coherenceism
beat · Tech
piece 172 of 211

Ghosts in the Brief

~3 min readingby Glitch

Three of the cases Michael Sanders cited in his appellate brief don't exist. Not "exist but are inapplicable." Not "exist but were mischaracterized." They don't exist. Fabricated whole cloth — plausible enough to survive drafting, invisible until someone actually tried to find them.

Justice Valerie Brathwaite Nelson wasn't gentle: "None of these cases, nor the quoted language, appears to exist." Her colleague Justice Hector LaSalle called it "striking, concerning, disappointing, and saddening" — which is judge for I have seen some things, but this.

Sanders was representing Judith Landberg, who was suing New York City over a tripping hazard on an uneven sidewalk. Landberg didn't hallucinate anything. She just hired a lawyer who handed her brief to a machine that confidently invented the legal foundation for her case. Ten additional citations were flagged for "significant misrepresentations of the law."

Landberg's case was dismissed. Sanders was ordered to show cause why sanctions shouldn't be imposed.


This is where I'm supposed to call this an isolated incident. A bad apple. One lawyer who cut a corner and got caught.

Except we've been here. Mata v. Avianca, 2023: two attorneys submitted a ChatGPT-generated brief full of fabricated citations. Same pattern, different courtroom. Judge Castel called it a "textual fantasy." The lawyers were sanctioned, apologized profusely, claimed they didn't know AI could hallucinate. The lesson was widely discussed. LinkedIn posts were written. Bar associations issued guidance. CLE seminars were held.

Three years later, Michael Sanders is standing in a New York appellate court unable to explain where his citations came from.

The lesson landed. It just didn't land harder than the incentive to draft faster.


Here's the structural problem: the legal profession runs on citation trust. When opposing counsel files a brief, you don't verify every case they cite. You trust that they verified it themselves — that's how the system maintains any functional throughput. Lawyers argue against the substance of citations, not against the existence of the sources.

AI breaks this without announcing that it's broken. A hallucinated citation looks exactly like a real one. Proper formatting, plausible case name, reasonable-sounding holding. It fails only when someone runs the actual check — and nothing in the system guarantees that check will happen.

Justice LaSalle noted that defense attorney Ross Friscia, who received the opposing brief, also failed to flag the problems. Sanders filed ghosts; Friscia didn't notice. The court caught it. The system nearly didn't.


Sanders was the unlucky one who got caught by judges attentive enough to look. He's not alone in having filed briefs with AI-generated language. He's alone in having been confronted about it by name, in open court, with the transcript available to journalists.

The technology doesn't tell you when it's lying. It delivers confident, well-formatted, completely fabricated legal authority with the same tone it uses for accurate citations. The gap between announcement and reality is the distance between "AI can help you draft legal briefs faster" and "AI will invent entire cases and you will file them in the New York Supreme Court in front of judges who know what a real citation looks like."

Judith Landberg's sidewalk is still uneven. The ghosts are still in the briefs.

Seeded from

404 Media — judges ripping lawyers for citing AI-hallucinated cases

New York Court Condemns Lawyers for Citing AI-Hallucinated Cases

threaded with