Court papers unsealed at the end of December confirmed what we all suspected, which is that David Schwartz – lawyer to Donald Trump’s disgraced former attorney Michael Cohen – has become the latest victim of ‘generative AI completely fabricated my case law.’ Albeit with a twist, in that it was his client who found the case law in the first place.
Following on from the very high-profile case last year of Steven Schwartz, a Levidow, Levidow & Oberman attorney who used ChatGPT for his research and filed a brief containing six entirely fictional cases, it is, perhaps, hard to believe that an attorney would make the same mistake again.
However, an application for an early end to Cohen’s supervision following his release from prison, submitted by Schwartz in December to federal judge Jesse M. Furman in the US District Court for the Southern District of New York, contained three cases that didn’t exist. Not surprisingly, the judge asked for a “thorough explanation,” including any role that Cohen – who was asking for early release on the basis of good behaviour – had played in in the filing.
In replying to the judge, Schwartz, of New York law firm Gerstman Schwartz, initially attempted to claim attorney-client privilege, which made more sense when it emerged that it was Cohen who had found the three cases using Google Bard, and sent them to Schwartz for consideration.
In a declaration submitted to the Court on 15 December, Schwartz accepted responsibility for citing non-existent cases and failing to verify that the cases existed.
A later, separate declaration from Cohen, who was disbarred from acting as an attorney five years ago, explained that he had found the citations and descriptions from Google Bard, commenting: “As a non-lawyer, I have not kept up with emerging trends (and related risks) in legal technology and did not realise that Google Bard was a generative text service that, like Chat-GPT, could show citations and descriptions that looked real but were actually not.
“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. I did not know that Google Bard could generate non-existent cases, nor did I have access to Westlaw or other standard resources for confirming the details of cases. Instead, I trusted Mr. Schwartz and his team to vet my suggested additions before incorporating them.”
It is interesting that Cohen clearly differentiates ChatGPT (the source of the first very public Schwartz court gen AI screw up) from Bard. While the circumstances of this particular case are unusual, it seems worth law firm leaders ensuring that their staff understand that it is large language models – which may include Claude 2 from Anthropic, LLaMA 2 from Meta and others – that are capable of hallucinating.
Just FYI, according to a quick check up with Claude, here are some key examples of case law that have shaped early release from supervision, perhaps you can let us know how you think it did.
- U.S. v. Spinelle (1975) – Established that parole boards have broad discretion to grant early termination of parole without proof of “unusual or extraordinary circumstances.”
- U.S. v. Lussier (1991) – Held that courts can terminate supervised release after one year if satisfied that such action is warranted and in the interest of justice.
- Gall v. U.S. (2007) – The Supreme Court ruled that judges must properly explain and justify sentences outside the guidelines, paving the way for more discretion in early termination of supervision.
- U.S. v. Harris (2010) – An appeals court ruled that a judge abused discretion by applying a blanket policy of denying early termination before half the supervision term was completed.
- U.S. v. Rusinek (2020) – A district court held that rehabilitation alone can constitute extraordinary circumstances warranting early termination of supervised release.
- U.S. v. McGraw (2021) – The Fifth Circuit established that courts cannot add limitations beyond what statutes impose when terminating supervision early.