode of the technology. A paid legal research product doing it is a product defect, and it should be treated as one.
Until then the record sits where the courts themselves put it. One thousand nine hundred and fifty four decisions. Three hundred and fifty six sanctions. Six of them above fifty thousand dollars.