back
85 comments
Nilay Patel argues that law is undeterministic (and its application ambiguous) to begin with:

> But law isn’t actually code, and society and courts aren’t computers. [...] the law is not deterministic. You simply cannot take the facts of a case, the law as written, and predict the outcome of that case with any real certainty, even though the formality of the legal system makes people think it works like a computer — that it’s predictable.

> [...] it’s actually ambiguity that’s at the very heart of our legal system. It’s ambiguity that makes lawyers lawyers. Honestly, it’s ambiguity that makes people hate lawyers because it’s always possible to argue the other side, and it’s always possible to find the gray area in the law. That’s why prosecutors end up working as defense attorneys and why our regulators tend to end up working for big corporations.

https://www.theverge.com/podcast/917029/software-brain-ai-ba...

IMO, as with most domains, AI _tools_ will save a huge amount of time, but it's the human specialist making judgment calls based on real world context.

You can apply this same argument to everything. Code is deterministic but what is being made is often not because people don't know what they want to make. Society can choose just to make everything boring and deterministic so that computers can do everything.
It’s not just that law is ambiguous. That’s something that can be worked around. It’s that lawyering deals with human disputes and everything that comes with that. Imagine asking a couple that’s getting divorced “why are you getting divorced?” You’ll get two completely different stories with different recollections of the facts. Business disputes aren’t really any different. People have selective recollection, they don’t remember things, they shade the truth or hide things or just lie.

Oftentimes, the law is reasonably clear. The hard part is getting the facts out of the witnesses to figure out what really happened, so you have facts to apply to the law.

This isn't really true though. This is how the law used to work, until people did the research and discovered it let to absolutely loads of mad variation in outcomes, with people with similar offences getting totally different sentences based on random luck. Hence most countries not have pretty strict sentencing guidelines, with a bit of space for judgement on top (despite a lot of protesting from judges).

https://www.ubs.com/global/en/our-firm/what-we-do/our-brand/...

You should be able to predict the outcome of a court case if you have all the facts available. That's what fairness means.

> Nilay Patel argues that law is undeterministic (and its application ambiguous)

I argue that of all things, law should be as deterministic as possible.

I've always thought that we (as a country) should maintain one single ordered list of specific crimes and punishments. Every new case that wants to set a punishment must insert it into this ordered list and explain convincingly why it fits into the list at the proposed position.

This would prevent the outrageous differences we see today where someone gets a few days of house arrest for murder and another guy gets a decade of solitary confinment for stealing a pen.

Counter argument - even stone age Chat GPT 3 was great at making reasonably convincing sounding arguments and newer models are great at aligning those arguments to souces (laws or cases). Chat is IMO better at ambiguous nonsense than objective analysis.
Lawyers bill by the hour. It is not in their interest to speed up what they do, because they only have so many clients. They would need far more clients to bill the same hours.

I used AI to save my last company thousands of dollars, and more importantly weeks of time. When I had to negotiate a contract, I'd just have Claude legal make the redlines against the counterparty for me. Sometimes their lawyer even complimented my "lawyer" on how thorough it was.

Only when a final contract was agreed on did I engage my human lawyer for final review (and usually they didn't find anything of concern).

For standard contracts, AI is pretty good.

> Lawyers bill by the hour. It is not in their interest to speed up what they do, because they only have so many clients.

This isn't how markets work.

If one lawyer starts doing 10x as much work in the same number of hours, then all the customers will move to that lawyer as soon as they find out, and the other lawyers will have to adapt to remain competitive.

> Lawyers bill by the hour. It is not in their interest to speed up what they do, because they only have so many clients.

Just like in IT, in law billed hours by one person have little correlation to hours spent working by that person (or others!). The billing does affect customer/client retention and reputation though.

If you involve a lawyer only for the final review, they will understand you don't want them to rethink your approach nor to advise a different strategy. So, they will just bill you for the time, which seems to be fine for both parties of the trade. And, where it's mandatory to have a law firm draft a deed or a contract (like for mortgage or real estate transactions, depending on the state you're in), your Claude-made first draft will not make your bill lower. Just don't try to rely on this for adversarial work, like in litigation. BTW, just curious about the timing: how could Claude for Legal save you thousands of dollars at your last company if the product came out in February? (Not that Claude for Legal is any special in its legal output compared to "ordinary" Claude)
These are probably contracts where a lawyer would struggle to add value anyway, or you wouldn’t have hired them in the first place. Seems more likely a Jevon’s paradox example to me than anything.
Your future legal counterparties will thank you for using AI to draft your contracts. It's going to make lawsuits so much easier...for them.

AI is great at extremely simple contracts for which most lawyers just use standardized templates. Incidentally the same templates used to train the AI you used.

I think the most likely situation with AI lawyers is simply Jevon's paradox. We'll simply ask for lawyer support in 1000's of more situations where we didn't before.

It's so counter-intuitive that even seasoned AI researchers get it wrong. It happened to radiologists, it's about to happen to Lawyers too [1].

[1]:https://fortune.com/2026/05/04/godfather-of-ai-geoffrey-hint...

Maybe you are right but I don't think the comparison holds that well. I can't take a scan of myself and send it to some AI because I don't have the hardware. I have to go to some highly consolidated business. But the last time I did something legal with AI there was no such choke point. It walked me through filing a trade mark and I didn't have to leave my desk.

The difference in capital deployed is huge. A common laptop on my desk vs big bucks for a real estate full of medical imaging equipment. If radiology becomes less profitable the industry won't deploy as many imaging machines. Legal doesn't have that moat.

Jevon's paradox is a 19th century economic observation about coal that has turned into a meme that people are using to support the most profound bullshit that they already believe to be true a priori.

Lawyers are exactly the counter example and how stupid this meme is.

If I was convicted of a crime and had access to Claude in jail, of course I would use profoundly more "law research". You wouldn't hire more lawyers though, you would do the law research yourself because you would be empowered to do law research in ways you wouldn't have been able to do previously and an order of magnitude cheaper than a human lawyer.

Food/farming is really the best example of the stupidity of this meme though. As if the increase in farming automation leads to more farmers as opposed to an increase in the consumption/quality/diversity/output of food.

The meme is a category error as it relates to the output vs the paid for human activity.

I think legal search and intelligent document search tools are where the money is. Not "AI lawyer" stuff.

Legal document templates and generating them with given constraints existed before the LLM boom. They are accurate and predictable. Many easy and clear cases are already automated. You get a basic case, you take a template, and if the case has something specific, you add it.

I can see LLMs helping with: some paralegal work, legal searches when humans are there to judge the results, spotting and reporting errors in writing. Small modifications for templates.

The problem with the "AI lawyer" idea is that after things are written down, most of the thinking is already done. The text is the output of a hard-to-automate process involving:

* "Asking questions," being curious, and spotting things visually or by listening.

* figuring out the angle (formulating a case theory),

* identifying what is special in the case (the core anomaly),

* what the client wants (the true client objective). That's almost never what they say unless it's another corporate lawyer. You need to figure out emotional drivers, risk tolerance, and what constitutes a "win."

* what the opponent wants (adversarial motives),

* identifying ambiguities. Society is always shifting, and new ambiguities are created steadily.

A lawyer does all this and writes down their thinking. Lawyers think in writing. The legal profession has a really amazing blogosphere.

> Only three entities in the United States have anything approaching complete coverage

> They sell the editorial infrastructure built on top: headnote taxonomies that organize millions of opinions into searchable categories, practice guides written by specialists over decades, and treatises that synthesize primary law into usable guidance.

> The Free Law Project’s CourtListener provides free access to millions of federal and state court opinions, oral arguments, and PACER documents.

I think issue of a data-moat is somewhat overstated, or at least it is not argued very well here. If secondary organization and interpretation of open data is their moat, and if it is mostly focused on guiding humans through the complex web of knowledge, then AI should make short-order of that.

But, as usual, the issue of structural and organizational barriers is definitely convincing. Sometimes existing players are too entrenched to change. A new kind of AI-oriented law-firm might need to emerge and show itself to be competitive to either make mainstream firms truly change or push them out of the market.

Precedent in law is like discovery in science. It advances the boundary of human understanding so a probabilistic regurgitator will have trouble applying it without treatise roadmaps.
There may also be a Bitter Lesson element here. Ultimately, if law is like other domains, we may be able to solve legal applications with more compute on the limited freely-available data.

That's a load-bearing "if", though, given the incohesiveness of legal systems compared to typical Bitter Lesson examples.

A lot of the applications of Ai are going to have to go through "normal" innovation routes.

Eg low-end disruption. I have already seen "Ai lawyer" at play here.

A colleague of mine is involved in a long class action against a builder. The group chats have gone absolutely chaotic this year... as members consult heavily with LLMs and the (real) lawyer can't deal with the volume of action.

Another friend is a wholesaler and does a lot of small-scale commercial deals. Contracts have gotten bigger and negotiation has gotten more involved as "Ai lawyers" read and write these contracts.

Employment contracts are much more likely to be negotiated, referenced, etc.

So... These are all routes to "classic" disruptive innovation. It's not replacing billable hours at law firms. It is replacing non-consumption.

Law is adversarial. A formal legal letter requires a form of legal letter in response. Law generates its own demand.

I would be watching a lot more for ground up innovation, rather than adoption at firms.

This does not mention either interpretation or hermeneutics. For a computer to function as a lawyer it would have to be able to perform interpretation.

I would expect such an article to start there, or at least make some argument that concludes that a computer actually could perform professional legal tasks. Which I don't think they can, just as they can't do philosophy.

> This does not mention either interpretation or hermeneutics. For a computer to function as a lawyer it would have to be able to perform interpretation.

Don't overwhelm engineers with hermeneutics

AI can't do novel research either right?
Law is a very large and varied field. I have met other lawyers in the past and thought to myself, it's crazy that people would consider us to have the same job. In terms of the kind of work involved, the expected output, the training materials required and the economics of the business, law is in reality more like several distinct fields with varying degrees of overlap.

Usually when people talk about the legal field online they focus on the act of advocating in court, appealing to juries etc. But that is a relatively small fraction of the actual work that lawyers do. Drafting, negotiating and advising is the bread and butter of legal work.

A lot of the value added by lawyers isn't just in knowing how a particular clause would be interpreted by a court or whether it is in compliance with statute, but also knowing what terms are "market", what would likely be acceptable to counterparties or regulators, etc. LLMs could learn that stuff but it isn't enough that they are trained on case law and statute (which in itself is not a particularly difficult problem IMO). They would also need to be trained, at the very least, on actual contracts, and ideally also on deal correspondence and notes of advice (all of which are going to be much more difficult to come by). Otherwise, expecting an LLM to output quality legal product is like expecting it to produce quality C code when it has only been trained on the spec and no actual code.

Don't get me wrong, there is a lot of legal work that could be automated relatively easily. But probably a lot of that is at the low end of the market, where there just isn't that much money to be made by doing so. Your local small town lawyer who will draft your will or work on the sale of your home is not in the same bracket as the New York or London lawyer who advises on blockbuster private equity deals or IPOs. The latter is probably what will interest firms like Anthropic and OpenAI. Within those high value fields, I do think AI will make significant inroads but it will be through the firms themselves adopting AI rather than human lawyers being replaced en masse by AI lawyers.

I assume also that LLMs are not very good at chronology.

Since law evolves, I wouldn’t be suprised that LLMs would spit out arguments that are out of date.

The good thing is new law usually references old law and explains why it has evolved so I wonder if arranging the law in chronological order will help the LLM follow the thread.
What I always wonder is why we don't have more standardisation around end-user contracts. E.g. something equivalent to YC's SAFE (https://www.ycombinator.com/documents) but for employment agreements, leaseholds etc.

We understand what we can & can't do with software licences and creative commons because we know "this is MIT" or "that is CC, no commercial, with attribution" and we don't need to delve further.

If we had similar for employment terms - ACME Ltd want to hire me for £x at Y location using the standard "UK employee contract" - it feels like you could sidestep a lot of the need for AI parsing individual documents that are all subtly different.

Lawyers are _already_ using templates, but they're all using bespoke templates and it means that you've got ambiguity by virtue of the fact that the sentence in my contract has never been tested in court.

Interestingly this already does happen a lot, but usually in more complex and high-value B2B contexts. ISDA Master Agreements, (G)MRAs, GMSLAs, LMA documentation, EFET General Agreements, Incoterms, etc. The fact that it is more common in B2B contexts than B2C is IMO a reflection of the power imbalance in the latter as it benefits a large corporation to have consumers contract on its terms. Where there is standardisation in consumer contracts it is usually driven by law or regulation as the market isn't otherwise incentivised do it. For example some countries do have template tenancy agreements published by the government or a regulatory body.
A massive firm in the UK has reported itself to the regulator for using AI not once, but twice in a submission to a court.

https://www.ft.com/content/5ba4690b-8b98-43b3-ba0b-f2ec5591a...

I suppose one way is that the Lawyers and Legal Assistants use Legal AI as a replacement for standard search. Instead of parsing content and creating new notes, let AI search and create but humans spend the same amount time instead for verifying what was created.

That way the billable hours can match, but like the article says, who does this benefit? Ultimately the transfer of time to another task will keep law just as expensive. Perhaps there is room to save time on verification vs creation. Is it worth all the investment though?

The problem with measuring AI productivity is that the people doing the actual job (paralegals, developers, etc.) are doing it for someone else (judges, managers, etc.). More work, or even a speedup does not actually benefit them. So when you give professionals a tool that speeds them up, they increase their slack and/or focus on other, less productive activities rather than work more.

The article captures this too, mentioning a couple of examples of startups where presumably this feedback loop is tighter.

Courts are already overflowing with years of backlogs. I would argue if everyone had a legal bot to represent their interests against others according to the law, and they could come to an automatic traceable agreement, that would be an overall benefit.

However there are highly (self) regulated industries like lawyering that will try to protect their business model with tooth & nails before yielding to what's good for the population.

I've booked an appointment with a real lawyer, not going to rely on AI for anything important legal wise.

I am, however, going to be extensively using AI to prepare for the appointment. A list of questions to ask, what to bring etc. I've also used AI to research in advance what the likely answers will be, so I'll have an idea of follow up questions to ask.

That should hopefully save additional appointments (and billable hours).

I've successfully sued my neighbor over land disputes. I've won small claims against companies. I've used AI to do it. I'm no longer afraid.

Obviously I'm not going to be taking on huge law firms but AI has opened up a whole new vector for me in that I am able to sue people at will without paying for any legal fees and I think that is the most powerful outcome of all.

My days are spent now not asking what mobile apps or SaaS I need to make but who I can apply pressure to. The fact is most companies do not like to go to court as it costs them a lot of money. For me I spent about $1000/month on various LLMs for software mostly but I am amused the amount of sway I can have on companies now.

So far I've gotten free internet after suing my ISP for overcharging me, won injunctions against a neighbor over encroaching fences and trees, and now pursuing legal actions against my ex coworkers and employers in an attempt to garnish their wages.

It truly is incredible how much value one can extract from frontier models and what was out of reach due to exorbitant legal fees AI allows me to do.

My days are spent looking for ways to sue companies, employers, people to the same vigor a security vulnerability researcher would.

Please note that I've had some legal education and that its probably not for everyone and I understand if some are upset my choices.

Before suing all your coworkers for garnishing their wages, I suggest you first consider how much fun you will have when you have to pay for the legal costs of the other party. But you will find that out for yourself. Anyways, I don't think it's your choice whether you want to take on huge law firms - that's usually the other party's choice. (I do hope most security researchers have a bit more training than "some education".)
I know someone who is with a leading firm. He is involved in a new multibillion USD matter every month.

The clients simply do not care about the multimillion dollar legal bills, since it is just a rounding error at that scale.

I find it hard to see AI being integrated at that end of the market.

But these figures measure exposure, not integration.

Ach, it's probably a mostly human-generated piece, but any time I see the 'Not [x] but [y]' formulation, I tune out.

There's just a whole bunch of the wrong people involved. AI will replace a majority of lawyers. It is inevitable. It won't be an LLM alone that does it though.
> As AI becomes capable of producing entire work products, the profession that has spent decades treating “I reviewed it myself” as the standard of care has no framework for what happens when that review becomes economically irrational. The ethical rules assume a human at the center. The technology is moving humans to the periphery.

What horrendous morals behind this article. Why would anyone advocate for prioritizing economics and technology before ethics, especially in something as important as law?

The "won't someone think of the all the poor people with no access to legal counsel" part sounds a lot worse once you realize they actually mean "won't someone think of the money we're leaving on the table by not getting some revenue from selling cheaper AI slop, uh I mean AI legal representation, to those who can't afford anything else".

I’m thinking the structural barrier is always going to be AI confabulations and misalignment.
> When AltFee won the ABA TECHSHOW 2024 Startup Alley competition, company representative Scott Leigh offered a simple proposition: our product will help you divorce from the billable hour.

This has been the promise of legal tech for more than a decade, and I have yet to see it deliver. Many lawyers got divorced from the billable hour long ago. For plaintiff's lawyers, for example, attorney hours are a cost center. Every hour a lawyer sinks into a contingency case is money out of the partners' pockets. They have every incentive to use legal tech to reduce their capital investment into these cases. But plaintiff's firms litigate cases more or less the same way defense-side firms that bill hourly do. That's a telling indicator of the shortcomings of legal tech.

AI might be different. I'm cautiously optimistic that agentic AI could do for lawyers the same thing it does for programers: act as a force multiplier that allows doing more work with fewer bodies. I'd love to be able to point an agent to years' of emails and prepare me a deposition outline for a witness. But the barriers to that are technological, not structural.

It would seem these structural barriers are increasingly become either more porus or more malleable as AI has brought more OSS legal initiatives to empower both attorneys and regular users. The Law and lawyers are being dragged kicking and screaming into e/acc.
AI Lawyers paid by crypto coins with robots as executioners.

The moment to off oneself...

Once again there seems to be this deep misunderstanding by engineers that all lawyers do is draft docs like an engineer writes code. And if you can automate the doc creation process, you have automated the attorney.

But this not true. It greatly depends on what type if lawyer it is. They can operate very differently.

When it comes to docs, lawyers try not to reinvent the wheel on every doc. It’s very risky. They want to use language that has stood up to the test of time. So they have templates and old contracts and they will piece together a contract from that - it takes no time at all. AI writing new stuff every time using new language doesn’t save much time and puts a client at risk because nobody knows how that language holds up in court.

Many attorneys have paralegals who do the actual work. Those paralegals are much cheaper. Efficiency arguments don’t often take them into account. Most legal tech companies don’t think about paralegals. But they are often the real workhorse in a legal firm. But automating them doesn’t really speed anything up: they do a lot of varied things and are not just mindlessly following some process like a factory worker.

So what do lawyers do? They advise clients, they project manage, they research, they talk with other lawyers, sometimes they go to court, they review work, they work on getting new clients, and yes sometimes they write legal docs. Then at the high end of things, they come up with legal strategies and solutions that AI has never seen before.

In corporate law, legal is often budgeted in. It’s expensive but it is often ROI positive. Legal mistakes are incredibly costly.

Having said this, are lawyers pretty slow to adopt new tech? 100%. Word processors replacing typewriters was seen with a lot of suspicion. Email too. The cloud too. I think AI will have a big impact on law - but it won’t be as simplistic as a little doc generator an engineer cooks up one night in his bedroom having never even met a lawyer. The law is quite gray and so even if you have found something impactful, attorneys manage risk by looking to see what others are doing. If there is social proof, they are more likely to accept it.

Data as an obstacle to legal AI is a human rights violation.

As citizens we are subject to federal, state, and local law. This consists of statutes, regulations, and common law. It is insane that the content of this law is locked away behind private paywalls.

How can one comply with laws when you aren't told what those laws are?

Lawyers, especially appealing to juries, require appealing to humanity. I'm sure the state will force software defense on us but it will fuck our neighbors over.

I do not understand why we do not abandon english common law and find common sense.

The LLMs are not even that good at law. I have a license agreement that I wanted to turn into General Terms and Conditions and they kept failing or rewriting the whole thing from scratch when a competent lawyer would just do a few pinpoint changes