News for hackers. Stuff that matters.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by seydor on Monday June 17, 2024 @04:15AM (#40703401)
There will be no French Meta AI , but the English one is getting better and better. And the head of Meta AI is a Frenchman. This is war
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Havoc on Monday June 17, 2024 @04:07AM (#40703355)
EU regulators sure do seem to be the only ones around with bite lately.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by mdrzn on Monday June 17, 2024 @03:13AM (#40703010)
Thank you EU! I haven't had to do anything and they defended my interests.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Hamuko on Monday June 17, 2024 @03:33AM (#40703124)
I feel like they're at best paying lipservice to my interests considering that they're planning on scanning every single message that I send to "protect the children".
Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @03:40AM (#40703175)
Given how ardently they defend them, they do more than pay lip service. I think what's happening is "we want to be the only ones who can invade our citizens' privacy".
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Hamuko on Monday June 17, 2024 @03:52AM (#40703249)
So, ardently defending my interests (privacy) meaning selectively defending them?
Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @04:04AM (#40703330)
Yes. The ones they care to defend, they do so ardently.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by hef19898 on Monday June 17, 2024 @03:42AM (#40703195)
I'd very much prefer my privacy to be respected universally. One has to note so, that there is a very significant difference between a democratic government having rules in place to intrude on it and some capitalistic corporation doing it for profit.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by pastage on Monday June 17, 2024 @03:56AM (#40703280)
Maybe what is missing is some simple rules, that we should all follow, with a good reason why they matter.

The only way to achieve that is to make the laws enforceable. They should probably try hard to fine some government organisations as well. The laws though littered with necessary exceptions for government are a step towards respect for privacy.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @03:48AM (#40703229)
I'd prefer it too, and sure, but it's not that the EU pays lip service to defending your privacy against corporations. They actually do defend you, very effectively.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by hef19898 on Monday June 17, 2024 @03:52AM (#40703251)
On that we agree, and I am happy for it.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by freilanzer on Monday June 17, 2024 @04:21AM (#40703430)
> considering that they're planning on scanning every single message that I send to "protect the children".

That's one of the large issues with the EU. Against foreign surveillance and espionage, but internally it's a different story.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by vasco on Monday June 17, 2024 @03:43AM (#40703203)
I wonder if this will be one of those situations where because everyone else does the wrong thing, the people that try to do the right thing actually end up worse of. Mostly because if models will not be trained on european data, the internet and cultural hegemony of the US will go into overdrive with all models churning out only US-centric views.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by mrtksn on Monday June 17, 2024 @03:58AM (#40703289)
IMHO when the gap is too wide it sticks differently, it creates a caricature or an utopia in a similar way how Hollywood created an imaginary US utopia everywhere in the world or how the Anime culture created a caricature of Japanese people.

I don't know if it is good or bad thing in total. I remember idealising the USA as a teenager, not only its technological might but imagining that everybody lives in large detached homes only later to find out that Americans who live like that can't do anything without a car, don't have groceries, restaurants, cafes, museums, cinemas in walking distance.

My bet is that it will simply stick differently and will create an image of America through which Europeans will reflect on themselves and have conversations about how things can be done differently.

I don't believe that hearing different viewpoints automatically makes you to subscribe to them and that's why I'm very ant-censorship and I think it's a mistake to block or remove content from the internet. Real people don't get influenced the way priests influence villagers in Age of Empires 2.

Actually, I think a bit of cultural decoupling from the US will have positive effects as the US itself is in cultural crisis.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @04:03AM (#40703325)
That can be an issue, although - I think - the European approach will be to block private and semi-private data (like fb posts), but allow public data (articles etc).

In Poland we have a team of people who work on SpeakLeash which is 1TB of polish texts freely available for anyone to train their models on - http://speakleash.org/en/speakleash-a-k-a-spichlerz-english/

I can imagine models getting their knowledge and intelligence by getting trained on global - english and chinese - data, and then learning country specific languages and viewpoints from way smaller datasets.

You can see that already - gpt was trained on way less Polish language material than english, but it’s almkst just as smart and fluent in both.

As for cultural quirks - here it’s more about who does hrlf and alignment than the source material.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by mike_hearn on Monday June 17, 2024 @04:36AM (#40703516)
That isn't the European approach. From the article:

"Meta on Monday said it hoped to use Europeans' data to train its models. It promised to only use public posts and comments — not private chats and DMs — and to not use any content from anyone under the age of 18. Crucially, the biz said it would give Euro folks a chance to opt out; a safeguard not extended to the rest of the world."

Private messages have never been in play here. The regulators are arguing that stuff posted for the whole world to see is not OK to use for training purposes if it was posted to Facebook, but it is OK if it was posted to a blog.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by jstummbillig on Monday June 17, 2024 @03:57AM (#40703283)
I don't see Europeans not wanting to use the AI products build with US data.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Rinzler89 on Monday June 17, 2024 @04:30AM (#40703489)
Because US tech companies have no restrictions in selling the US data to EU consumers.

And how can EU consumers know which of the data they get is US data or EU data? I'm European but almost everything I write online even about my home country, including this comment you're reading, is in English on US developed platforms, not in my mother tongue. Now, is that US data or EU data?

Re:Meta won't train AI on Euro posts after all, as watchdogs... by cynicalsecurity on Monday June 17, 2024 @06:25AM (#40704111)
Don't feel so puffed up, my American friend. We really don't want to use it, but we have no other choice.

AI is devaluing you as an employee. We are only at the beginning of the AI apocalypse. If things keep going this way, all that is left to meatbags like you or me will be hard manual labour under strict AI supervision. This is not the future we asked for.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by hef19898 on Monday June 17, 2024 @03:55AM (#40703268)
Nah, China already sealed of their internet as did Russia and to a degree India. If we have, in the end, a dead US-centric internet, well, as soon as people start ignoring that it wpuldn't be a real loss to humankind. Until we reach that point so, we are in deep shit. And LLM spam and crap content being US centric is least of our problems.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @04:05AM (#40703338)
Did they really? Can’t we scrape chinese or russian internet just as well as western internet?

Honest question. I know they have firewalls, but these are one sided as far as I understand. No?

Re:Meta won't train AI on Euro posts after all, as watchdogs... by mike_hearn on Monday June 17, 2024 @04:38AM (#40703527)
Bandwidth in and out of the country is heavily throttled in both directions and packet filtering applies on both. It's very hard to crawl the Chinese internet from the west and it has always been that way. If you want to do it you have to do the crawl from inside China, and then you're open to having your datacenters raided and software stack stolen (this isn't a theoretical concern).
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Swizec on Monday June 17, 2024 @03:55AM (#40703266)
> the internet and cultural hegemony of the US will go into overdrive with all models churning out only US-centric views

I saw a fascinating outcome of this at a conference recently. Polish speaker made an IRS joke to a room full of Europeans at a conference in Europe. Everyone got it.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @04:11AM (#40703377)
That’s true. As a Polish person I will often say “IRS” or “polish IRS”, when I mean a more generic “tax office”.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by cynicalsecurity on Monday June 17, 2024 @05:59AM (#40703974)
Do you really live in Poland? I'm a foreigner (well, not any more) and I've been always calling it urzad skarbowy even when speaking with other foreigners. If I called it IRS, no one would understand it.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by vasco on Monday June 17, 2024 @07:06AM (#40704368)
I have the weird situation where in my country the IRS is called the IRS even though the letters stand for words in my language for "income tax" it ends up being the same acronym and its used interchangeably with the actual name of the government entity.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by abc123abc123 on Monday June 17, 2024 @04:12AM (#40703386)
And all they needed was an opt-in. That's really all they needed.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by posix_monad on Monday June 17, 2024 @04:17AM (#40703415)
That solution doesn't work for Meta. Who on earth would choose to opt-in?
Re:Meta won't train AI on Euro posts after all, as watchdogs... by wcoenen on Monday June 17, 2024 @03:49AM (#40703237)
Note that this was about _public_ posts. What is stopping everyone else (other than Meta) from using your public posts as training material? How will you even know?

All that data is already being scraped and processed into datasets by many companies: https://research.aimultiple.com/facebook-scraping/

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @03:41AM (#40703183)
What's the big deal on AI being trained on my data? The only argument I can understand is that a megacorp makes money from content I created, even if that would only be a net a few dollars for me, if someone were interested enough to pay. What other issues are there?
Re:Meta won't train AI on Euro posts after all, as watchdogs... by johnnyanmac on Monday June 17, 2024 @03:59AM (#40703297)
>The only argument I can understand is that a megacorp makes money from content I created

for professional content creators, that is exactly the biggest problem. That's why the artist community is getting a huge brunt of the pushback. You don't want your content production tools like the Adobe suite to retroactively say "yeah, by the way we are scanning your content as its produced and using that to make more money".

Even if you as a person don't mind, that's a huge NDA breach for a companyu. You don't want AI training to produce assets similar to the in-production media that isn't even out yet. It'd be a huge leak at best and a DOA at worst. And I'm sure you've seen the year+ of debate on copyright issues.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by Rinzler89 on Monday June 17, 2024 @03:46AM (#40703221)
The EU level problem is that LLMs are trained on your publicly scrapable reddit or HN posts since you're not getting paid for your contributions either way.

The way I understand it is, the laws are made to cover cases of LLMs being trained on work of authors and artists who normally make a living from the content they create such as articles, books, music, images, videos, etc. whose work is normally copyrighted but LLMs don't give a shit and can just scrape and "learn" your content, style and patterns and then regurgitate it with slight alterations without the original authors getting paid.

It's no secret that some LLMs have been trained on pirated books off libgen.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @04:01AM (#40703312)
Sure, but this is about FB and Instagram posts, posts that Meta owns anyway (due to the ToS).
Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @04:36AM (#40703518)
In Europe at least, laws override ToS and other agreements. You can sign all kinds if things with your clients/users, but laws can render parts of these agreements void.

In Poland we have a whole catalogue of voided agreement paragraphs, that gets updated regularly. It’s a safety mechanism against companies exploiting the legal system to push people into signing things that are considered unfavourable to them - https://uokik.gov.pl/niedozwolone-klauzule

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @04:43AM (#40703561)
I agree, but they weren't voided before this law. Meta wasn't doing anything illegal, since you had signed your rights over your content over to them. That's why I have a hard time accept any arguments of legality/copyright against this.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @06:54AM (#40704283)
Govt can void it after the agreement was signed. That is because companies keep working around existing laws, and pushing clients to sign unfavorable deals.

We’ve had banks for example that had bad terms (or trading platforms), and then government step in and make such terms void.

The reasoning is that a consumer/user should expect fair treatment when agreeing to ToS, without spending money on legal advice before clicking “I agree”.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @06:57AM (#40704310)
The conversation went like this: I asked why people are against AI training on their posts, some people said "because of copyright", I pointed out that Meta has a license from them to use the posts for anything, and now your post is "but the government can force them not to". Yes they can, but that wasn't the original question.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @07:23AM (#40704468)
Meta, nor anyone else doesn’t have a license to do anything they wish according to EU law - we don’t have wildcard copyright assignments - you need to specify the fields in which a person grants the copyright. Exatly for the reason that when a new tech is invented, the original author may decide to not grant the copyright for that field, or grant it to someone else.

That is, each legally binging copyright assignment needs an enumeration of fields, like radio, tv, streaming and so on, and when a new field appears the copyright belongs to the original author by default. And statements like “in all fields current and future” are not legally binding.

There is no such thing as “people signed copyright to Meta so Meta can use it in whatever fashion”.

You can read an exercept of Meta’s ToS here - these ToS do not allow them to train models based on the content people publish. https://www.quora.com/When-I-post-a-photo-on-Facebook-or-Ins...

Re:Meta won't train AI on Euro posts after all, as watchdogs... by ehnto on Monday June 17, 2024 @03:52AM (#40703254)
Privacy is another, if they are training with chat messages in a sense it "knows" the information in your chat. Data exfiltration would be really interesting in this space, a weird blend of tech and social engineering.

I don't really care about the money but as a society we should care a lot about such a blatant heist of personal effort. For many their words online are their works of the day, it's not at all fair for a company to take that and use it, and even more insidious that it will be used to replace the people who created those works.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @04:02AM (#40703320)
But this is about FB posts, not personal writings, for example. Meta owns your FB posts, don't they?
Re:Meta won't train AI on Euro posts after all, as watchdogs... by ehnto on Wednesday June 19, 2024 @05:29AM (#40726503)
I don't think it is fair to assume your average person understands that their personal rants are ending up in an LLM/Publically accesible AI, even if the T&Cs say they can.

That's what the fuss is about I guess, we are working out the ethics our societies will accept on a pioneering front of technology. I don't think Facebook gets away on the letter of the law this time, they will need to adjust to new regulations as we decide on what's ethical.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by prasoonds on Monday June 17, 2024 @04:03AM (#40703324)
The article very clearly mentions training on your public posts.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by hef19898 on Monday June 17, 2024 @03:51AM (#40703244)
A) Copyright laws exist for a reason, even big tech should respect them (considering their utter disregard for anything but money, well...) B) Especially Meta knows aleeady too much about everyone, now they can add AI trained on all of that (but what could possibly go wrong with that, right? C) I don't want AI to be trained on my stuff, to the point I don't share anything publicly anymore and only use encrypted clouds for back-up (and not even sure I'll do so any longer)

By the way, not everything is about money. Convincing people it is, is even a better trick than the devil convincing people he doesn't exist (that is assuming one believes in such things, and then considers the devil to be evil instead of the people ending up in hell, but I digress).

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @04:03AM (#40703322)
You don't have a copyright claim against Meta posts you made on FB and Instagram, though.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by hef19898 on Monday June 17, 2024 @04:43AM (#40703559)
That depends on tge terms of service and the jurisdiction you are based in. And whether or not you want to go to court over it or not. IANAL, but I am fairly certain terms of service do not trump national copyright laws. At max, you might forgoe some usage rights, copyright will stay with you (at least were I live, as you cannot sell copyright, only usage rights if I understood, and remember, the finer points of our laws aroubd that stuff correctly).

That bog tech gets away with, is not necessarily because it is legal, but rather that people don't care, law makers only start to care and even if people would care, it is nigh impossible to successfully litigate.

None of the above is written in stone, nor should it stop us from doing something about it.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by stavros on Monday June 17, 2024 @04:46AM (#40703575)
IANAL either, but AFAIK you grant meta an irrevocable, perpetual, worldwide license over your content. Meta isn't likely to have overlooked such a basic legal requirement. What they were doing wasn't illegal.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @04:30AM (#40703482)
Let’s say I tweet that I hate cats bc I’m allergic to them.

I would be ok with that tweet being findable through search, but if a model is trained on the information and starts spitting out that I’m a cat hater period (without context), then that’s a different thing.

Also, there is an expectation of privacy on platforms like FB or Insta, because the posts require at least login to be read. Because of that people may be posting things that they expect only their friends to read, and friends of friends - not to become a part of a world knowledge canon.

For me personally - on platforms like HN and Reddit, I have no issue with models getting trained on. Fb/insta I treat as an extended group chat and out of limits for cralwers and ml datasets.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by im3w1l on Monday June 17, 2024 @03:55AM (#40703270)
There are pros and cons to this. I wonder if the EU will come to regret chosing oblivion.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by johnnyanmac on Monday June 17, 2024 @04:00AM (#40703305)
Those are two oddly juxtaposed sentences.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by im3w1l on Monday June 17, 2024 @04:50AM (#40703605)
Well the EU believes in the right to be forgotten right? That's the philosophical underpinning of it's pushing for privacy. This in contrast to the US that more comes from an angle of people resisting an evil government. That's why all the EU privacy protections go out the window as soon as the police gets involved.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by callalex on Monday June 17, 2024 @03:19AM (#40703042)
What in the furry is going on here?
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Piskvorrr on Monday June 17, 2024 @03:22AM (#40703053)
Meta: "Privacy is dead, we own your data now." EU: "No you don't." Meta: "Uh, ok."
Re:Meta won't train AI on Euro posts after all, as watchdogs... by Moru on Monday June 17, 2024 @03:40AM (#40703171)
More like:

Meta: "Privacy is long dead, we always owned your data."

EU: "Hold up a moment, we need to check this before we can agree to this."

Meta: "Ok". Thinking: Eh, whatever. This will happen later anyway.

Re:Meta won't train AI on Euro posts after all, as watchdogs... by raverbashing on Monday June 17, 2024 @03:42AM (#40703196)
Yes

The AI will be very good at that it seems

Re:Meta won't train AI on Euro posts after all, as watchdogs... by genericacct on Monday June 17, 2024 @04:04AM (#40703334)
nice footgun, EU. This way AIs will speak european languages worse than they speak english
Re:Meta won't train AI on Euro posts after all, as watchdogs... by input_sh on Monday June 17, 2024 @04:16AM (#40703404)
So what you're saying is absolutely nothing would change?
Re:Meta won't train AI on Euro posts after all, as watchdogs... by mrighele on Monday June 17, 2024 @04:06AM (#40703350)
Given the level of the discourse on social media, I think this ban will actually rise the quality of the output.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by kolinko on Monday June 17, 2024 @04:38AM (#40703528)
It speaks EU languages well enough without this data already.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by cynicalsecurity on Monday June 17, 2024 @05:53AM (#40703939)
Surprise, English is a European language.
Re:Meta won't train AI on Euro posts after all, as watchdogs... by rickydroll on Monday June 17, 2024 @11:51AM (#40706916)
We are missing a big opportunity here! Training datasets should be considered a worldwide Commons and any profits made from those training sets should be fed back into national (or even worldwide) sovereign funds. If they go to profit from our work, we should get a share of those profits equivalent to or in excess of what shareholders make.