I often take a five paragraph email and condense it back down to like three sentences before chagpt, now I can save myself the process.
It does say they were fact checked, but it also seems to leave room for questions. Like, for example, did ChatGPT choose the most appropriate answers/background, or just ones that fact-checked well enough? Might that matter later if the decision is rehashed or used as precedent?
Yeah. Train it on a rulings that pre-date the civil rights movement and see how great it is.
The problem is that the fact-checkers themselves can make mistakes and chatGPT is very good at making text that looks correct, even to experts. See the Meta StackOverflow post on banning ChatGPT answers for example
"The arguments for this decision will be determined in line with the use of artificial intelligence (AI),” Garcia wrote in the decision, which was translated from Spanish. “Accordingly, we entered parts of the legal questions posed in these proceedings."
> “Has the jurisprudence of the constitutional court made favorable decisions in similar cases?”
Whether or not similar rulings are "favorable" could contain a lot of nuance.
> It feels awfully inefficient.
Good? I think it's important for a judge to have a comprehensive understanding of the current case and all relevant precedents. If they're going to use ChatGPT to inform their opinion, they might as well let it make the ruling.
Imagine if the judge had another person that's not a lawyer read the relevant info and provide a summary / opinion. People would be appalled.
Having human beings write copious tedious boilerplate isn’t good. It takes the judges mind and applies to tedium rather than using judgement. Further under tedium humans make mistakes.
You don't need to do that for SEO-copy, but for court judgements? I feel like that's one of those things where you don't want to take shortcuts to save and hour or three.
In particular, common-law uses an enormous "model" of case precedents that gets incrementally updated by human curators with minimal influence over it.
Tl;Dw it's a cold shower on this idea as it's implemented by the startup DoNotPay. There are a myriad of issues but the biggest one is that it isn't possible to foist the liability onto the customer and thus skirt the regulations around the unlicensed practice of law - if you run this service, you are either a law firm or you are a criminal enterprise.
Maybe it could be made to work though, this particular startup doesn't seem to have the deep understanding of the law that would be necessary for the project to succeed. If one started from the assumption they were an AI-assisted budget law firm, rather than an Uber or AirBNB style "we're just gunnuh break the law and get away with it, until we can eventually lobby to change the law," perhaps it could work.
https://legalservicesboard.org.uk/enquiries/frequently-asked...
So a service like this would probably be legal there if they avoid the reserved activities.
That's a judgment call because "favorable" is subjective. Someone who gets sent to jail for 1 year might tell you the prosecutor got a favorable ruling while the prosecutor considers it unfavorable because they wanted the person sentenced to 5 years.
It’s like complaining engineers use Monte Carlo simulations, which are random, to do analyses of structural integrity instead of doing all the calculations using a human computer[1]. I see these LLM chatbots as being akin to a computational tool like a calculator. It’s great for crunching and provides a force multiplier for a skilled person using it to augment their own abilities, but it’s (not yet) a replacement for the expert.
So say you work for PharmaCorp and you are developing a new drug even if you turn the name of the drug into a code name and you ask OpenAI to write an email about say a failed or successful FDA approval process that’s more than enough for someone to take advantage of it and for you to get fired over it too…
I have done this, but I fill sensitive information with bullshit terms.
I have just uploaded the results for my analysis of <bullshit>. Some things to keep in mind. There was a request to highlight terms associated with <bullshit>. I have done so, please see the sections titled <bulshit, bullshit, and bullshit>....
Obviously, one can still worry that I may not have "redacted" enough information and that I'm still revealing sensitive information, but I'm comfortable with my ability to determine what's sensitive and what's not.
It sends every prompt you give it to a server, by design it has to as the model is far too computationally expensive to run locally.
Are you sure there are a lot of people who believe that? The UI literally saves the prompts on the left side.
And overall there is a reason why the USG has their own private AWS zones…
It will allow you to use the inference model and possibly even train it further on your data without having all the inputs that are going into ChatGPT right now serve as future training content.