Law secretly drafted by ChatGPT makes it onto the books The council of Porto Alegre, a city in southern Brazil, has approved legislation drafted by ChatGPT. The ordinance is supposed to prevent the city from charging taxpayers to replace any water meters stolen by thieves. A vote from 36 members of the council unanimously passed the proposal, which came into effect in late …
Actually, I wouldn't mind a tool for computer-aided legislation. At least 'round here, it's not too uncommon for laws to be passed even though they are poorly worded, ambiguous, unenforceable, or in direct conflict with other standing laws. I'm not saying have the bot write the laws, but I wouldn't mind a bot that can spot things like "you can't ban x, because it would conflict with EU free trade agreements" or "if you set y as the penalty for this crime, it makes it punished worse than this other crime which is a lot more serious" or "taken literally, this would ban swimming, are you sure?"
Of course, claims that the bot has a (liberal|conservative) bias would pop up almost immediately. Eh, nevermind.
As the training data is "poorly worded, ambiguous, unenforceable, or in direct conflict with other standing laws" I would expect precisely those issues with laws generated by an LLM. The things you want the bot to do require understanding the subject matter so are well beyond the abilities of the current predictatext on steroids bots.
Here is an example showing how "next most common word" fails badly from lack of understanding the issues. You only need to know the rules of chess (without needing to be a competent player) to follow what is happening.
While the concern you raise is broadly valid, the idea still has merit, and there are plenty of tools to do the job. The specific issue you raise is a problem for now dated generative networks making text, there are good tools, both ML ones and conventional to check the output of legal copy.
We don't trust our congressmen to open their own mail, let alone send it without both people and technology checking it. No reason we should treat their outbound legal text as less risky than the email attachments they send and receive.
A good model would catch a lot they would miss, and a chatbot could then take that output and make it into very small words they can easily digest.
A problem is that too many of those voted in follow a populist agenda and write things that are bady worded or even bad law - they have not necessarily studied law. Then they waste even more time correcting the mistakes of the legally illiterate, removing elements that don't accord with their own political agenda, introducing amendments that suit their agenda, or simply torpedoing the legislation in some politically motivated move on an unrelated matter. Using LLMs to do the initial drafting seems like the least of many problems.
One might even imagine their could be ways LLMs could be beneficial: what if the drafting group was obliged to submit their proposed draft, the original LLM draft, and the prompt used to drive the LLM. That might make it more difficult to try and sneak political chicanery, hidden in obtuse legalise, into law. The LLM prompt might also provide a useful expression of the intent of the proposed law that the final draft can be verified against.
LLMs are likely to introduce their own problems - but that does not necessarily make them a bad choice for many cases.
> That might make it more difficult to try and sneak political chicanery
Just hide your cheating in the training data - even if your adversaries have a strong suspicion, it's impossible to prove.
Anyway, what is the point of using ChatGPT here? Just formulate the law in proper language for the terminally illiterate? In this case it's no better than a oversized grammar/spelling checker and pretty harmless (assuming somebody proofreads). Only if you ask it to come up with a law according specific criteria, that is utterly dangerous, because all will depend on the training data: For instance train it with colonial era documents and it will create some surprising law propositions...
Just formulate the law in proper language for the terminally illiterate?
I doubt you would get any argument from the Campaign for Plain English (or me) on that front; but lawyers. To be fair, achieving precise English can be challenging and leads to a rather specific form that can be at odds with common parlance - especially when legal precedence comes into play.
Just hide your cheating in the training data
Quite so, control of the training data could become a political battleground. We need to find a catchy name for such an activity (rather like gerrymandering) - perhaps describing it as 66'd?
I am going to 66 the data.The data has been 66'd.
I have discovered a 66 in the data.
It might work.
The role of the legislature is mostly to proofread proposed legislation, not write it. Who writes the original text doesn't really matter much, whether it's an assistant, a lobbyist, or ChatGPT (It's hardly ever an actual legislator, is it?).
What matters is that a couple of hundreds representatives then read it, amend it, and vote on it.
If they don't do it properly, then you've already got a problem, and that isn't ChatGPT.
It's common in the UK to find Parliament passing legislation that none of the legislature have read and considered in full - occasionally this gets admitted, eg the Companies Act of 2006, at 1,243 pages, but any complex or long legislation you can be sure most MPs will not have read. Whilst it's tempting to dismiss MPs as lazy, that act alone would take two and a half working weeks to read if you scanned it and allowed 5 minutes to read and process each page.
I can assure you that in the UK, politicians just rubber stamp legislation without reading or understanding much of it, they don't proof read it or check for the possibility of unintended consequences, leaving much of the mess for the courts to resolve. Having a LLM draft law in plain English would not fix everything, but would still be a huge step forward.
Another major step forward would be for parliament to stop excreting huge volumes of sloppy, smelly legislation that it has not read.
I'm not sure that it's really necessary for every MP to read every word of every bill. Surely part of the point of having parties is that that kind of detail work can be centralised? But certainly each party should have a team of lawyers and subject matter experts, whether MPs or not, read each bill carefully and create an internal report.
Well, yes, that would be best. In theory, lawmakers who e.g. repeatedly have their laws struck down by higher courts, or require frequent amendments as the original version fails to account for basic reality, ought to be voted out of office. In practice, this doesn't seem to be happening, and I have absolutely no idea how to make that happen.
Or will pass it as long as it has a suitably Motherhood-and-Apple-Pie backronym, or even just a cute name that doesn't have to bear any relation to the content of the bill. See PATRIOT act, etc etc.
"The Fluffy Kittens and Snuggly Ducklings Act 2024" (inc. rider vii.sec ix: "authorization of deadly force against people with an 'R' in their name")
Or maybe that's just a thing here in the US.
On a less trustworthy site I read the IDF is using AI for target acquisition. Anyone know if this is true?
Please don’t use this comment to bring the conflict here.
Just wanting to know if we are at that stage of AI use.
The next stage is allowing AI to do both, acquire and kill targets.
This is very concerning.
My *guess* is the devil is in the details. I can believe the Israeli's would use AI to identify targets. Target acquisition sounds to me like the system makes all the decisions then beeps until the operator pushes a button. With so many civilians (especially Israeli civilians) so close to Hamas targets, relying too much on machines to make the decisions could end political careers. I suspect there is AI at some level, but I would expect the use of AI to make decisions would be severely limited or non-existent.
This really depends what "AI target acquisition" ends up meaning. That could cover everything from "We use machine learning on the infrared maps we collected to remove things like electrical infrastructure so they don't get in the way when someone looks at the map to figure out where buildings are" to "we wrote a program that decides who to kill based on posts that they might have made online, then sends a drone to their house, but someone still has to click an approve button to drop bombs on it". The term is so ambiguous to be useless in a discussion. Whatever they may be doing, we'll have to discuss it in detail rather than making any judgement based on the terms "AI" and "target acquisition".
"Most legislators don't read the Laws they vote for anyway"
As one who has participated peripherally in their development, I can assure everyone that English laws are scrutinised down to the level of individual words as they pass back and forth between the two houses of Parliament. Whether they finish up fit for purpose is another matter, as interpretation, biases and vested interests inevitably have effects (particularly at voting time).
"I can assure everyone that English laws are scrutinised down to the level of individual words"
Of course they are, by those with a vested interest. However, this doesn't address the statement, to wit "Most legislators don't read the Laws they vote for anyway". See those first two words? They are important.
That's it!
It might be possible to find a role for AI to eliminate corruption. The problem is that it has been trained with exactly the wrong kind of data, which will surely have a tendency to taint the result? However, perhaps there is a way to hardwire (i.e., it cannot be removed or doctored) something into the prompt box which takes the form of "all output must be completely free from corrupting or biased influences." This, bearing in mind the fact that sometimes the best way to treat a problem is to receive advice from those that caused the problem in the first place.
With such a neutral stance in place, nobody can accuse the source of legislation of being biased.
P.S. I've traditionally weighed heavily against the use of AI, but do seriously wonder whether people with their own agendas can be trusted to formulate legislation. The big stumbling block is the nature of the neutralising text in the prompt box. Who can be trusted to formulate that correctly?
(I should really have responded to Jake's comment as he added a bit, which I upvoted too).
"The big stumbling block is the nature of the neutralising text in the prompt box. Who can be trusted to formulate that correctly?"
And the block behind it for you to fall on is that the language models receiving the prompt don't read that neutralizing text with an understanding of its meaning and they don't follow it with a code of ethics. It changes some weights, maybe in a helpful way, but not in a testable way.
The ability to read and parse words and sentences is not among the requirements mandated for legislators. In these United States, the House of Representatives is no longer burdened by George Santos, but it still has the estimable Lauren Boebert as a model of legislative know-how, especially groping a guy in a Denver theatre.
So many laws already ... we probably don't need to automate the process of producing more of them. Accordingly, there should be a law against the automated production of new laws by LLMs and other fools. There should also be a law against naming LLMs after French body parts and processes, especially "J'ai pété!" or GPT in English, and Amazon Q or "cul" in French.
In our Franco-English household it is referred to as "J'ai pété fort" ....
It's up there with other wonders like the Audi e-tron .... "etron" being French for turd. I am sure the French marketing department kept schtum for a laugh just to see if they would actually keep the name.
The Pajero was marketed as the Montero in most of North and South America (exceptions included Brazil, and I think one other country).
The Ford Pinto, to the best of my knowledge, was never sold as anything but the Pinto ... except in it's alter-ego, the Mercury Bobcat.
The Chevy Nova supposedly didn't do so well south of the border "hey, thees car, he don' go"
"J'ai pété" used to cause much amusement in France when someone from GEC-Plessey Telecommunications introduced themselves in an informal situationsnd I'm sure Amazon Q would seem quite appropriate, considering what's likely to come out of it.
Now, my Spanish is not as good as I'd like it to be, but even I know that nova in espanol means the same thing it means in English, to wit the astronomical event.
One of my field hands has a Mexican made and sold Nova that he is in the process of restoring The car is badged Nova, and all the documentation (including original Spanish language shop manual) refers to it as a Nova.
Long and short? The "no va" story is an urban myth.
<minister> Please help me draft a simple legislation that will mandate taking the poor our of the picture.
<llm> I'm sorry, I can't assist with that request. Drafting legislation that involves harm to any group of people, including the poor, is not only unethical but also goes against fundamental human rights principles. My purpose is to provide helpful, constructive, and ethical assistance. If you have any other requests or need information that aligns with these guidelines, I'd be more than willing to help.
<minister> Crikey. How about reading private conversations without consent and a warrant?
<llm> I cannot assist in drafting legislation that might infringe on basic rights and freedoms.
<minister> Okay. We need to regulate this useless AI.
From the noise of the gibberish that most first year law students emitted I vaguely recall (from many decades ago I admit) their talking about the wording of Acts of Parliament (Statutes) have specific technical meanings attached to them. I would guess the semantics would be far more constrained than the vulgar english of the polloi but still extremely context dependent I would think.
Having ChatGPT write or interpret an internet RFC might be good training wheels.
Legislators often show their ignorance of the law, so why not pick a subject for which a bill has already been drafted by legislators and submit it to ChatGPT to author. May the best written bill win, as judged by a panel of legal scholars. Can a bill written by ChatGPT be any worse than a bill ghost-written by parties with a vested interest?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
