Comment by broken-kebab | Hacker News Reader

Comment by broken-kebab | Hacker News Reader

Parent Full thread

I'd say that for now LLM-generated content is detectable, still at least. But anyway, "proven guilty" doesn't work as many people seem to believe. Often, if not most of the times, courts deal with cases with no full direct information which 100% prove guilt or its absence. If enough pieces of general picture suggest something, it usually considered proven.

In the same way a suspect may be declared guilty when e.g. there were reasonable expectation to double check a published image, a possibility to do it, obvious benefits from spreading fakes. Even if that image is genuinely looking convincing.

Probably, eventually somebody will try to use it to silence real whistleblowing. And it's unclear how this will work at all with expected advancements of generation capabilities. But I believe there's a public demand for such law.

I expect pretense of uncertainty to be used to get around: "we got this video from unknown source and we are not sure if it's AI-generated, but look like how this politician f*cks a pig"

> I'd say that for now LLM-generated content is detectable, still at least.

It’s really not, if someone spends a little bit of time to make something look human. For comments, literally say to your agent “write in the style of the HN user broken-kebab”, and it will pass pretty much any sniff test

(Cool nickname btw)

it's not about having 98% confidence. but 100% irrefutable proof.

You can go to reddit and cherry pick coments, that are made before LLMs, that have that exact LLM style.

Can you elaborate? I'm not sure I'm getting what is the point here. As I said, legal system rarely deals with 100%, it's not a requirement to make a law enforceable. Not telling the particular law is good or bad, just that "you can't certainly distiguish between human and LLM output" is not a strong argument against it. The conflict emerges when e.g. you are attributed words you have never said, or presented in a fake video, and there ways to argue about such cases in court other than just trying to analyze bit-by-bit, or doing histograms.

Simple - you need to assign blame, and you need to do it in a fair way.

Any assingement of blame of using unauthorized LLM generated content could be shifted to external contractor(preferably outside of EU's jurisdiction) where you yourself had no awareness of it. Contract of course explicitly forbids LLM content and requires labeling.

This is and extremely basic loophole that gives enough plausible deniability to any sane publisher/content farm.

Now you need to prove they were aware of LLM content being generated, as they cannot rely on LLM content detection methods due to high amount of false positives.

Even easier if you publish content from people writing under pseudonyms/anonymously. It devolves into he said she said kind of situation, and burden of proof is on accusers side.

Basically depending on enforcement this system either:

- only punishes small players that cannot afford to setup such chain

- is used as an arbitrary excuse to shut some outlets if the system is corrupt

Thanks, now I see. No, that's not how it works. If a legislation designates you as responsible, you can outsource compliance implementation, but not the responsibility itself. Banks can hire external companies to verify operations wrt embargoes, but if relevant authority found a violation, the most probable outcome will be penalty againt bank, not contractor. Moreover, at any time a bank may be ordered to demonstrate that they organized the whole process in such a way that they have substantial reasons to believe it works as required according to the current best practices (whatever it means), and be fined if they don't. Laws are quite often stupid, but placing blame is a honed process, and when the letters say they will kick your balls if anything, it's not easy to avoid this. Possibly we even will see publishers overrecation (as it observed in banking where clients or operations often blocked for no strong reason) and human content will be marked as LLM-produced, or "possibly LLM-produced" if they are given such option. Your concern about disadvantage to smaller players is not without reason as legally complicated environments often have such effect, but to be sure one needs to study the whole thing deeper. E.g. there could be exceptions for individuals/small companies.

Reply on news.ycombinator.com