The AI That Ended a Law License.

In February 2026, Omaha attorney Greg Lake stood before the Nebraska Supreme Court to argue a divorce appeal. Of the 63 citations in his brief, 57 were defective. Twenty were hallucinated. Three cited cases that did not exist in any jurisdiction. When the court asked whether he had used AI, Lake denied it. Then he admitted it, calling the episode a grave error of judgment.

In April 2026, the Nebraska Supreme Court suspended his license indefinitely. It was the first indefinite suspension in U.S. history tied to AI hallucinations in a court filing.

His client got no do-over. Jason Regan, fighting for custody of his daughter, was left facing $52,000 in opposing counsel fees. The sanctions against his lawyer did not erase them.

Lake is not an outlier. Raja Rajan of Cherry Hill, New Jersey, a solo practitioner, was sanctioned $2,500 and then $5,000 by the same federal judge for two separate AI-tainted filings. Rajan told the Philadelphia Inquirer he could not remember which chatbot wrote the brief. He had asked a second AI to verify the first AI's citations. It confirmed them. All six were false; one cited a case that did not exist at all.

This is not a handful of careless lawyers.

Researcher Damien Charlotin maintains the most comprehensive public database of these cases. The count is not static: it passed 1,780 decisions worldwide in July 2026, and new ones are added most days, sometimes several from several different courts at once. Open the database on any given week and the most recent entries are days old, not months.

The penalty arc is steep and getting steeper. It started with a $5,000 fine in the 2023 ChatGPT case that made national news. By 2025 it was $3,000 fines for Mike Lindell's attorneys and a revoked pro hac vice admission for a Morgan & Morgan attorney whose fake citations came from the firm's own in-house AI platform. Then the first quarter of 2026 alone produced at least $145,000 in sanctions, including roughly $109,700 in combined penalties and fees in an Oregon winery dispute, believed to be the largest aggregate penalty tied to a single matter. In March 2026, the Sixth Circuit fined two attorneys $15,000 each and dismissed their case entirely. In June 2026, a federal judge in Mississippi suspended two lead attorneys from practicing in the district for two years and canceled the trial after both sides filed hallucinated citations.

“I use a legal AI tool” is not a defense.

You might assume this only happens to lawyers pasting briefs out of free ChatGPT. It doesn't.

Stanford's RegLab put the purpose-built tools under peer review. Westlaw's AI hallucinated on roughly one in three legal queries. Lexis+ AI hallucinated on roughly one in six. In early 2026, the Fifth Circuit sanctioned an attorney $2,500 for defective citations drafted with vLex and Thomson Reuters CoCounsel, commercial legal platforms. The court raised the fine because she did not accept responsibility, and the commercial pedigree of the tools bought her nothing.

Courts have been explicit: the duty to verify is non-delegable. It does not matter which tool you used, how reputable the vendor is, or who on your team ran the prompt. Your signature on the filing is the liability.

And the reason no tool can promise otherwise has now been demonstrated formally. In 2024, researchers formally proved that hallucination cannot be fully eliminated from large language models. It is not a bug awaiting a patch. It is how the technology works. Every product that generates fresh text at the moment of the question carries the risk, every single time.

So what do you actually have?

Here is the honest inventory of what a solo attorney can do today.

Manual verification. Check every citation against a primary source before filing. This works. It is also exactly the time-consuming work you adopted AI to escape, and the documented cases show that under deadline pressure, lawyers skip it. Rajan called it taking a shortcut. Multiple federal district courts now have standing orders requiring you to certify that AI-generated content was verified, so the shortcut is now also a certification violation.

Trusting the tool's own confirmation. The lawyer in the original ChatGPT case asked the chatbot whether its cases were real. It said yes. Rajan asked a second bot to check the first. It approved the fakes. AI verifying AI is the pattern that produced the sanctions, not the pattern that prevents them.

Post-hoc citation checkers. Useful for briefs. They catch fabricated citations after generation. But they only address one narrow slice of the problem: court filings. They do nothing about the other places AI now speaks for your practice.

Which brings up the exposure most solo attorneys have not thought about yet.

The sanctions wave is about briefs. But the underlying failure, an AI confidently inventing facts, follows the technology wherever it goes. And right now it is going onto law firm websites.

Solo and small firms across the country have been adding AI chatbots to capture leads and answer visitor questions after hours. Nearly all of them work the same way: retrieve some content, then have a language model generate a fresh answer on the spot. Every answer is generated live. Which means every answer can hallucinate, exactly like the tools that produced the sanctions wave.

Picture the questions a prospective client types into your site at 11pm. How much is a retainer for a custody case? What's the deadline to file after a car accident? Do I qualify for Chapter 7? Can my landlord evict me without notice?

If your chatbot invents a statute of limitations, misstates your fee structure, or confidently gives wrong jurisdictional guidance, you have a problem no citation checker will catch. Depending on the answer, you may have an unauthorized practice of law problem, a fee dispute, a malpractice allegation from someone who relied on it, or a bar complaint. The disclaimer at the bottom of the widget will not feel like much protection when the transcript is Exhibit A.

You would never let a first-year associate answer client questions unsupervised on your letterhead. That is what a generative chatbot is.

What a solo attorney actually needs.

Not a smarter model. A different architecture. One where nothing reaches a visitor unless you approved it first. Where answers come only from your documents: your fee schedule, your intake policies, your practice-area content. Where the system says so when there is no approved answer, instead of guessing, and logs the question so you can add a real one. And where every published answer has a record of who approved it and when.

That last item matters more every month. Courts and regulators across every profession are converging on the same standard: human oversight only counts if it is genuine, documented, and provable. An approval log is the difference between saying you supervise your AI and showing it.

It runs on your server with your own API key. Your documents are processed at upload; Truebe is never in the data path and never stores your client-facing content. For a profession bound by Rule 1.6 confidentiality obligations, that is not a feature. It is a requirement.

It also works inward, and for a solo attorney this may be the half you use daily. The same system runs a private research chat over the documents you upload: your case files, briefs, contracts, precedent memos, firm policies.

Ask a question and get an answer drawn only from those documents, with the source cited and the verbatim passage shown, ready to copy with its citation. No open-web guessing, no invented authority. It cannot cite a case that is not in your files, because it cannot answer from anything but your files.

The research shortcut that suspended a license in Nebraska is the exact behavior this is built to make impossible.

The bottom line.

The profession's rules have not changed. You have always been responsible for every word that goes out under your name. What changed is that a new technology now generates words under your name at scale, and it is documented, in more than 1,780 court decisions and counting, to lie confidently.

You cannot make a language model stop hallucinating entirely. Nobody can. What you can do is make sure no unapproved answer ever reaches a client or a prospect.

That is the entire product. Truebe is $49 a month, with a free tier covering five documents.