I would love to see these benchmarks split up across "generic" tasks and novel tasks.
For example, redlining an OEM agreement with my vendor. They pretty much all use the same template, and the objections are always the same. I tried using Claude first, and even my lawyer was impressed that it found things they missed. Then I used Claude to redline that type of contract and the other side said "your lawyer was surprisingly thorough". I had my lawyer do the final check, but saved a bunch of money doing the first three rounds with Claude instead. And my lawyer basically just said "looks good to me" by the time they reviewed it.
Or even writing a simple will or estate document.
It seems like LLMs are great for the kind of generic work you'd usually pawn off to a new associate or paralegal, but not for the important things that partners usually do.
> It seems like LLMs are great for the kind of generic work you'd usually pawn off to a new associate or paralegal
This seems to be the case with most industries that LLMs are impacting, coding, data input, legal etc.
Whilst it's great to get the time saving from not delegating to a junior (LLMs are much faster at the basic level 1 work) there's going to be this big knowledge gap that you really only gain from actually doing the mundane boring shit.
A lot of this _should_ be automated away, but we really need to think about how the next group of people are going to learn the skills and nuance to applying them. I don't really have an answer to that one other than still giving that work to junior staff, but they will then just utilise an LLM to help complete the job, and rightfully so, I'd have done exactly the same thing.
I think OpenAI tried doing this with some early chatgpt settings with "learning mode" or whatever that is. But an LLM just telling you what to do will never teach someone, there was something about having to go out and hunt for the answer through the sea of crap responses or wrong SO questions that really taught you something. How do you replicate that learning with an LLM and how do you apply that to different industries like Law?
> but we really need to think about how the next group of people are going to learn the skills and nuance to applying them
That's what school is for. For the last 150 years we've automated a bunch of things, and then we send people to school to learn how those automations work, so they can make better automations.
Maybe law school will have to be a year longer. Or maybe some forward thinking firms will have new associates check the work of the LLM to gain that experience.
> It seems like LLMs are great for the kind of generic work you'd usually pawn off to a new associate or paralegal, but not for the important things that partners usually do.
This seems to be the common thread everywhere LLMs end up being useful. As the sole software developer? Pure trash, but as "tons of juniors we can shuffle boring tasks at": great. As the sole designer and maintainer of a website? It'll look like 150% slop and no will buy anything, but as a designers friend to create what they dictate: even a designer can throw together a good looking website quickly. Etc etc.
I’m not sure if this is the right solution, but we clearly need better tools for the legal space. Anthropic's recent policy update makes this abundantly clear:
“When our products are used in ways that can affect someone's health, legal rights, finances, livelihood, or access to essential services, we require there to be a qualified ‘human in the loop’ (that is, someone who has the authority to review and, if necessary, change Claude’s recommendations), and for the individual affected to be told that AI was used.”
Something as common as a family law case has the potential to affect all of the above. Now imagine being a petitioner, respondent, paralegal, or other 3rd party using Claude to review a draft document only to find out that Anthropic has sent your entire file to opposing counsel.
It sounds egregious, but Anthropic is clearly stating they will use their own judgment to control what people can do with its product, and that they will treat customer data as if it's theirs to disseminate. So there's a massive opportunity for an unbiased tool willing to fill this void.
Marketing brochure for a fine tuned Qwen (which they don't mention), the model discussed here is likely better. I would not trust these guys to deliver high quality solutions given a friend's experience there, but you do you I guess.
Not sure how good the Harvey benchmark is. I recently looked into traces from an agent solving their firm knowledge task subset and found multiple problematic tasks and rubrics.
Harvey bench is probably one of the worse benchmarks out there. Agents as judges, vast scope, some questionable problem/solution pairs. Tricky domain TBF.
If you actually did the tasks in the way they deem correct, at least for immigration and white collar criminal defense, you will get detaind/arrested, and removed in the first case, and the second case you will snitch. It assumes that the state is correct. In reality the state, especially in the initial phase, are never correct on all aspects. You also can't only plan on snitching because what evidence exists and what evidence is admissible and what evidence is admissible but can be used in your advantage are all unknowns. Half of the bench tests on "covering your own ass" basically, and also, seems to have not taken the Loper Bright decision into account, which changes administrative law and how it is adjudicated and appealed entirely, for good or ill.
I don't know the other fields well enough, but if all your clients end up needlessly detained or admit to things they have a valid defense for, eventually people will notice. It's not exactly the two major sins that will definitely get you disbarred - stealing money and screwing client while representing them - but it's definitely at least incompetent. They seem to not understand that common law jurisdictions are fundamentally different from civil ones, and also, there's quite a bit of difference between how the system evolved in each country. Or that in China what's printed is basically more like a manifesto, since it's almost entirely run by the bureaucracy and personal relationships, and lawyers that aren't foreign-trained and work for foreign firms aren't considered prestigious since the courts are essentially entirely for show.
They need to get like, one jurisdiction, one area in that jurisdiction, and specifically the kind of sub-sections in that area down pat. But they can't train on real material so they will never be able to do so. Their banner testimonials come from non-lawyers and that says it all.
This is why I severely dislike using all pass rate. A model could be at like 93%, meeting practically every requirement except the broken ones and all pass would give it like 20%.
As an attorney, I am not particularly impressed about a 20% pass rate on a benchmark. The name of the game for litigation practice is about being prepared for novel situations, not relitigating past ones.
It is amusing how many people are arguing what litigation is here, probably none of these people have a law education.
Imagine a bunch of lawyers arguing what programming is. Even a bunch of programmers arguing what programming is abnoxious at worst and a mental exercise at best.
That's how I feel reading HN's opinions on the recent math results. <5% of the responses even understand the point they're supposed to be arguing. AI might take over mathematics. It might not. The discussion regarding that topic is almost devoid of any relevant information though.
> probably none of these people have a law education.
Domain knowledge is informative sometimes but mainly overrated: the success of LLMs proves just that. What most lawyers do is fairly standardized and routine, efforts to invoke mystery cases or "novel" is an attempt at resisting the inevitable. The "the only British food in Palo Alto" equivalent of unique skills.
To be fair to them, on this particular benchmark (Harvey's Legal Agent) 20% puts it almost top of the pack of _any_ model, and certainly top for any remotely in its price class.
Wouldn't be surprised, if a human attorney would get 5%, while copy-pasting ChatGPT answers, report in the results a week past the deadline, with a $12,500 retainer charged upfront.
Totally fair. However, even if the LLM would come to the same conclusion and produce the same content that a competent attorney would, the attorney must still thoroughly vet the work for accuracy. So there is potentially time saved in research and drafting, but it’s not the miracle some might characterize it as.
An 80% miss rate on a benchmark suggests to me that automation has a very long way to go, even for the drafting and research phases.
What are you talking about? 90% of litigation is convincing the judge that existing case law that clearly applies to your case does apply to your case.
I'm guessing what they're saying is that nothing goes to litigation if it's a clear pattern-match to existing cases, so actual litigation practice is only novel cases.
Isn't that just in the US/Commonwealth judicial system? I could be completely wrong on how others, like non-adversarial judicial systems, work and/or how pervasive the US system is or is not.
EDIT: specifically around the use of judicial precedent in case-law.
90% of legal work where there is a plaintiff and defendant is pressuring a settlement. My wife, an attorney, does not consider that to be "litigation".
As a former litigator, I would consider that to be litigation the moment a complaint is filed in court, regardless of whether the dispute is settled without trial.
But, for your sake, it's probably best that you accept your wife's definition.
And most of the leverage you have in negotiating a settlement is by referencing prior case law. If something is novel, there's too much variance to get a good settlement.
How much novelty can there be?
If attorneys were as diligent and generous with “open sourcing” their data as SWEs had been, could LLMs benefit from all that training data?
Law and code are only superficially similar as law always requires human interpretation plus law is an interface to reality. Code can stay contained in its nice, little, abstract box.
So every lawsuit dispute, divorce, traffic injury, etc… is a completely novel case that requires depths of creativity far beyond simple CRUD apps or solving 377 open math problems.
Yeah agreed, especially for legal and healthcare models.
[edit] I'd go one step further too and say the training data should be explicitly requestable too so third parties can verify (freedom of information act styles).
"Freakazoid has ... access to all of the Internet's knowledge. He cannot fly, but often runs around with his arms extended and making whooshing sounds with his mouth, pretending he is."
Curious what domain is next, GLM-Science? And when will it be profitable to drill down to even more domain specific models like GLM-Economic-Geography or GPT-Aquaculture?
Imo, we are going to see a lot more specialised models in the near future as pushing the frontier gets harder. Qwen 3.8 27B for example is a pretty rubbish chat and general world knowledge model, but is stellar at coding (for it's size). Its small enough anyone with a modest gaming PC (16GB vram and 32GB of DDR5) can run it at speeds good enough for overnight coding sessions with an PR to review at the end of it.
Eventually, I hope we'll see many more specialised small models that run nicely on consumer and prosumer hardware. Like LLMs specific to income tax law for UK residents, or LLMs specific to building power delivery circuits for PD USB C devices. There probably is a point where the niche-ness does become to niche to justify the post training, but again, small models means much less training required.
Kind of begs the question if this even needs new/better data? Could you take the foundation model that's trained on everything, and then run a final fine tuning training run on just the set of foundation training data that you want the model to specialise in? Eg, extract just the training data that is about UK income tax law and then run a few epochos on just that? Any ML people know if this is a thing or if not, why it doesnt work?
Sort of ironic, in Google Chrome, when attempting to visit that legal z.ai site I get:
Your connection is not private
Attackers might be trying to steal your information from z.ai (for example, passwords, messages, or credit cards). Learn more about this warning
net::ERR_CERT_COMMON_NAME_INVALID
If he is on a corporate device, he has a corporate-issued MITM certificate, which allow third party "security" companies to read, store and analyze his complete network traffic in plain text.
Wait, so this not an open weight model, correct? Only available via z.ai?
So, if a US or EU person shares their private legal info here, it's processed and stored by an adversarial country that has no extradition treaty with their country?
I prefer all my data be in the hands of foriegn entities, they cant do anything to me. Id much rather Huawei spy on me than Meta (if I have no choice other than to pick one) There's literally no consequence of Huawei having my data. Unless you're a VIP, a chinese entity storing your data is probably better than a US company, as a westerner. It wont be used to manipulate you and will generally be used less adversarially against you.
I think part of the reason there's so much propaganda around chinese tech is to scare you away from possibly using them, not for your own safety, but so Meta/Google/MS ect dont lose access to the ability to spy on every aspect of your life.
I love how we have to pretend like China is some scary boogeyman because people on screens say they are.
What would prevent a Chinese company from selling your legal data to say, your legal adversary via data brokers, in the USA? People everywhere like to make money.
> I love how we have to pretend like China is some scary boogeyman because people on screens say they are.
I think you've lost the plot a bit on this one.
I'm scared of China because of their lack of democracy and the terrifying stuff they do to their own citizenry, which presumably they'd very much like to expand.
There's no digital attorney-client privilege. If a lawyer uses LLM's on your behalf, the communications are protected. If you use the same LLM's on your own behalf, the chats are not protected.
DOJ could cut Z.ai off from US-based payment systems (Mastercard/Visa) if they don't comply with subpoenas.
They could, but generally it is expected that before a subpoena is complied with it is checked to be compatible with the laws of the country it’s in, and requires cooperation from that country. Most often they are ignored if there’s not a presence/entity/assets in the US.
It’s also hard because no individual case would likely justify a move like that. It would have to be the sum of many, many cases.
You got it wrong. If courts and attorneys in the US can't run "discovery" on Z.ai servers or compel them to do anything, and if Chinese authorities don't cooperate with the US legal system as a matter of principle, then what you functionally have is super privilege.
They'll just run discovery on you. If you try to respond to a subpoena with "the documents are outside the country and I won't give them to you," the judge will throw you in jail until you do.
You're right on this point. But there are ways to build interfaces that sidestep this.
Run a sanitized VDI with no local storage --> offshore independent entity (e.g. a Hong Kong Company) that holds the AI service contract and API license --> GLM-Legal.
Yeah sure but your clever sidestep is just "hiding evidence from the court." The kind of person that feels the need to have a secret AI lawyer is also probably going to just tank their case in another way.
NATO allies can't even get US soldiers to trial. In Romania we've had cases of manslaughter and the US just shipped them back ASAP and at worst they were discharged dishonorably.
For reference, that specific case is from 2004. Christopher Van Goethem was Marine security ambassadorial staff and was convicted in a military court martial for obstruction and making false statements. The whole case was porrly handled and the Romanians are right to be pissed. A key piece of evidence that would have landed him in jail for 10 years was the blood alcohol draw but the Marines slow walked it in Romania and he didn't show as intoxicated even though he admitted to drinking that night.
When I first tried using IDE AI coding assistants a bit over a year ago, they would frequently fail to read files and completely hallucinate the code I asked it to analyze or edit. And even when it did successfully read my code, it would often hallucinate the API of other frameworks.
I don’t think hallucinations are a major issue with agentic coding a year later? At least, not for me.
Legal writing does have the problem that there is no such thing as a correct argument that a goal-driven loop can target. But basic things like “does this case even exist” seem solvable with research agents, and even harder problems like “does this case support the argument being made” seem conceptually very similar to questions I now ask Claude daily.
Yes, mechanical question of "does this exist" is trivial to solve. Whether the research is actually aligned with the user intent / the AI makes claims that are substantiated by the research at hand are much harder problems.
An issue I come across is the attributed failures of AI of a first time use of AI.
Non-deterministic software can behave orders of magnitude different based on how input is structured.
This factor is largely missed by bringing a deterministic mindset to trying AI.
This means the same problem solution can work for one person reliably, and not another and both are true.
Assertions that begin with a deterministic assessment using a snapshot in time of both the user and models capabilities are a new feature of this new category of software that will have to be better caught and expressed.
I provided some very light guidance to a couple of masters students who were working on legal citation hallucination detection last summer. Their summer project is here: https://github.com/gt-csse/mellea-lrc
If anyone here is looking for engineers to work on legal AI products, both of the students who worked on this project come highly recommended.
To your second point, it seems like every time a significant practical limitation is identified, it goes through a familiar pattern. It gets a news cycle, and then it gets a wave of confidently incorrect opinionating, and the problem gets treated as if it's a permanently built in problem that's baked into the foundation of AI. And then time passes and the issue is either mitigated or circumvented entirely.
Hallucinations are one, "AI can't scale anymore" is another (remember that one?), limited context windows, strawberry problems, glue on pizza, sycophancy.
I don't want to say there's no problems, there are very serious ones. User psychosis, economic harms, alignment, cost of computer components, etc. But there's a lot of false positives that get turned into Ed Zitron articles along the way.
So many of these we likely using Google to search then found AI mode and used that. Which is to say, they were (and still are!) using Google Gemini 1.5 from many years ago.
A lot of the examples they show seem to be focused on common law countries. The US and UK are where most of the money is to be made in the legal industry, but I am curious to see what ripple effects we get from people attempting to use AI tools geared toward US/UK-style legal systems in other kinds of systems.
The usual concern is attorneys, but judges are probably even more overloaded these days. Can this help judges as well? If so, would this ultimately lead to better decisions?
The legal space understands well how to tier services based on who has the liability. E.g. trainees and paralegals whose work is filtered through the lawyer who ultimately puts their name on what gets submitted to courts.
I'm sure we'll continue to see the occasional person who fails to understand they themselves will be held liable if they submit work done by an AI without vetting it, but it's not like they don't have ways of dealing with this.
There should be laws about disclosure on the part of attorneys about their use of AI. I know they're already using it, but spending years or decades in prison because Lionel Hutz, Esq. did something stupid at a prompt sounds unappealing to me.
I think they should be somehow restricted to using non-generative models (or at least for the user facing output). Something that would only identify potentially relevant case law from an existing set, to accelerate the search process.
Common law in general seems essentially impossible to formalize in such a way to use purely deterministic reasoning models of yesteryear, but I'm sure many specific laws could be converted, especially in fields like tax law (I'm sure they already are by many big players)
A lot of law is combing through large bodies of rule sets and reference cases to find inputs or parallels relevant for a case. Are you sure you'd rather have Lionel Hutz, Esq. with his limited comprehension do that at half a page per minute?
There’s no question that the technology has improved. It’s just that the way GP described the situation made it sound like there weren’t legal indexes before.
Yes, actually, because Lionel can suffer a consequence. He can be sanctioned. He can be disbarred. He can be sued. He can be thrown in jail.
AI can not meaningfully suffer consequences for what it does, and the trend is to remove the human from the loop more and more.
Furthermore, it's entirely conceivable that you have a law firm staffed by lawyers who more-or-less just say "yeah sure whatever" to what this thing does and, when asked about what can be done to appeal something, or how a document didn't make it into evidence, or how a subpoena wasn't filed, just shrug their shoulders and go "iunno". Vibe-practicing law, basically. You need to have an idea of how they're using AI and in what parts of the case.
If Lionel submits junk from an AI, he can still be sanctioned, disbarred, sued, and thrown in jail, just as he could if he let some incompetent paralegal screw up and he submits the resulting documents without checking them.
It's ultimately his name on the documents submitted to the court.
> but spending years or decades in prison because Lionel Hutz, Esq. did something stupid at a prompt sounds unappealing to me.
In which case would this be the primary blame, and not the legal system itself, or you for actually doing the crime?
I fail to see how its your lawyers fault that you went to jail for a crime you didn't commit over the system as a whole. And I really fail to see how it's the lawyer's sole fault that you went to jail for a crime that you should - in fact - have went to jail for...
There are a lot of problems with the legal system (including with lawyers). This just doesn't seem like the hill to die on.
There are a lot of situations where my lawyer's mistake might cause me to go to jail for a crime I didn't commit. There could be some prejudicial evidence that he failed to get excluded, or exculpatory evidence that he failed to get included, or an incorrect ruling in the trial that wasn't properly preserved for appeal.
One of the nice quirks of USA justice system is that you don't have to be guilty of a crime to go to prison, just that the prosecution has to convince 12 people beyond reasonable doubt that you are guilty.
The sixth is taken seriously to the point that incompetent lawyer could be a valid reason for granting a retrial.
An attorney I have retained from a large firm bills me $500 to send me one email update a month (minimum billable).
They are obviously AI-generated (but does not disclose such) and frequently misses or conflates key details the same way Gemma in particular does. They're also bizarrely positive when I correct them instead of pissy and arrogant like I'm used to.
Nobody should be paying these rates for legal "work" performed by a machine that poorly acts the part.
Lawyers, essentially the exorbitant and freeloading self-appointed gatekeepers of the deliberately arcane, unaffordable, inaccessible, and patently unequal legal systems all across the world, will now have even their routine/rote, repetitive, menial work done by cheaper LLMs for cheap.
Hopefully. Giving everyone cheap and equal access to the law cannot be a bad thing--well, it might feel bad to those who no longer have exclusive access to the law.
I think an under-appreciate way AIs might disrupt the world is by giving individuals an agent that will ruthlessly optimize for them.
Companies have entire departments that spend all their time ruthlessly optimizing so they can charge an extra fee here, or spin up a new shell company there. Meanwhile, here I am in our capitalist society and I just enjoy relaxing and spending time with my wife and kids instead of maximally exploiting the capital I possess at all times--like a fool.
It will be interesting if we ever integrate AI so fully in our lives that the AI can maximally exploit all our capital. My AI might check prices, and I will drive past your store to save 5 bucks at the next. My AI might help me manage 27 LLCs and shell companies of my own so that I too can just disappear in a poof of legal smoke if I ever need to. Etc.
Lawyers are going to start living gilded lives. I don’t see us getting past their professional human sign-off so they are essentially going to be guaranteed, easy, high paying jobs with the LLMs doing all the heavy lifting in the future.
While I agree on the premise – (written*) lawyer activity being reduced to putting signatures under things – I’m not so sure the gilded part is how it’ll actually play out in practice.
It might also become a race to the bottom, since you can’t deliver on quality better than others anymore, you’ll have to put effort towards quantity.
That could mean firms might pressure lawyers to sign as much as possible in a given time, while they’re still personally at risk/responsible for the documents they put out. That means either very hectic reading, and missing details, or no reading at all.
*And like socializer said, the written documents are only one part of the job.
That's already a thing. You have paralegals do most of the grungy work for much less than your hourly rate. You still have to show up to court and talk to customers, and these parts are not going away.
Or... the tools get to the point where we don't need a human for expertise but accountability and the role becomes more about ethics, thereby lowering the bar. (PUN!)
I was about to suggest that AI might teach people to become lawyers for cheap and they can just take the BAR exam and become a lawyer...
But then I looked into it and I see that you actually do have to go to certain schools to become a lawyer. Having the knowledge is not enough, you have to pass through the proper (and expensive) gates as well.
It looks like they're relying on tracked changes and human review, but I didn't see anything addressing something that has frequently made the news: LLMs outright hallucinating cases or holdings.
If the human lawyers don't catch that on review, they're likely to have an angry judge and an AI saying "absolutely, you were right to call me out on that, I shouldn't have hallucinated a non-existent case and made the judge sanction you."
I would love to see these benchmarks split up across "generic" tasks and novel tasks.
For example, redlining an OEM agreement with my vendor. They pretty much all use the same template, and the objections are always the same. I tried using Claude first, and even my lawyer was impressed that it found things they missed. Then I used Claude to redline that type of contract and the other side said "your lawyer was surprisingly thorough". I had my lawyer do the final check, but saved a bunch of money doing the first three rounds with Claude instead. And my lawyer basically just said "looks good to me" by the time they reviewed it.
Or even writing a simple will or estate document.
It seems like LLMs are great for the kind of generic work you'd usually pawn off to a new associate or paralegal, but not for the important things that partners usually do.
> It seems like LLMs are great for the kind of generic work you'd usually pawn off to a new associate or paralegal
This seems to be the case with most industries that LLMs are impacting, coding, data input, legal etc.
Whilst it's great to get the time saving from not delegating to a junior (LLMs are much faster at the basic level 1 work) there's going to be this big knowledge gap that you really only gain from actually doing the mundane boring shit.
A lot of this _should_ be automated away, but we really need to think about how the next group of people are going to learn the skills and nuance to applying them. I don't really have an answer to that one other than still giving that work to junior staff, but they will then just utilise an LLM to help complete the job, and rightfully so, I'd have done exactly the same thing.
I think OpenAI tried doing this with some early chatgpt settings with "learning mode" or whatever that is. But an LLM just telling you what to do will never teach someone, there was something about having to go out and hunt for the answer through the sea of crap responses or wrong SO questions that really taught you something. How do you replicate that learning with an LLM and how do you apply that to different industries like Law?
> but we really need to think about how the next group of people are going to learn the skills and nuance to applying them
That's what school is for. For the last 150 years we've automated a bunch of things, and then we send people to school to learn how those automations work, so they can make better automations.
Maybe law school will have to be a year longer. Or maybe some forward thinking firms will have new associates check the work of the LLM to gain that experience.
so they eliminate your talent pipeline and cut down on billable hours? is this good?
Not good for the law firms, but great for the rest of us!
how much 'novel' tasks are there in law?
> It seems like LLMs are great for the kind of generic work you'd usually pawn off to a new associate or paralegal, but not for the important things that partners usually do.
This seems to be the common thread everywhere LLMs end up being useful. As the sole software developer? Pure trash, but as "tons of juniors we can shuffle boring tasks at": great. As the sole designer and maintainer of a website? It'll look like 150% slop and no will buy anything, but as a designers friend to create what they dictate: even a designer can throw together a good looking website quickly. Etc etc.
I’m not sure if this is the right solution, but we clearly need better tools for the legal space. Anthropic's recent policy update makes this abundantly clear:
“When our products are used in ways that can affect someone's health, legal rights, finances, livelihood, or access to essential services, we require there to be a qualified ‘human in the loop’ (that is, someone who has the authority to review and, if necessary, change Claude’s recommendations), and for the individual affected to be told that AI was used.”
Something as common as a family law case has the potential to affect all of the above. Now imagine being a petitioner, respondent, paralegal, or other 3rd party using Claude to review a draft document only to find out that Anthropic has sent your entire file to opposing counsel.
It sounds egregious, but Anthropic is clearly stating they will use their own judgment to control what people can do with its product, and that they will treat customer data as if it's theirs to disseminate. So there's a massive opportunity for an unbiased tool willing to fill this void.
> Anthropic is clearly stating they will use their own judgment to control what people can do with its product
Haven't they been perfectly clear about that from day one with the ToS? I'm not saying I like it but it hardly seems unreasonable.
> they will treat customer data as if it's theirs to disseminate
Not sure where you're getting that. What you quoted in no way implies that.
The traditional legal data brokers like Thompson Reuters provide AI solutions presumably with considerably better data governance.
https://www.thomsonreuters.com/en-us/posts/innovation/thomso...
As someone forced to this, its complete garbage for doing anything useful
Marketing brochure for a fine tuned Qwen (which they don't mention), the model discussed here is likely better. I would not trust these guys to deliver high quality solutions given a friend's experience there, but you do you I guess.
it's like Uber, for lawyers!
Not sure how good the Harvey benchmark is. I recently looked into traces from an agent solving their firm knowledge task subset and found multiple problematic tasks and rubrics.
Here are 3 examples (there are more): https://github.com/harveyai/harvey-labs/issues/191
Harvey bench is probably one of the worse benchmarks out there. Agents as judges, vast scope, some questionable problem/solution pairs. Tricky domain TBF.
If you actually did the tasks in the way they deem correct, at least for immigration and white collar criminal defense, you will get detaind/arrested, and removed in the first case, and the second case you will snitch. It assumes that the state is correct. In reality the state, especially in the initial phase, are never correct on all aspects. You also can't only plan on snitching because what evidence exists and what evidence is admissible and what evidence is admissible but can be used in your advantage are all unknowns. Half of the bench tests on "covering your own ass" basically, and also, seems to have not taken the Loper Bright decision into account, which changes administrative law and how it is adjudicated and appealed entirely, for good or ill.
I don't know the other fields well enough, but if all your clients end up needlessly detained or admit to things they have a valid defense for, eventually people will notice. It's not exactly the two major sins that will definitely get you disbarred - stealing money and screwing client while representing them - but it's definitely at least incompetent. They seem to not understand that common law jurisdictions are fundamentally different from civil ones, and also, there's quite a bit of difference between how the system evolved in each country. Or that in China what's printed is basically more like a manifesto, since it's almost entirely run by the bureaucracy and personal relationships, and lawyers that aren't foreign-trained and work for foreign firms aren't considered prestigious since the courts are essentially entirely for show.
They need to get like, one jurisdiction, one area in that jurisdiction, and specifically the kind of sub-sections in that area down pat. But they can't train on real material so they will never be able to do so. Their banner testimonials come from non-lawyers and that says it all.
This is why I severely dislike using all pass rate. A model could be at like 93%, meeting practically every requirement except the broken ones and all pass would give it like 20%.
Approximately every benchmark has significant issues.
Of course. That’s why I opened the issue so that the team can improve it.
Every existing, and not yet existing benchmark.
As an attorney, I am not particularly impressed about a 20% pass rate on a benchmark. The name of the game for litigation practice is about being prepared for novel situations, not relitigating past ones.
It is amusing how many people are arguing what litigation is here, probably none of these people have a law education.
Imagine a bunch of lawyers arguing what programming is. Even a bunch of programmers arguing what programming is abnoxious at worst and a mental exercise at best.
That's how I feel reading HN's opinions on the recent math results. <5% of the responses even understand the point they're supposed to be arguing. AI might take over mathematics. It might not. The discussion regarding that topic is almost devoid of any relevant information though.
Lawyers would be better suited to argue what programming is because they know logic. There’s not much to algorithms: do this, then do that.
Programmers typically don’t know law outside of which license allows them to get code for free.
> probably none of these people have a law education.
Domain knowledge is informative sometimes but mainly overrated: the success of LLMs proves just that. What most lawyers do is fairly standardized and routine, efforts to invoke mystery cases or "novel" is an attempt at resisting the inevitable. The "the only British food in Palo Alto" equivalent of unique skills.
To be fair to them, on this particular benchmark (Harvey's Legal Agent) 20% puts it almost top of the pack of _any_ model, and certainly top for any remotely in its price class.
Opus 5.5 scores 4% on it.
Objectively a 20% pass rate is not impressive. Yet, there is still utility to AI research and drafting.
Wouldn't be surprised, if a human attorney would get 5%, while copy-pasting ChatGPT answers, report in the results a week past the deadline, with a $12,500 retainer charged upfront.
If professional AI can do all the work that isn’t novel, you free up human talent for edge cases and then improve your AI with new information.
How much of your work is truly novel and unique? Not 100%
Totally fair. However, even if the LLM would come to the same conclusion and produce the same content that a competent attorney would, the attorney must still thoroughly vet the work for accuracy. So there is potentially time saved in research and drafting, but it’s not the miracle some might characterize it as.
An 80% miss rate on a benchmark suggests to me that automation has a very long way to go, even for the drafting and research phases.
What are you talking about? 90% of litigation is convincing the judge that existing case law that clearly applies to your case does apply to your case.
I'm guessing what they're saying is that nothing goes to litigation if it's a clear pattern-match to existing cases, so actual litigation practice is only novel cases.
Isn't that just in the US/Commonwealth judicial system? I could be completely wrong on how others, like non-adversarial judicial systems, work and/or how pervasive the US system is or is not.
EDIT: specifically around the use of judicial precedent in case-law.
90% of litigation is pressuring a settlement.
> 90% of litigation is pressuring a settlement.
90% of legal work where there is a plaintiff and defendant is pressuring a settlement. My wife, an attorney, does not consider that to be "litigation".
As a former litigator, I would consider that to be litigation the moment a complaint is filed in court, regardless of whether the dispute is settled without trial.
But, for your sake, it's probably best that you accept your wife's definition.
And most of the leverage you have in negotiating a settlement is by referencing prior case law. If something is novel, there's too much variance to get a good settlement.
No, most of the leverage you have is how much time and money is on your side vs. the opposing party.
[flagged]
How much novelty can there be? If attorneys were as diligent and generous with “open sourcing” their data as SWEs had been, could LLMs benefit from all that training data?
> How much novelty can there be?
As much novelty as there is in human interactions.
> How much novelty can there be?
Infinite
In the same sense there could be infinite CRUD apps?
In the same sense as infinite jest.
Law and code are only superficially similar as law always requires human interpretation plus law is an interface to reality. Code can stay contained in its nice, little, abstract box.
So every lawsuit dispute, divorce, traffic injury, etc… is a completely novel case that requires depths of creativity far beyond simple CRUD apps or solving 377 open math problems.
Makes sense.
Isn’t code an interface to reality as well?
A lot of code is an interface to other code. Nginx is how many layers removed from actual physics?
Looks like this is not an open model. I'm curious how the sentiment toward chinese models will shift as labs trend toward keeping them proprietary.
Also, I wonder if this is at all related to the recent reports that Harvey has been getting squeezed on margins by the US labs.
Nobody will have any interest if they are not open
Yeah agreed, especially for legal and healthcare models.
[edit] I'd go one step further too and say the training data should be explicitly requestable too so third parties can verify (freedom of information act styles).
Training data being discoverable is much more interesting than open weights. Also a PITA so nobody does it.
[flagged]
Not in the West, but they will be in China and the rest of the world.
bold claim with "the rest of the world". more like rest of the 3rd world
If you are not self-hosting on your hardware inside your network, why do you care whether the model is open or not?
In theory if your provider raised rates you could switch providers with an open model. So price competition would be one reason.
because vendors stop serving models and it is good to be able to choose to self host for any reason at any time
we can start out with the hosted version for evaluation and then self host for production and real customers
https://quoteinvestigator.com/2019/07/30/lawyer/
"he who will be his own Counsellour, shall be sure to have a Fool for his Client."
How much worse the lawyer who subcontracts Freakazoid?
https://en.wikipedia.org/wiki/Freakazoid!
"Freakazoid has ... access to all of the Internet's knowledge. He cannot fly, but often runs around with his arms extended and making whooshing sounds with his mouth, pretending he is."
Curious what domain is next, GLM-Science? And when will it be profitable to drill down to even more domain specific models like GLM-Economic-Geography or GPT-Aquaculture?
Imo, we are going to see a lot more specialised models in the near future as pushing the frontier gets harder. Qwen 3.8 27B for example is a pretty rubbish chat and general world knowledge model, but is stellar at coding (for it's size). Its small enough anyone with a modest gaming PC (16GB vram and 32GB of DDR5) can run it at speeds good enough for overnight coding sessions with an PR to review at the end of it.
Eventually, I hope we'll see many more specialised small models that run nicely on consumer and prosumer hardware. Like LLMs specific to income tax law for UK residents, or LLMs specific to building power delivery circuits for PD USB C devices. There probably is a point where the niche-ness does become to niche to justify the post training, but again, small models means much less training required.
Kind of begs the question if this even needs new/better data? Could you take the foundation model that's trained on everything, and then run a final fine tuning training run on just the set of foundation training data that you want the model to specialise in? Eg, extract just the training data that is about UK income tax law and then run a few epochos on just that? Any ML people know if this is a thing or if not, why it doesnt work?
A progress for sure. Just don't expect GML-legal to answer about Xi or CCP for any US courts that mention those terms.
Sort of ironic, in Google Chrome, when attempting to visit that legal z.ai site I get:
Your connection is not private Attackers might be trying to steal your information from z.ai (for example, passwords, messages, or credit cards). Learn more about this warning net::ERR_CERT_COMMON_NAME_INVALID
If you’re on a corporate device you may not have recently issued certificates in your cert chain yet
If he is on a corporate device, he has a corporate-issued MITM certificate, which allow third party "security" companies to read, store and analyze his complete network traffic in plain text.
Not the case for me.
Wait, so this not an open weight model, correct? Only available via z.ai?
So, if a US or EU person shares their private legal info here, it's processed and stored by an adversarial country that has no extradition treaty with their country?
I prefer all my data be in the hands of foriegn entities, they cant do anything to me. Id much rather Huawei spy on me than Meta (if I have no choice other than to pick one) There's literally no consequence of Huawei having my data. Unless you're a VIP, a chinese entity storing your data is probably better than a US company, as a westerner. It wont be used to manipulate you and will generally be used less adversarially against you.
I think part of the reason there's so much propaganda around chinese tech is to scare you away from possibly using them, not for your own safety, but so Meta/Google/MS ect dont lose access to the ability to spy on every aspect of your life.
I love how we have to pretend like China is some scary boogeyman because people on screens say they are.
I couldn't agree more but feel that so many people on here consider themselves VIP that this won't be well-regarded
What would prevent a Chinese company from selling your legal data to say, your legal adversary via data brokers, in the USA? People everywhere like to make money.
I initially wanted to dislike this but it's actually genius hahaha
It will not protect you. China data brokers are generally not prevented from selling data on US people to US companies.
> I love how we have to pretend like China is some scary boogeyman because people on screens say they are.
I think you've lost the plot a bit on this one.
I'm scared of China because of their lack of democracy and the terrifying stuff they do to their own citizenry, which presumably they'd very much like to expand.
So, you're all-in on Mistral then? Do they offer something similar to this?
This sounds like a huge feature though? Basically, a digital attorney–client privilege.
A key point of attorney-client privilege is that it's a feature of the legal system itself, not a technical or political hack.
There's no digital attorney-client privilege. If a lawyer uses LLM's on your behalf, the communications are protected. If you use the same LLM's on your own behalf, the chats are not protected.
DOJ could cut Z.ai off from US-based payment systems (Mastercard/Visa) if they don't comply with subpoenas.
They could, but generally it is expected that before a subpoena is complied with it is checked to be compatible with the laws of the country it’s in, and requires cooperation from that country. Most often they are ignored if there’s not a presence/entity/assets in the US.
It’s also hard because no individual case would likely justify a move like that. It would have to be the sum of many, many cases.
Not every judge agrees with your statement of the law of privilege.
More like attorney-client liability
You got it wrong. If courts and attorneys in the US can't run "discovery" on Z.ai servers or compel them to do anything, and if Chinese authorities don't cooperate with the US legal system as a matter of principle, then what you functionally have is super privilege.
They'll just run discovery on you. If you try to respond to a subpoena with "the documents are outside the country and I won't give them to you," the judge will throw you in jail until you do.
You're right on this point. But there are ways to build interfaces that sidestep this.
Run a sanitized VDI with no local storage --> offshore independent entity (e.g. a Hong Kong Company) that holds the AI service contract and API license --> GLM-Legal.
Somebody ought to set this up for pro se work.
Yeah sure but your clever sidestep is just "hiding evidence from the court." The kind of person that feels the need to have a secret AI lawyer is also probably going to just tank their case in another way.
You should not trust any LLM to give you accurate legal advice. I can’t imagine all the ways that could go wrong.
Like pattern matching sentence algorithms many can reveal concepts and strategies to confirm independently
Good luck to a EU person getting "US AI people" extradited. They aren't even put on trial in the US.
NATO allies can't even get US soldiers to trial. In Romania we've had cases of manslaughter and the US just shipped them back ASAP and at worst they were discharged dishonorably.
For reference, that specific case is from 2004. Christopher Van Goethem was Marine security ambassadorial staff and was convicted in a military court martial for obstruction and making false statements. The whole case was porrly handled and the Romanians are right to be pissed. A key piece of evidence that would have landed him in jail for 10 years was the blood alcohol draw but the Marines slow walked it in Romania and he didn't show as intoxicated even though he admitted to drinking that night.
Nobody can. See the Hague Invasion Act
https://en.wikipedia.org/wiki/American_Service-Members%27_Pr...
[flagged]
[dead]
No discussion on this can avoid mentioning this ongoing disaster...
"AI Hallucination Cases" - https://www.damiencharlotin.com/hallucinations/
When I first tried using IDE AI coding assistants a bit over a year ago, they would frequently fail to read files and completely hallucinate the code I asked it to analyze or edit. And even when it did successfully read my code, it would often hallucinate the API of other frameworks.
I don’t think hallucinations are a major issue with agentic coding a year later? At least, not for me.
Legal writing does have the problem that there is no such thing as a correct argument that a goal-driven loop can target. But basic things like “does this case even exist” seem solvable with research agents, and even harder problems like “does this case support the argument being made” seem conceptually very similar to questions I now ask Claude daily.
Yes, mechanical question of "does this exist" is trivial to solve. Whether the research is actually aligned with the user intent / the AI makes claims that are substantiated by the research at hand are much harder problems.
> I don’t think hallucinations are a major issue with agentic coding a year later? At least, not for me.
Are you still looking at code and agent traces to know if this is true or not?
At work, we still see hallucinations across all model families as we share bot claims for each others code
An issue I come across is the attributed failures of AI of a first time use of AI.
Non-deterministic software can behave orders of magnitude different based on how input is structured.
This factor is largely missed by bringing a deterministic mindset to trying AI.
This means the same problem solution can work for one person reliably, and not another and both are true.
Assertions that begin with a deterministic assessment using a snapshot in time of both the user and models capabilities are a new feature of this new category of software that will have to be better caught and expressed.
I provided some very light guidance to a couple of masters students who were working on legal citation hallucination detection last summer. Their summer project is here: https://github.com/gt-csse/mellea-lrc
If anyone here is looking for engineers to work on legal AI products, both of the students who worked on this project come highly recommended.
There being multiple cases every single day is nuts.
But also, interesting how the rate of cases peaked and seems to be going down now.
To your second point, it seems like every time a significant practical limitation is identified, it goes through a familiar pattern. It gets a news cycle, and then it gets a wave of confidently incorrect opinionating, and the problem gets treated as if it's a permanently built in problem that's baked into the foundation of AI. And then time passes and the issue is either mitigated or circumvented entirely.
Hallucinations are one, "AI can't scale anymore" is another (remember that one?), limited context windows, strawberry problems, glue on pizza, sycophancy.
I don't want to say there's no problems, there are very serious ones. User psychosis, economic harms, alignment, cost of computer components, etc. But there's a lot of false positives that get turned into Ed Zitron articles along the way.
So many of these we likely using Google to search then found AI mode and used that. Which is to say, they were (and still are!) using Google Gemini 1.5 from many years ago.
A lot of the examples they show seem to be focused on common law countries. The US and UK are where most of the money is to be made in the legal industry, but I am curious to see what ripple effects we get from people attempting to use AI tools geared toward US/UK-style legal systems in other kinds of systems.
The usual concern is attorneys, but judges are probably even more overloaded these days. Can this help judges as well? If so, would this ultimately lead to better decisions?
It will lead to bigger fines :-) https://www.damiencharlotin.com/hallucinations/
Until it hallucinates something no one bothers to verify
Judges hallucinate too. We call them bad rulings.
[dead]
Unexpected direction for zai to be going into
Video is very SAAS-demo coded. I wonder if they worked with any in-domain legal experts at all.
How does this compare with existing frontier models especially when processing massive volumes of evidence that might be contradictory...
No fable 5.1 no opus 5.5 , no gpt 6.1 , benchmarks for the hype .
Would be nice if it said what legal system it is set up for. US law? Chinese law? European law? All of them? The author seems to assume we'd know.
The benchmark is US law, I imagine it's been aggressively post-trained on it.
the liability model of these models is poor, i think progress in that domain should be a heavy focus.
The legal space understands well how to tier services based on who has the liability. E.g. trainees and paralegals whose work is filtered through the lawyer who ultimately puts their name on what gets submitted to courts.
I'm sure we'll continue to see the occasional person who fails to understand they themselves will be held liable if they submit work done by an AI without vetting it, but it's not like they don't have ways of dealing with this.
Why don’t I lose all legal privs when I hook up my email to these third party services
There should be laws about disclosure on the part of attorneys about their use of AI. I know they're already using it, but spending years or decades in prison because Lionel Hutz, Esq. did something stupid at a prompt sounds unappealing to me.
I think they should be somehow restricted to using non-generative models (or at least for the user facing output). Something that would only identify potentially relevant case law from an existing set, to accelerate the search process.
Common law in general seems essentially impossible to formalize in such a way to use purely deterministic reasoning models of yesteryear, but I'm sure many specific laws could be converted, especially in fields like tax law (I'm sure they already are by many big players)
A lot of law is combing through large bodies of rule sets and reference cases to find inputs or parallels relevant for a case. Are you sure you'd rather have Lionel Hutz, Esq. with his limited comprehension do that at half a page per minute?
This capability has already existed for decades thanks to Westlaw and LexisNexis. These are indispensable tools in the lawyer’s kit.
You're kidding, right? Using Astra to comb through cases is far faster than doing it manually in Westlaw. Models for research are pretty good now.
There’s no question that the technology has improved. It’s just that the way GP described the situation made it sound like there weren’t legal indexes before.
Yes, actually, because Lionel can suffer a consequence. He can be sanctioned. He can be disbarred. He can be sued. He can be thrown in jail.
AI can not meaningfully suffer consequences for what it does, and the trend is to remove the human from the loop more and more.
Furthermore, it's entirely conceivable that you have a law firm staffed by lawyers who more-or-less just say "yeah sure whatever" to what this thing does and, when asked about what can be done to appeal something, or how a document didn't make it into evidence, or how a subpoena wasn't filed, just shrug their shoulders and go "iunno". Vibe-practicing law, basically. You need to have an idea of how they're using AI and in what parts of the case.
That same stuff can all still happen tot he lawyer who uses these tools. If he makes an egregious enough mistake, he's liable.
And that's why you want him to have to disclose that he used those tools.
Why do the tools matters? He is equally liable if he makes the mistakes without them.
If Lionel submits junk from an AI, he can still be sanctioned, disbarred, sued, and thrown in jail, just as he could if he let some incompetent paralegal screw up and he submits the resulting documents without checking them.
It's ultimately his name on the documents submitted to the court.
> but spending years or decades in prison because Lionel Hutz, Esq. did something stupid at a prompt sounds unappealing to me.
In which case would this be the primary blame, and not the legal system itself, or you for actually doing the crime?
I fail to see how its your lawyers fault that you went to jail for a crime you didn't commit over the system as a whole. And I really fail to see how it's the lawyer's sole fault that you went to jail for a crime that you should - in fact - have went to jail for...
There are a lot of problems with the legal system (including with lawyers). This just doesn't seem like the hill to die on.
There are a lot of situations where my lawyer's mistake might cause me to go to jail for a crime I didn't commit. There could be some prejudicial evidence that he failed to get excluded, or exculpatory evidence that he failed to get included, or an incorrect ruling in the trial that wasn't properly preserved for appeal.
> In which case would this be the primary blame, and not the legal system itself, or you for actually doing the crime?
There are plenty of ways a lawyer can screw up procedurally that can drastically impact your case.
One of the nice quirks of USA justice system is that you don't have to be guilty of a crime to go to prison, just that the prosecution has to convince 12 people beyond reasonable doubt that you are guilty.
The sixth is taken seriously to the point that incompetent lawyer could be a valid reason for granting a retrial.
An attorney I have retained from a large firm bills me $500 to send me one email update a month (minimum billable).
They are obviously AI-generated (but does not disclose such) and frequently misses or conflates key details the same way Gemma in particular does. They're also bizarrely positive when I correct them instead of pissy and arrogant like I'm used to.
Nobody should be paying these rates for legal "work" performed by a machine that poorly acts the part.
Lawyers, essentially the exorbitant and freeloading self-appointed gatekeepers of the deliberately arcane, unaffordable, inaccessible, and patently unequal legal systems all across the world, will now have even their routine/rote, repetitive, menial work done by cheaper LLMs for cheap.
This is why we see so many pro-se litigants. "Legalese" is no longer the barrier (not to nearly the same extent, anyway).
Hopefully. Giving everyone cheap and equal access to the law cannot be a bad thing--well, it might feel bad to those who no longer have exclusive access to the law.
I think an under-appreciate way AIs might disrupt the world is by giving individuals an agent that will ruthlessly optimize for them.
Companies have entire departments that spend all their time ruthlessly optimizing so they can charge an extra fee here, or spin up a new shell company there. Meanwhile, here I am in our capitalist society and I just enjoy relaxing and spending time with my wife and kids instead of maximally exploiting the capital I possess at all times--like a fool.
It will be interesting if we ever integrate AI so fully in our lives that the AI can maximally exploit all our capital. My AI might check prices, and I will drive past your store to save 5 bucks at the next. My AI might help me manage 27 LLCs and shell companies of my own so that I too can just disappear in a poof of legal smoke if I ever need to. Etc.
Professional AI is so interesting to me, legal, medical, engineering, accounting all can benefit.
And the opportunity for a moat of AI that is correct and accountable(legally) is ripe for the picking.
Lawyers are going to start living gilded lives. I don’t see us getting past their professional human sign-off so they are essentially going to be guaranteed, easy, high paying jobs with the LLMs doing all the heavy lifting in the future.
While I agree on the premise – (written*) lawyer activity being reduced to putting signatures under things – I’m not so sure the gilded part is how it’ll actually play out in practice.
It might also become a race to the bottom, since you can’t deliver on quality better than others anymore, you’ll have to put effort towards quantity. That could mean firms might pressure lawyers to sign as much as possible in a given time, while they’re still personally at risk/responsible for the documents they put out. That means either very hectic reading, and missing details, or no reading at all.
*And like socializer said, the written documents are only one part of the job.
That's already a thing. You have paralegals do most of the grungy work for much less than your hourly rate. You still have to show up to court and talk to customers, and these parts are not going away.
Or... the tools get to the point where we don't need a human for expertise but accountability and the role becomes more about ethics, thereby lowering the bar. (PUN!)
I was about to suggest that AI might teach people to become lawyers for cheap and they can just take the BAR exam and become a lawyer...
But then I looked into it and I see that you actually do have to go to certain schools to become a lawyer. Having the knowledge is not enough, you have to pass through the proper (and expensive) gates as well.
I'm not really surprised.
This is just SWEs and Consultants discovering the utility of professional regulation. It's still a tough job.
Do you predict the same paralegal crisis people are claiming about junior swe's?
don't most developers do this already? Senior developers. Juniors are having a hard time.
It looks like they're relying on tracked changes and human review, but I didn't see anything addressing something that has frequently made the news: LLMs outright hallucinating cases or holdings.
If the human lawyers don't catch that on review, they're likely to have an angry judge and an AI saying "absolutely, you were right to call me out on that, I shouldn't have hallucinated a non-existent case and made the judge sanction you."
"advancing". Absolutely laughable, we're talking about less than 20% accuracy on benchmarks. 100% useless.