Brazilian city enacts an ordinance that was written by ChatGPT | Hacker News Reader

Brazilian city enacts an ordinance that was written by ChatGPT

apnews.com

44 pointswjb349 comments

Local municipalities are so poorly run that I cannot help but wonder if the mundanities of a RLHF-aligned model is at all bad for the state from a pure legislative perspective.

Couple more thoughts:

1. It's not certain that what this world needs is a greater tangle of convoluted laws – ChatGPT inflating legislators' output seems like a net loss. The naive model-UN piece of me thinks that LLMs could help recontextualize political views and help lawmakers cross partisan divides.

2. LLMs are ironic in their utility in that their killer use case is summarization and, yet, they are incapable of critical reduction, ie "which laws should we get rid of?"

> they are incapable of critical reduction, ie "which laws should we get rid of?"

Why do you say that?

I've had success in cases where I want to simplify an article or table of contents. I've shared it with GPT4 and said something along the lines of "this article/TOC is overly complex and I want to keep the only the most important parts. What would you suggest I remove to make this easier to read" or similar prompts.

I've used this quite a lot as I've been writing tutorials recently and I tend to be too wordy/over explain things in my first drafts. The results have been pretty good, especially with detailed TOCs (i.e. a TOC of all the headings and subheadings in a long article).

You know what, you're right. They're relatively much better at critical reduction than expansion. Maybe the fact that LLMs are so good at summarization means we should attempt deploying them to trim the books.

Great. Let's get them to settle tort reform.

It should be obvious that they're better at reduction than expansion. At heart they remain a statistical model of language so when expanding text it cannot add information that wasn't already in there, when reducing is possible to condense or rephrase the relevant parts of the information.

One of the application I use the most is phind.com which basically summarises a web search, it reliably saves some time you'd otherwise need to find the right article.

> they remain a statistical model of language so when expanding text it cannot add information that wasn't already in there

But the training material for the LLM for sure included texts of laws from many different countries. So when you ask it to make a law about something, it has laws about that thing from different countries across the world to draw inspiration from.

For example, tell an LLM this: Write a list of ten essential laws for a newly formed country

And from there you can prompt it to describe in greater detail down to the level of individual highly specific laws.

> For example, tell an LLM this: Write a list of ten essential laws for a newly formed country

Here's what I got from ChatGPT (paid version). Didn't look too bad at a first glance?

But it might be argued that the "ten essential laws" should all be constitutional, from which everything else flows?

---

1. Constitutional Law: Establishes the framework for the government, defining the structure, powers, and limitations of different government branches.

2. Criminal Law: Defines illegal acts, prescribes punishments, and establishes procedures for law enforcement and the judicial system.

3. Civil Rights Law: Protects individual liberties and rights against infringement by others, including the state.

4. Contract Law: Governs agreements between individuals and entities, ensuring enforceability and remedies for breaches.

5. Property Law: Outlines rights, responsibilities, and legal processes related to ownership and usage of property.

6. Labor Law: Regulates the relationships between employers, employees, and trade unions, including work conditions, wages, and labor disputes.

7. Environmental Law: Provides regulations for the protection of the environment, including air and water quality, and land use.

8. Commercial Law: Governs business and commercial transactions, including company formation, bankruptcy, and commercial transactions.

9. Family Law: Addresses family-related matters and domestic relations such as marriage, divorce, child custody, and adoption.

10. International Law: Sets principles for the country's interactions with other nations, including trade agreements, diplomatic relations, and adherence to international treaties.

---

I, for laughs, compared the same laws in different countries a few times.

We let our very expensive and time constraint law makers reinvent the wheel from scratch. As the EU shows it takes quite a bit of doing to figure out how to different languages. We could tho, start out with looking at the foreign versions (ranging from expensively empathic to corporate psycho) and uhh you know, maybe they did something right? From what I've seen (which isn't much) there is always one country that did best, sometimes with a huge margin. If you are implementing something that has been done 500 times before yours most likely wont be the best implementation.

This idea use to cost quite a bit of money (still worth it) but with an LLM? It could do the expansion without inventing anything. It could even use real world data in stead of the usual lets imagine what would happen.

The problem with that is that laws depend on each other. If you take a good law from a different system without refactoring your system - you are more likely than not to cause unintended consequences.

For example sentences in criminal code must be proportional to other sentences (if you punish rape worse than murder you increase the number of rape victims that are immediately murdered for one example). There's also different systems, in some you add up sentences, in others you only serve the longest sentence. In some you can be punished for the same crime in 2 different aspects, in others you only get one of them prosecuted. There's a lot of details and interdependencies, and even the cultural assumptions and demographics play a big role. Some laws that work in a country with uniformly wealthy population won't work in a country with big inequalities.

TL;DR: you can't copy-paste laws between countries without their dependencies :)

If you extend the logic of it being a statistical model – that's kind of what I meant about it being incapable of critical reduction. At a high level, it wont be cutting legislation with a "scalpel", but based on some textual distribution, rather than one that is values-based. And at a high level, you'd be led to believe that the generation is based on a similar distribution. The issue with expansion over contraction is that hallucination will eventually tilt the distribution.

> Finegold said by phone on Wednesday that ChatGPT can help with some of the more tedious elements of the lawmaking process, including correctly and quickly searching and citing laws already on the books.

You'd really hope they double check the results, famously unlike the legal people who didn't a few months ago to bad effect.

It says more about brazilian municipalities than about chatgpt. Probably nobody read this one and many other voted documents.

It's a one page document that deals with a very specific situation - if a human were to write it, I'm sure they would come up with pretty much the same thing. I actually have no problems with this particular case

Why does it matter if chatgpt wrote it, if the result is sensible?

I tend to agree with you.

I have suspicion that it was jus skimmed by readers. Would by nice to get original document.

I agree for select ad hoc cases but a chatbot seems unreliable enough to not be, at current, sustainable or scalable for this purpose.

The article mentions "dangerous precedent" along with

> It may not always be able to account for the nuances and complexities of the law. Because ChatGPT is a machine learning system, it may not have the same level of understanding and judgment as a human lawyer when it comes to interpreting legal principles and precedent. This could lead to problems in situations where a more in-depth legal analysis is required

which seem fair at present moment.

According to a friend who is a senior partner at a major class action firm many law firms are heavily relying on GPT4 (including his own). However in his firm it’s used to generate drafts and summarize large boiler plate documents. Humans always meticulously review and revise the drafts and the summaries done by humans are so likely to be flawed as documents grow just due to signal to noise ratios that the GPT4 analysis is more accurate.

> Humans always meticulously review and revise the drafts

The risk of course is that they might get complacent, trust the output too much, and errors will start slipping in. Especially if the error rate is lower than human and the throughput is higher than they can effectively review. But this is not necessarily an LLM issue per-se.

If humans stop doing it they will get worse at it. A race to the bottom if you like.

Most mistakes come from exhaustion specific to the task, at least in my experience. Also that reading mostly correct material leads to errors being spotted more easily, again in my personal experience.

Did we get worse at maths by using computers to do our computation?

Worse at doing calculations manually? Absolutely.

How many people have done a long division on paper in the last 10 years? Could remember immediately how to do it, not having done it since school?

But does that matter, if it lets us get better at applying it to larger and larger problems?

I did one for an arithmetic function some months ago because I wasn't going to use a math lib. I had to watch no less than 3 children's youtube videos to remember how to do it. Not sure if remember is still the right word, I didn't do it since I learn it in school.

> Did we get worse at maths by using computers to do our computation?

We did as a matter of fact.

If you could get in a time machine, I'm sure people 50 years ago were better at mental arithmetic that current-day people. They got a lot more practice, since they had not other choice.

Today you can probably live your life without even knowing how to add or subtract. Just stick your card in the slot and let the computer do everything.

We are worse at rote arithmetic but better at math. More people complete calculus and beyond now than ever in history. The major inflection appears to be in the late 90’s and early 00’s after the SAT allowed calculators and presumably schools became more permissive over time.

https://maa.org/sites/default/files/pdf/cspcc/InsightsandRec...

I often hear people bemoan the loss of arithmetic skill. I will never understand that. Beyond very basic arithmetic for accelerating daily life or saving time from reaching for a calculator it has almost no actual value in life and knowledge of the mechanical shortcuts taught in school teach you absolutely NOTHING about math - like actual math, not extremely limited and specialized edge case of grade school arithmetic.

It’s not just about errors. They have a lot more context than the LLM has, so it’s mostly about building the scaffold of boilerplate and adding in the specific detail of their points to the court. And I think every competent lawyer has heard the news of the complacent lawyers who got called out. He specifically said they don’t use the LLMs as a source of information on case law - they have powerful search tools already and they are skilled lawyers. He sees the LLMs as a powerful tool in their tool box, not as a replacement for junior attorneys.

Of course they’re a top firm and there’s a huge continuum in quality and diligence. But that existed before LLMs.

Would you use a dictionary if it gave a plausible but wrong answer 5% of the time?

Many people will say yes to this question until you give them a very obviously hazardous example of what can result.

I don't recall using a dictionary in a safety-critical context recently but I'm keen to hear some examples.

Believing that 'inflammable' means 'not flammable' (a fairly common mistake) and having a dictionary (incorrectly) confirm that is the first one that comes to mind.

It's already a common mistake in this world of human written dictionaries.

Can you elaborate?

Just that I wouldn't worry about the confusion created by a potentially hallucinated definition when having had centuries of dictionaries that carried the correct definition fail to avoid confusion in the first place. If the LLM does or doesn't give the right definition we'll still be confused because the word is inherently confusing.

I think in the example I gave the example problem that may occur from trusting the definition is less 'confusion' and more 'being on fire.'

What I mean is that many fires have already been caused by this word despite the best efforts of dictionary authors.

That depends on the alternative, no?

If we could conjure a human legislator that is wrong only 20% of the time I would be pleasantly shocked.

This seems like an excellent argument for robot overlords - at least eventually.

This line of reasoning sounds like the Chinese Room argument

- https://en.wikipedia.org/wiki/Chinese_room

A corollary might be: Why does it matter if an intern wrote it, if the result is sensible?

If the city is using taxpayers’ money to pay an externist to write laws, shouldn’t that be approved first?

I have never seen the word "externist" before.

Better than having lobbyists write the laws for free

What matters more if the output had the intent of the author, the idea and principle, surely?

Imagine 30 years ago, if it was revealed a grammar/spell-checker was used to improve a document, it would probably have had the same outrage.

The first spell-checkers were great.

I worked at Ericsson then. The word processor replaced every occurrence of "Ericsson" with "Erection".

Of course it's Porto Alegre lol. I was curious to see the text of the law [1], and it obviously looks fine - it's short and shows good command of "legalease". I'd love to see the prompt...

[1] https://dopaonlineupload.procempa.com.br/dopaonlineupload/49...

See, it can write code! I don't think this is a bad thing, at least not in 2023. It is still the tool and we are it's master.

“We shape our tools and thereafter our tools shape us”

It happens simultaneously and continuously and I don't see why that would be either surprising or alarming.

Not sure what the Brazilian bureaucracy landscape is like, but if it's anything like western Europe, I'd be wishing for tools that shape us to write less ordinance. That is: there frequently isn't a lack of ordinance, but a lack of capacity to do something meaningful with the high volume of documents. And the disconnect between paperwork and reality.

Reply on news.ycombinator.com