site banner

Friday Fun Thread for July 4, 2025

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

1
Jump in the discussion.

No email address required.

I mean, yes, but the hallucination problem of putting in wrong cases and statutes is utterly disqualifying in advanced legal writing. Citing to a nonexistent case or statute compromises the entire brief or argument. A decent first year associate might misinterpret a statute or case, or miss that the case was overturned, but they wouldn't make up cases from whole cloth and build their arguments off those.

For a lot of tasks, you just need to go through and proofread or fix up the places where it filled in basic info that it obviously didn't have.

But citing a case that doesn't exist to build an argument is like asking it to design a bridge and it get the tensile strength of steel completely wrong, or perhaps it makes up a type of material that doesn't exist and hallucinates its properties as part of the specifications.

And maybe it does that, I don't know. But there's literally no reason for it to be doing that, either, when there is definitive information, easily available for reference. Its information it should never get wrong, in practice.

And it really shouldn't be hard to fix, the caselaw and statutes are already simple to look up. Just teach the thing to use WestLaw.

So I do expect them to solve that particular class of hallucinations pretty handily, even if it will still completely fudge its outputs when it doesn't have an easy way to check.

But there's literally no reason for it to be doing that, either, when there is definitive information, easily available for reference. Its information it should never get wrong, in practice.

Yeah this is something that gets me about the frequent code-based hallucinations too. The things will make up non-existent APIs when the reference docs are right there. It does seem like it wouldn't be hard to hook up a function that checks "does this actually exist". I assume it must not actually be that simple, or they would've done it by now. But we'll see what they can do in the future.

There's some technical parts to how LLMs specifically work that make it a lot harder to police hallucination than to improve produce a compelling argument, for the same reason that they're bad at multiplication and great at symbolic reference work. A lot of LLMs can already use WestLaw and do a pretty good job of summarizing it... at the cost of it trying to cite a state law I specifically didn't ask about.

It's possible that hallucination will be absolutely impossible to completely solve, but either way I expect these machines to become better at presenting compelling arguments faster than I expect them to be good researchers, with all the good and ill that implies. Do lawyers value honesty more than persuasion?

I mean, yes, but the hallucination problem of putting in wrong cases and statutes is utterly disqualifying in advanced legal writing.

One would think! And yet.

This is my biggest problem with rlhf aside from my free speech bullshit - due to the way llms work, rlhf means hallucination is impossible to solve - it is baked in.