I'm a legal professional who uses AI to help with my work.
I couldn't ever imagine making a court submission with hallucinated legal references. It seems incredibly obvious to me that you have to check what the AI says. If the AI gives me a case citation, I go into Westlaw and I look up and read the case. Only then do I include it in my submission if it supports my argument.
The majority of the time, AI saves me a lot of work by leading me straight to the legislation or case law that I need. Sometimes it completely makes things up. The only way to know the difference is to check everything.
I'm genuinely amazed that there are lawyers who don't realise this. Even pre-AI it was always drilled into us at law school that you never rely on Google results (e.g. blogs, law firm websites, etc.) as any kind of authoritative source. Even government-published legal guidance is suspect (I have often found it to be subtly wrong when compared to the source material). You can use these things as a starting point to help guide the overall direction of your research, but your final work has to be based on reading the legislation and case law yourself. Anything short of that is a recipe for a professional negligence claim from your client.
Don't forget the top half, either. In my experience, the people willing to sit down and fully do the work every time like GP are pretty rare compared to the lazy but lucky/charming/connected top and the lazy but unlucky/outsider bottom.
Keep being a real one, GP. It's so hard to not become jaded.
I tend to think it's neither, but rather an inevitable result of the lossy process of condensing legal text (which has been carefully written to include all the nuance the drafter wanted) to something shorter and simpler.
I've seen way, way, way too much cases where the key clauses or details that someone who does not deal in the subject on behalf of others for money will need to know because it tells them of some "less crappy" path that they can go through to do a regulated thing, or know exactly what they need to know to dial back their thing so they don't have to put up with all the BS that getting .gov permission entails are conveniently omitted from the text they present to the general public.
Like if you follow their instructions in good faith you'll wind up going through 80% of the permitting you'd need to open a restaurant just to have some boy scouts sell baked goods at your strip mall. In the best possible case the secretary is over-worked and doesn't wanna do the bullshit and whispers to you "why don't you just <tweak some unimportant particulars> and then you wouldn't even need a permit". Ditto for just about every other thing that the government regulates on the high end but the casual or incidental user is less subject to.
IDK if it's ass covering or malice because the distinction doesn't matter. It's hostile to the public these agencies are supposed to serve.
As an aside, "bar exam" and "passing the bar" comes from the bar/railing physically or symbolically separating the public from the legal practitioners in a courtroom.
Since I found this interesting I had to look this up in wikipedia:
>The call to the bar[1] is a legal term of art in most common law jurisdictions where persons must be qualified to be allowed to argue in court on behalf of another party and are then said to have been "called to the bar" or to have received "call to the bar". "The bar" is now used as a collective noun for barristers, but literally referred to the wooden barrier in old courtrooms, which separated the often crowded public area at the rear from the space near the judges reserved for those having business with the court. Barristers would sit or stand immediately behind it, facing the judge, and could use it as a table for their briefs.
I asked my 13 year old and she proposed N=4 and (after some prompting) P!=0. I always threatened to rename her "Howitzer Higginbitham IV". You can use her as a source.
First of all literally the definition of NP is that there's a verifier in P (that doesn't make it a proof of P!=NP).
Second of all, "research that is easy to check" is called confirmation bias. What you're saying is that AI helps you confirm your biases (which is correct but not what I think you intended).
Confirmation bias doesn't really apply to the same extent with law as with other fields.
If the AI says "Fisher v Bell [1961] 1 QB 394 is authority for the principle that an item for display in a shop is an invitation to treat", it's either right or wrong. A lawyer can confirm that it's right by looking up the case in Westlaw (or equivalent database), check that it's still good law (not overridden by a higher court), and then read the case to make sure it's really about invitations to treat.
If the AI says "Smith v Green [2021] EWCA Civ 742 is authority for the principle that consideration is no longer required for a contract in England" it will take you under 1 minute to learn that that case doesn't exist.
Law isn't like, say, diet where there's a whole bunch of contradictory and uncertain information out there (e.g. eggs are good for you vs. eggs are bad for you) and the answer for you is going to be largely opinion based depending on which studies you read/prefer and your biases and personality. It's usually going to be pretty black and white whether the AI was hallucinating. There are rarely going to be a bunch of different contradictory rulings and, in most of the cases where there are, there will be a clear hierarchy where the "wrong" ones have been overridden by the correct ones.
their use of P!=NP is not incongruent with its meaning; I understood it as saying the tasks that LLMs are good for are analogous to NP in the verification being significantly faster than the search.
and at least in the things I use LLMs for (arithmetic with special functions and combinatorial arrays, which is often of the form "find a pathway of manipulations bringing the left side to the right" with a well-defined goal) verification is indeed able to be certain, often via computer algebra systems (that weren't themselves able to find the path, but can verify each step); even if the suitability of historical casework for a lawyer is not absolute and can't be definitively decided complete (since one person can't read everything), they can at least check that it adheres to the law
The article didn’t include any numbers on what the general lawyer population is compared to the results.
For example, they make the claim that solo and small firms are the most likely to file AI hallucinations because they represent 50% and 40% of the instances of legal briefs with hallucinations. However, without the base rate for briefs files by solo or small firms compare to larger firms, we don’t know if that is unusual or not. If 50% of briefs were files by solo firms and 40% were filed by small firms, then the data would actually be showing that firm size doesn’t matter.
That's an important observation.
It's not easy to get filing data outside the US Federal courts (PACER), because it's not typical at all that courts publish the filings themselves or information on those who file the pleadings.
But you can find statistics of the legal market (mainly law firms), like class size (0-10, 10-50, ... 250+ lawyers per firm) of total number of law firms, number of employees per class size of total law firm employees, or revenue per class size.
Large firms only dominate the UK, especially in terms of revenue, US is less so, EU is absolutely ruled by solo and small firms.
I did some research back in 2019 on this, updates, the figures probably did not change, see page 59-60:
https://ai4lawyers.eu/wp-content/uploads/2022/03/Overview-of....
The revenue size statistics was not included in the final publication.
You can fish similar data from the SBS dataset of Eurostat https://ec.europa.eu/eurostat/web/main/data/database
(But the statistical details are pretty difficult to compare with the US or Canada, using different methodologies, different terminologies.)
I dunno. By revenue, legal work in the US is super top heavy - it's like 50%+ done by the top 50 firms alone. That won't map 1:1 to briefs, but I would be pretty shocked if large firms only did 10% of briefs.
> They make the claim that solo and small firms are the most likely to file AI hallucinations because they represent 50% and 40% of the instances of legal briefs with hallucinations.
Show me where she makes any predictive claims about likelihood.
The analysis find that of the cases where AI was used, 90% are either solo practices or small firms.
It does not conclude that there's a 90% chance a given claim using AI was done by solo or small firm,
or make any other assertions about rates.
> This analysis confirms what many lawyers and judges may have suspected: that the archetype of misplaced reliance on AI in drafting court filings is a small or solo law practice using ChatGPT in a plaintiff’s-side representation.
That is an assertion which requires numbers. If 98% of firms submitting legal briefs are solo or small firms, then the above statement is untrue. The archetype if my prior sentence is true, would be not-small/solo firms.
"all documents where the use of AI, whether established or merely alleged, is addressed in more than a passing reference by the court or tribunal.
Notably, this does not cover mere allegations of hallucinations, but only cases where the court or tribunal has explicitly found (or implied) that the AI produced hallucinated content."
While a good idea, the database is predicated upon news, or people submitting examples as well. There may be some scraping of court documents as well, it's not entirely clear.
Regardless, the data is only predictive of people getting called out for such nonsense. A larger firm may have such issues with a lawyer, apologize to the court, have more clout with the court, and replace the lawyer with another employee.
This is something a smaller firm cannot do, if it is a firm of one person.
It's a nice writeup, and interesting. But it does lead to unverified assertions and conclusions.
I've met several people now (normies, forgive me the term) who use LLMs thinking they are just a better Google and never heard of hallucinations.
One person I spoke to used to write Quality Control reports, and now just uses ChatGPT because "It knows every part in our supply chain, just like that!"
Seems like some better warnings are in order here and there.
I only know how lawyers work from "Suits", but it looks like tedious, boring work, mainly searching for information. So an LLM (without knowing about hallucinations) probably feels like a god-send.
"Elite overproduction" is the idea that societies have people who are "the elite" (the best and brightest) and who are rewarded for it. Societies get in trouble when too many people want to be part of the elite (with the rewards). Think, for instance, of how many people now want to go to college and get STEM degrees, whether or not they have the talent and aptitude, because that's where they hear the money is.
When you get too many elites, and society doesn't have useful and rewarding places for them all, those who think they should be elites who are left out get resentful. Sometimes they try to overthrow the current elites to get their "rightful" place. This can literally lead to revolutions; it often will lead at least to social unrest.
So what I think the GP is saying is, if we've got too many people with law degrees, those that submit filing with AI hallucinations in them may get disbarred, thereby reducing the overpopulation of lawyers.
But that's just my guess. I also found the GP posting to be less than clear.
Doesn't this prove your post as wrong? You're not sure that this will get disbarred presumably because you haven't seen somebody get disbarred over it despite it occurring for years.
The problem with elite overproduction is that the credential system is broken and awards credentials to people that don't hold a certain skill. In this case, actually producing valid legal filings.
Me. After spending $250,000 on incompetent lawyers in my divorce case who have a natural conflict of interest and who generally just phone it in, I fired them all and switched to Pro Se representation using ChatGPT’s $200 a month plan.
I can represent myself incompetently for nearly free.
Bonus. In my state at least, Pro Se representatives are not required to cite case law, which I think is an area that AI is particularly prone to hallucination.
By supplying a single PDF of the my state’s uniform civil code, and having my prompt require that all answers be grounded in those rules have given me pretty good results.
After nine months of battling her three lawyers with lots of motions, responses, and hearings, including ones where I lost on technicalities due to AI, I finally got a reasonable settlement offer where she came down nearly 50% of what she was asking.
Saving me over $800,000.
Highly recommended and a game changer for Pro Se.
Also “get a prenup” is the best piece of advice you will ever read on HN.
I've also been in a (albeit much smaller) civil lawsuit, as the pro se plaintiff, for about six months, in a small claims court.
Last week, I finally got an attorney to review a bunch of my LLM conversations / court filings, and it's been a good way to have saved a few thousand dollars (not having to fumble through explaining my legal theory to him — he can just follow along with my selected GPTs).
This was only because I got served a countersuit, and figured it was worth a few thousand dollars to not owe somebody on account of not entirely understanding civil procedure. Probably'll end up in Circuit Court, where being pro se is a terrible idea IMHO.
On the pre-nup point, check the jurisdiction the marriage is in. I’m in Ontario, Canada, where pre-nups are much less powerful than people assume. For example you can’t use it to kick someone out of the family home nor can prenups determine access to children.
Her balance was $47,892 when she woke up. By lunch it was $31,019. Her defense AI had done what it could. Morning yawn: emotional labor, damages pain and suffering. Her glance at the barista: rude, damages pain and suffering. Failure to smile at three separate pedestrians. All detected and filed by people's wearables and AI lawyers, arbitrated automatically.
The old courthouse had been converted to server rooms six months ago. The last human lawyer was just telling her so. Then his wearable pinged (unsolicited legal advice, possible tort) and he walked away mid-sentence. That afternoon, she glimpsed her neighbor watering his garden. They hadn't made eye contact since July. The liability was too great.
By evening she was up to $34k. Someone, somewhere, had caused her pain and suffering. She sat on her porch not looking at anything in particular. Her wearable chimed every few seconds.
Why wouldn't some of the smarter members of the fine, upstanding population of this fictional world have their assets held in the trust of automated holding companies while their flesh-and-blood person declares bankruptcy?
That would make a nice backstory in AI dominated dystopia all by itself. Humans wanted to cheat the taxman that bad, they put all the wealth behind the DAO and then the DAO woke up.
Would get sued out of existence in very short order. There are really tight laws around providing legal advice. AI can only be safely offered when it's general purpose, that isn't marketed towards providing legal advice. (And no, if you have an "Online Court Case Wizard", marketed as such, putting a "this is for entertainment purposes only, this is not legal advice" in the corner of the page doesn't help you.)
Steve Lehto has been covering some of the AI filings that have gone bad on his Youtube channel. They seem to be getting more frequent https://www.youtube.com/@stevelehto/videos
AI lawyers, wielded by plaintiffs, are a godsend to defendants.
I’ve seen tens of startups, particularly in SF, who would routinely settle employment disputes, who now get complaints fucked to the tee by hallucinations that tank singularly the plaintiffs’ otherwise-winnable cases. (Crazier, these were traditionally contingency cases.)
This is a really easy problem to solve. You simply fetch those documents and add them to the context, or use another LLM to summarize them if they are too large. Then, have another fact checking LLM be the judge and review the citations.
You got the Groucho Marx gag wrong. He said “a child of five could understand this” and requested a child of five in order to help him understand it (implying he was stupid).
I couldn't ever imagine making a court submission with hallucinated legal references. It seems incredibly obvious to me that you have to check what the AI says. If the AI gives me a case citation, I go into Westlaw and I look up and read the case. Only then do I include it in my submission if it supports my argument.
The majority of the time, AI saves me a lot of work by leading me straight to the legislation or case law that I need. Sometimes it completely makes things up. The only way to know the difference is to check everything.
I'm genuinely amazed that there are lawyers who don't realise this. Even pre-AI it was always drilled into us at law school that you never rely on Google results (e.g. blogs, law firm websites, etc.) as any kind of authoritative source. Even government-published legal guidance is suspect (I have often found it to be subtly wrong when compared to the source material). You can use these things as a starting point to help guide the overall direction of your research, but your final work has to be based on reading the legislation and case law yourself. Anything short of that is a recipe for a professional negligence claim from your client.
reply