In late 2025, I filed a brief that contained citation errors generated by an AI tool. One case did not exist. Eight quotations, drawn from a real decision, were inaccurate. The court issued an Order to Show Cause, and I was sanctioned.

I am writing about it on my own website, under my own name, because I think the profession is having the wrong conversation about AI — and because I’d rather you learn this from someone who has been through it than from a headline.

Here is what I will not do: blame the tool. The brief went out under my signature. Rule 11 doesn’t care that I was working against a deadline, and it doesn’t care which software I used. The duty to verify every citation and every quotation before it reaches a court is non-delegable. It was my obligation, I did not meet it, and I told the court exactly that. Most important I informed my client. I implemented a verification protocol the same week.

I am not an outlier, and that’s exactly the problem. A researcher in Paris, Damien Charlotin, maintains the most comprehensive public database of these incidents. As of this spring it catalogs roughly 1,500 court decisions worldwide addressing AIhallucinated content and this list continues to grow. What struck me most wasn’t the number. It was who’s in it. This started as a pro se problem: in 2023, most of the caught cases involved people representing themselves. By 2025, licensed attorneys were at fault in over half of new cases. And it isn’t just lawyers anymore — government lawyers on both sides, expert witnesses, and even judges and their law clerks have submitted filings and opinions containing fabricated citations. No one is shielded by their credentials. The technology does not care how many years you have practiced.

The temptation with AI is to treat it as a research associate. It is not. A junior associate who hands you a fabricated case has committed misconduct. An AI tool that hands you a fabricated case has followed its programming — to generate fluent, plausible, confident text based on probabilities. Let the word probability sit with you for a second. AI predicts what comes next in a sentence based on probable outcomes; a second-year associate does not predict, they write with an eye towards the truth.

The 2025 Stanford study on legal research tools found hallucination rates between 17% and 33% even in the retrieval-based systems marketed as “hallucination-free.” General chatbots run far higher. The error rate is a feature of the technology, not a bug you can train your way out of by being careful.

So the question is not whether to use these tools. Used well, they save real time. The question is what discipline you wrap around them. Here is the protocol I now follow, and that I would urge on any firm:

Pull every case before you cite it. Not the AI’s summary of the case — the actual decision, in a real reporter or on the court’s own site. If you can not find it in 60 seconds, assume it doesn’t exist. Frankly, I am not sure it is a good idea to depend on even Westlaw or LEXIS-NEXIS.
Read the quote against the source. The dangerous error is not the fake case; those are easy to catch. It is the real case, accurately named, with a quotation that’s been subtly altered or pulled from dicta and dressed up as a holding. That requires opening the opinion and reading the language yourself.
Keep a verification log. For every AI-assisted filing, document what you checked and when. If something slips through anyway, a contemporaneous record of diligence is the difference between a mistake and a pattern.
Disclose early if you find an error. Courts have been consistent on this: the attorneys who get hammered are the ones who deny, minimize, or blame staff. The ones who come forward and own it fare far better. Honesty after the fact is the cheapest insurance you’ll ever buy.
Treat AI output as a draft, never a result. The tool drafts. You are the lawyer. Nothing changes that.

I represent workers against some of the largest companies in the country. I am not giving up the efficiencies these tools offer. But I have stopped pretending the technology will protect me from myself. The verification is the work. There’s no version of competent practice that skips it.