Egregoros · Phoenix Framework

Post

Remote status

AI6YR Ben

21w

LA Times: O.C. immigration attorneys suspended for filing briefs filled with AI-hallucinated errors

"...The ruling was clear that the discipline wasn’t issued simply for the use of AI, but “due to their filing of briefs containing multiple nonexistent cases, misattributed quotations, and gross misrepresentations of real cases, and due to their lack of candor in revealing that these fabrications and inaccuracies were the result of hallucinations by generative AI.”...

(paywall)
https://www.latimes.com/california/story/2026-06-04/suspended-o-c-immigration-attorneys-filed-briefs-filled-with-ai-slop

#legal #law #aislop #AIpocalypse

Replies

Replying to @ai6yr@m.ai6yr.org

Jim Hubbell

@Dougfir@m.ai6yr.org remote

21w

@ai6yr
Why would an attorney even want to use AI? Every lawyer i ever dealt with was paid a very handsome amount per hour. Do a job quicker and you get paid less. Do a job poorly and quicker and never get paid at all.

Replying to @Dougfir@m.ai6yr.org

AutumnOut Boy

@feld@friedcheese.us remote

21w

@Dougfir @ai6yr

> Why would an attorney even want to use AI?

According to the recent Stanford study, it's better than the average lawyer when you use models with strong reasoning capabilities (Gemini and NotebookLM).

> Perhaps most striking: professors flagged AI responses as pedagogically harmful only 3.5% of the time, compared to 12% for peer-written answers.

https://law.stanford.edu/press/ai-outperforms-law-professors-in-stanford-law-study/

Replying to @feld@friedcheese.us

Blurry Moon

@sun@shitposter.world remote

21w

@feld @Dougfir @ai6yr based on personal experience, using chatgpt and claude for legal questions has had such bad hallucinating that it was totally worthless. it straight up lies frequently.

Replying to @sun@shitposter.world

JollyWizard

@JollyWizard@poa.st remote

21w

grok hallucinated constantly and couldn't give me an accurate link to a legal citation, but chatgpt has been extremely helpful to me in self representation.

the issue is it wants to constantly moralize and pussy out, trying to tell me it knows how the judge thinks, or strategies to use procedure rather than litigate the point I'm making. Then there is also the ask your physician/ lawyer warnings.

None of that stuff is helpful, but the actual legal citations work perfectly, if I kick it in the balls hard enough.

I have to be like don't tell me how to play it, give me the information I want and actually review the documents and give me links to them. every link and analysis checks out.

Grok could only get 1 out of 10 citations correct, no matter how I argued with it.

Replying to @JollyWizard@poa.st

Orkin_Awk

@Orkin_Awk@poa.st remote

21w

@Terry

Replying to @sun@shitposter.world

AutumnOut Boy

@feld@friedcheese.us remote

21w

@sun @Dougfir @ai6yr yeah this study used only Gemini and NotebookLM and I don't know if there was anything else special about the setup.

general purpose models are stupid and we need to kill them off. narrow focused models will do so much better.

Replying to @sun@shitposter.world

mangeurdenuage 🌿

@mangeurdenuage@shitposter.world remote

21w

@sun @feld @Dougfir @ai6yr
I think the issue for the legal models is because they use internet sources rather than limit themselves to train on purely out of archives of justice cases.
It's a lot less data but it's also a lot less bullshit.

Replying to @feld@friedcheese.us

Jim Hubbell

@Dougfir@m.ai6yr.org remote

21w

@feld @ai6yr
I know nothing about how lawyers work but if I ever turned in a consulting arborist report on some tree problem that claimed fictional sources, i would have never been hired again.