I think it's more accurate to say that the EU legislates labelling AI content. Enforcement is another matter, as we've already seen with things like WEEE.
Isn't the whole problem that it is impossible to determine if content is AI or not? Doesn't this make it impossible to prove under 'innocent untill proven guilty' unless you're really blatant about it?
And isn't this easy to sidestep via plausible deniability and using middleman outside of EU borders?
So bad actors, especially ones outside of EU, can just continue to operate as is and ones within can just outsource the 'bad' work outside.
AI disclosure makes sense, but it should be a short list in standardized label format.ie AI used for code, assets, text, etc.
The problem is what we(as society )try to sidestep: is to have content creators(including major traditional media) be held responsible for content they provide.
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
Does it require innocent until proven guilty? Other standards. Not familiar with most EU countries, but common law countries have a balance of probabilities for civil cases.
"May contain X" labels are usually not allowed here with some exceptions. This has some downsides (like deliberately adding allergens to food because it's impossible to make sure there are absolutely none) but in general you can't just hedge your way past regulations.
> in general you can't just hedge your way past regulations.
Of course you can. The internet is absolutely polluted with cookie banners, including on the EU's own government websites, precisely for the reason to hedge your way past regulations.
Is it malicious compliance? Sure it is. But it's the EU's ill-conceived and misguided laws that have encouraged it.
Honestly don't even know how to feel about this regulation. I think in principle I like it. If one uses AI to generate content on a topic and is not actually checking its accuracy, why not force them to label it.
What practical benefits this might have in the long term I do not know. Perhaps an implicit trust in orgs/individuals who actively tag AI-generated content.
If one is known to tag AI-generated things as AI-generated, the un-tagged things have more weight as being original. To be honest I don't think this is a likely outcome.
Looking forward to all the AI bots on Facebook, Instagram, Twitter, HN, etc to suddenly have a label added to their posts.
I have a vision of millions of social media "influencers" suddenly finding out they have three actual followers. And they're all their own grandmother.
What if you find ex images that is not labeled should you assume its not AI? This will make everything not labeled as AI get more legitimate even though it's still at a similar level of trustworthiness. this a big problem.
The actual EU regulation on labeling AI content:
https://digital-strategy.ec.europa.eu/en/policies/code-pract...
I think it's more accurate to say that the EU legislates labelling AI content. Enforcement is another matter, as we've already seen with things like WEEE.
Isn't the whole problem that it is impossible to determine if content is AI or not? Doesn't this make it impossible to prove under 'innocent untill proven guilty' unless you're really blatant about it?
And isn't this easy to sidestep via plausible deniability and using middleman outside of EU borders?
So bad actors, especially ones outside of EU, can just continue to operate as is and ones within can just outsource the 'bad' work outside.
AI disclosure makes sense, but it should be a short list in standardized label format.ie AI used for code, assets, text, etc.
The problem is what we(as society )try to sidestep: is to have content creators(including major traditional media) be held responsible for content they provide.
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)
Does it require innocent until proven guilty? Other standards. Not familiar with most EU countries, but common law countries have a balance of probabilities for civil cases.
It's impossible to determine whether a killing is a murder or not. We still do it. And it's easy to sidestep by doing the murder outside the country.
I wonder if this will just end with "may contain AI content" labels on everything instead of achieving something useful.
Like everything in California may cause cancer.
"May contain X" labels are usually not allowed here with some exceptions. This has some downsides (like deliberately adding allergens to food because it's impossible to make sure there are absolutely none) but in general you can't just hedge your way past regulations.
> in general you can't just hedge your way past regulations.
Of course you can. The internet is absolutely polluted with cookie banners, including on the EU's own government websites, precisely for the reason to hedge your way past regulations.
Is it malicious compliance? Sure it is. But it's the EU's ill-conceived and misguided laws that have encouraged it.
Honestly don't even know how to feel about this regulation. I think in principle I like it. If one uses AI to generate content on a topic and is not actually checking its accuracy, why not force them to label it.
What practical benefits this might have in the long term I do not know. Perhaps an implicit trust in orgs/individuals who actively tag AI-generated content.
If one is known to tag AI-generated things as AI-generated, the un-tagged things have more weight as being original. To be honest I don't think this is a likely outcome.
Looking forward to all the AI bots on Facebook, Instagram, Twitter, HN, etc to suddenly have a label added to their posts.
I have a vision of millions of social media "influencers" suddenly finding out they have three actual followers. And they're all their own grandmother.
[dupe] Discussion: https://news.ycombinator.com/item?id=49132341
You will now be able to witness evolution akin to the one that "made in Germany" and "made in Japan" went through and in just short few years. Fun!
What for?
What not for? That people can't post AI fake stuff and pretend it's true.
What if you find ex images that is not labeled should you assume its not AI? This will make everything not labeled as AI get more legitimate even though it's still at a similar level of trustworthiness. this a big problem.