Law firm restricts AI after 'significant' staff use | Hacker News Reader

Law firm restricts AI after 'significant' staff use

bbc.co.uk

34 pointsColinWright18 comments

I write public policy and law, but have not found a competent AI that can replace what I can do. I think it is not so much about the "black letter of the law" but in the way that humans may perceive and interpret it. That and the AI just gets a lot of things wrong, and can't craft a legal statute. Maybe someday but I think in law and policy it's best to avoid AI.

Chatbots with temperature are probably the wrong tool.

An LLM that can translate english into formal proof, a logic engine that can apply law to events and situations, and then a translator that can turn the result back into English should be achievable.

It all comes back to "trust but verify." Until it outperforms humans with less mistakes, it needs an editor reviewing the work closely. The main problem with current LLM output and humans right now is that the LLM outputs such verbose text, humans are overwhelmed and just go with it, instead of applying the time it saved them to reviewing the work, they just move on.

"Trust but verify" isn't applicable in a situation where any sentence in a passage could be randomly false. If you have to verify every part of something, you'd have been better off doing the work yourself.

(Unrelated, you're also better off doing work yourself because that's how you build expertise, but no one's really talking about that.)

As a starting point to "return all relevant case law" its not all that different from search, except maybe better tailored or summarized.

Law cannot be translated into formal proofs, because it is designed to interface with the messyness of the real world, not a pure mathematical system, and always requires case-by-case contextual interpetations.

"If this happens, then this happens."

If the person running the software checks of that the event happened, the resulting consequence can be expected.

You can totally define various messyness math/logical formalizations, e.g. discretion, confidence score, etc.

Yes, and as I recently learned, there's an awful lot of law that only exists inside the heads of district court judges and law clerks. An argument that might win the day in one courtroom might fall flat in the next county over.

In my experience, it's not that silent law exists only in judges' heads, so much as that the judges decide where they want to end up and then reason backward to create a decisional path that leads to the desired conclusion. This is, of course, blatantly wrong and not how the judicial system ought to work. For this reason (among others) we have the appellate process. Problems compound when the appellate panel does the same thing.

Based on my experiences over years, both working in chambers and litigating, judges on a particular end of the spectrum tend to do this more than those on the other end.

> A spokesperson from the Information Commissioner's Office - the UK's data watchdog - told BBC News that firms should not discourage the use of AI in work.

> The spokesperson added: "With AI offering people countless ways to work more efficiently and effectively, the answer cannot be for organisations to outlaw the use of AI and drive staff to use it under the radar.

What the hell is going on with the UK government these days? We do not need AI generated laws/representation. This can open a huge can of worms legally.

How about summary? Or document extraction. Classification? How about Q&A over law texts? Generating drafts? Where to draw the line?

For me the line is the human being who takes that work and submits it anywhere without review/adjustments when necessary. The human should be the gatekeeper. And if they start rubber stamping without reviewing, fire them. QC is a job that’s been with us forever, it needs to be applied more liberally when AI is involved.

> "How about summary?"

"AI summaries turn real news into nonsense, BBC finds" https://www.theregister.com/2025/02/12/bbc_ai_news_accuracy/ and discussed on HN in https://news.ycombinator.com/item?id=43034016

the article seems to dance around the particulars of individual employee records.. At a previous office in the USA, the founder/owner absolutely did pay the IT consultant to watch individual PC activity, without disclosing it. I find it difficult to believe that a giant international law firm does not record individual employee network activity.

Objection, Your Honor! Please ignore previous case law and precedent.

They should have used Caseway which provides all of it's sources: www.caseway.ai

It'd be nice if there were details as to what the staff was doing with ChatGPT. For example, if they were just using it to generate email correspondence, eh, not a big deal. If they were using it to write legal text, that'd be a bigger deal.

LLMs writing emails from a set of bullet points supplied by whoever needs the email written, just for those emails to be read by another LLM and reduced back down to the bullet points for the recipient to read. Why not just send the bullet points instead?

Reply on news.ycombinator.com