What AI Still Can't Tell Us
Hallucination and the Quest for Truth
On February 3, 2026, Omaha attorney Gregory Lake went before the Nebraska Supreme Court to argue an appeal in a divorce and child custody case, Prososki v. Regan, S-25-0295.
Thirty-seven seconds into the argument, Lake was interrupted by one of the justices. “Before we get into that, I’d like to ask you about your brief.”
Lake: “Of course.”
Justice: “And your brief had a number of errors in it that were submitted. Can you explain to us how that occurred?”
Lake: “Absolutely, Your Honor. I was on my 10th wedding anniversary. While flying down there, my computer broke. And I uploaded the incorrect version of my brief.”
Opposing counsel then informed the court that of the 63 references in Lake’s brief, 57 were either fabricated, contained misquotes from other cases, or misquotes from the statutes themselves.
Justice: “The brief that was submitted had misquotes from cases, fictitious cases, and misquotes from statutes. How were those all generated in your, I guess, the version that you did submit to us?”
Lake: “Sure. It was a draft. And when I... My writing process is when I’m drafting, I stick in things that I know wouldn’t pass muster.”
Justice: “The elephant in the room is whether or not you used artificial intelligence. Did you?”
Lake: “No, I did not.”
Justice: “With respect, if you didn’t use artificial intelligence, how do we end up with a citation to cases that don’t exist? I mean, it’s frankly a little hard to believe that’s just a citation error.”
Lake: “Certainly, Your Honor. And again, like I said, I was... My computer was broken.”
Later in the argument:
Justice: “With a number of mistakes and basically misleading comments that were made in the brief, why shouldn’t this appeal just be treated as frivolous?”
Lake: “Your Honor, I don’t have a great answer for that.”
At stake was the custody of the child of Lake’s client, Jason Regan. “This was supposed to be where I felt my story would be heard,” he said.
Lake isn’t the first attorney to be caught using AI in a legal brief, nor is he the last. The consequences can be severe. In April of this year, Oregon attorney Stephen Brigandi was sanctioned and fined $110,000 in the largest AI-hallucination penalty in US history. The case he represented was dismissed with prejudice, preventing it from being refiled. A website built by AI/legal Senior Research Fellow at HEC Paris, Damien Charlotin, has identified 1730 legal decisions in cases in which generative AI produced hallucinated content, with at least 3-4 cases being added each day. While the majority (59%) of these cases involve pro se (self represented) litigants, nearly 40% are cases represented by an attorney. The consequences can be life-altering for those caught in the crosshairs, like Jason Regan.
In spite of the risks, AI adoption is accelerating and with it, an implicit if tentative trust in the models, even as hallucination persists like a metastatic cancer through the body of collective knowledge. Hallucination - the production of a statistically plausible but often inaccurate response - creates pseudo-answers that, even to a trained eye, can look and feel like the real thing. And because the leading LLMs (Large Language Models) are trained on publicly-available internet data, the hallucinations themselves become part of the corpus that trains subsequent outputs, risking a cascade of compounding misinformation.
Leading frontier AI labs have taken steps, or at least a stand, to stem the tide of “slop” as it is commonly called. In September 2025 OpenAI (GPT) published a paper in which they identified the problem, if not a complete solution. At its root, according to OpenAI, the standard training and evaluation mechanisms reward confident guessing over a safer “I don’t know” response. OpenAI’s initial response has been to provide grounding options via web search as well as file search, and they have called for a new mechanism that prioritizes “I don’t know” responses over plausible guesses. Google/Gemini has also opted for the grounding method by pointing Gemini to live search results in Google, with citations to the pages from which the information was pulled. According to Google, hallucinations have been reduced by 40%. Anthropic (Claude) opted to build a grounding layer directly into Claude itself, offering connectors that point to verified sources (e.g. Westlaw, HealthEx), as well as a Citations API that ties assertions to original sources.
Though the solutions proffered by each lab are different, they’re efforts to solve the same problem. But none of them address the root: the provenance and authority of the data the models retrieve from. What’s needed is a multi-domain, canonicalized corpus, sourced through a governed methodology, that is verified before it’s ever retrieved. Where a source can’t be confirmed, the model should say so.
ATLAS is our response to that gap, and the reason we’ve devoted ourselves to this mission. AI models can tell you what’s plausible. They can’t yet tell you what’s true. We believe that distinction is worth fighting for. And with stakes so high, as Jason Regan learned, we believe everyone has a right, as well as a responsibility, to expect nothing less.
- S
Sources
Prososki v. Regan, 321 Neb. 38, No. S-25-295 (Nebraska Supreme Court, filed March 20, 2026)
Nebraska Judicial Branch — case docket (oral argument date, parties, counsel)
Nebraska Supreme Court questions attorney about AI use in court brief — WOWT (the oral argument exchange)
Nebraska Supreme Court suspends Omaha attorney over AI use — WOWT (the suspension and admission)
AI Hallucination Cases Database — Damien Charlotin (case count, jurisdictions, party breakdown)
Federal judge hands down $110K penalty against 2 lawyers for AI errors — ABA Journal (Stephen Brigandi, Oregon)
Why Language Models Hallucinate — OpenAI (Kalai, Nachum, Vempala, Zhang, Sept. 2025)
Shumailov et al., “AI models collapse when trained on recursively generated data,” Nature, July 2024. DOI: 10.1038/s41586-024-07566-y