The Fixer Who Trusted the Wrong Search Engine

Michael Cohen spent his career fixing problems for other people. He had a new one of his own, and it came from a search bar.

Cohen was disbarred five years ago, which means he lost access to the legal-research tools every working lawyer takes for granted. So when he wanted to build a case for ending his own court supervision early, after serving more than a year behind bars, he did what anyone without those tools now does. He went looking on the internet.

He landed on Google Bard. He believed it was a search engine. A good one, even, the kind that had worked for him before, in other contexts, for other questions. He asked it to find him case law. It did. It handed him citations, clean and specific. He handed them to his attorney, David M. Schwartz. Schwartz filed them with a federal judge.

None of the cases existed.

The motion that wasn't supposed to be complicated

The filing itself should have been routine. Cohen had already pleaded guilty in 2018 to tax evasion, campaign finance charges and lying to Congress, the case where he arranged hush money payments, $130,000 to Stormy Daniels and $150,000 to Karen McDougal, to keep both women quiet ahead of the 2016 election. He'd done his time. Schwartz's job was to argue that Cohen's court supervision should end early, on the strength of citations that would help the argument along.

Judge Jesse Furman found the citations first. He asked a lawyer to explain how court rulings that do not exist ended up quoted in a motion filed on Cohen's behalf, and what role Cohen himself had played in writing it. That question is the legal equivalent of a smoke detector going off at 3am, technically just a question, practically an accusation with better manners.

"I understood it to be a super-charged search engine"

Cohen's answer came in a declaration submitted to the judge. He said he'd found the citations by doing research through Google Bard, and that he had no idea the tool could invent cases that didn't exist. In his own words:

"As a non-lawyer, I have not kept up with emerging trends (and related risks) in legal technology and did not realize that Google Bard was a generative text service that, like Chat-GPT, could show citations and descriptions that looked real but actually were not."

And this part is the whole tragedy in one sentence:

"Instead, I understood it to be a super-charged search engine and had repeatedly used it in other contexts to (successfully) find accurate information online."

That's the trap in miniature. Google Bard rolled out earlier that year as an answer to ChatGPT, the same way Microsoft was folding ChatGPT into Bing. Both tools generate text from a prompt. They build a plausible answer word by word. Fetching a document that exists somewhere never enters into it, and a plausible legal citation looks exactly like a real one. The tool has no internal flag that lights up when it's making something up. It has succeeded, from its own point of view, the moment the sentence sounds right. Whether the case is real was never the assignment.

Cohen had used it before and gotten good results, because plenty of what these tools produce is fine, or close enough that nobody checked. That's worse than failing constantly. Constant failure would teach you to distrust it. Success most of the time just teaches you to stop looking.

The part where a lawyer was supposed to be the safety net

Here's where the story becomes about a system with no backstop.

Cohen filed nothing himself. He passed his research to Schwartz, his lawyer and longtime friend, and Schwartz put it in front of a federal judge. Schwartz said in his own declaration that he thought the draft had already been reviewed by E. Danya Perry, a former federal prosecutor also representing Cohen. He said he never checked the research himself, because he believed someone else already had.

Perry had not. She found out the citations were fake the same way the judge did, by reading the filing after it was already in, and she was not gentle about Schwartz's account of events. Her words:

"incorrect and I believe, far-fetched, as I had no involvement in any back-and-forth, not directly with Mr. Schwartz or his paralegal and not even indirectly through Mr. Cohen."

Read that chain again. Cohen assumed Bard was a search engine. Schwartz assumed Perry had reviewed the research. Perry had never seen it. Everyone in the chain was relying on someone else to be the person who actually checked, and the case citations sailed straight through every checkpoint because every checkpoint believed it was somebody else's job.

Perry reported the fake citations to the judge and to federal prosecutors herself, and she made a point of saying Cohen had engaged in no misconduct and shouldn't take collateral damage for Schwartz's misstep. Cohen, for his part, blamed Schwartz for not checking the citations before filing them, while asking the judge to go easy on him anyway, calling the whole thing an "honest mistake."

There's something almost poetic about the man who pleaded guilty to lying to Congress asking a federal judge for leniency on someone else's behalf over an honesty problem. Everyone here meant well. The cases just weren't there, and everyone had assumed somebody else had looked.

Not exactly a first

Judge Furman noted something worth sitting with: this was the second time that year a judge in the same Manhattan federal courthouse had confronted lawyers over fake AI-generated citations. In an unrelated case, two lawyers had already been fined $5,000 for citing cases invented by ChatGPT.

Twice in one year, in one courthouse, is a pattern forming in real time, and the pattern is simple: a tool that writes confidently, a person under deadline and nobody between them whose actual job is to check.

The 2am version

Cohen was resourceful, in the wrong direction. A disbarred man with no formal research access, he did the only research available to him, with a tool he genuinely believed was just a very good search engine. That belief was reasonable enough to be repeated, under oath, in a declaration to a federal judge, and it was still wrong.

Schwartz was trusting, in the wrong direction too. He assumed someone else, a former federal prosecutor no less, had already done the checking. That assumption was also reasonable, and it was also wrong, because everyone in the chain had made the same assumption about somebody else.

That's the actual lesson, and it runs past "don't use AI for research." A generative tool will hand you something citation-shaped with total confidence whether or not it's real, and somewhere between the search bar and the judge's desk, a human being has to stop and read the source rather than trust the last link in the chain or the tool that produced it. Just check.

That's the boring, unglamorous work we built the whole product around. Every citation gets checked against the live record before it goes anywhere near a filing, a term paper or a judge. The paper exists and says what you think it says. Nobody downstream finds out the hard way that it didn't. Cohen found out via a federal judge's question. Most people find out in a smaller room. Both rooms sting more than checking would have.