an incomplete list of fediverse instances scraped by meta to train AI
an incomplete list of fediverse instances scraped by meta to train AI
cyberpunk.lol
:pona_plush: #FediPact :pona_plush: (@FediPact@cyberpunk.lol)
26 comments
-
-
I see that shitposter.club is on the list. Good to know they're using only the highest-quality training material.
-
I tried to visit but their security certificate is expired. Are they still a legit site?
-
Moved to shitposter.world according to their site with the expired cert, but I haven't seen as much on fedi from the new domain as I used to from the old one.
-
-
-
Copy and pasting my own list from here
-
off-topic but wow, it's great to see so many lemmy instances up and running 🥰
it really looks like we're well on the way to hitting critical mass
-
-
This is only a loosely related thought, but are there any new foss licenses or anything that prohibit ai usage? I know it'll be ignored but it feels like explicitly disallowing things could be important in opening the door to successful legal challenges to ai scraping and theft...
-
Case law is still pretty young in this area, but it's looking like there's nothing actually against copyright about the training of AI on copyrighted content. It's not something that a license can restrict because the trainers can simply reject the license and carry on training under the basics of what the law allows them to do anyway.
Open source licenses only have power because they grant permissions that people normally wouldn't have and put conditions on those permissions. If you don't need those permissions then you don't have to be bound by those conditions.
-
Ahhh, that sucks ass :(
Thank you for expanding my understanding of the problem!
-
-
-
Those tasteless frauds!
-
Can we poison our posts by putting a nonsense “signature” at the end of each of them?
-
They're training on Hexbear
That's... amusing.
26 comments
I don't see why everyone's surprised about this. The Fediverse is running on ActivityPub, an open protocol whose purpose is to broadcast the content we post here to anyone who wants it. Of course it's being used to train AI, why wouldn't it?
Except iirc, they aren't scraping "properly" (read: efficiently at least, setting aside morality for the sake of discussing this component in isolation), and are causing traffic troubles. If only they took the time to install an actual instance themselves then nobody would care in the slightest (again, ignoring the morality part, for now).
TLDR: they are being dicks about it, bc offering everything we have for free is not enough for them.
But if they do it the “proper” way, they won’t be able to grab the data if instances defederate from them, right? And that’s what the majority of instances will do.
of all the scrapers we see, the requests identified as originating from Meta seem to be well behaved overall. they appear to (mostly) be respecting robots.txt where present and their request volume to Lemmy.World is only averaging slightly above 5 requests per minute over the last 2 weeks. they also don't spoof their user agents to pretend to be web browsers, or at least I have not seen credible accusations of this happening.
i mean, that’s exactly what they did with threads, and many instances defederated from it because they didn’t want to have their data scraped by meta
At this point, I appreciate that anyone can scrape it. Not just Reddit or Meta exclusively, but any start up that’s wants to compete. Sure, meta and the biggies have an easier time of it, but at least they don’t get it all only for themselves.
That doesnt necessarily mean that training AI on this data is legal. Especially when multiple of these instances had legal documents in place specifically forbidding this kind of use.
There are some lawsuits in motion about this and the early signs are that it is indeed legal. For example, in Kadrey et al v. Meta the judge issued a summary judgment that training an AI on books was "highly transformative" and fell under fair use, and similarly in Bartz, Graeber and Johnson v. Anthropic the judge ruled that training an AI on books was fair use. I always expected this would be the case since an AI model does not literally contain the training material it was trained on, it learns patterns from the training material but that's not the same as the literal expression of the training material. Since the training material isn't being copied there's nothing for copyright to restrict here.
it isn't about surprise silly goose its about moving the interaction from a suspected unknown to a known interaction in our collective threat models
silly goose