It doesn't matter that Bluesky/ATProto is an open protocol. That might make scraping easier but the scrapers don't care. I promise they're scraping Twitter too, and the Fediverse/Mastodon. Anything that is public. The only way to solve third party scraping is by making it illegal and enforcing it. - ThreadSky

lina.yt • 168 days ago

It doesn't matter that Bluesky/ATProto is an open protocol. That might make scraping easier but the scrapers don't care. I promise they're scraping Twitter too, and the Fediverse/Mastodon. Anything that is public.

The only way to solve third party scraping is by making it illegal and enforcing it.

Comments

coreh.bsky.social•164 days ago

Making scraping illegal would be really bad though, there are many legitimate usages for it, including archival/preservation (Wayback Machine etc), adversarial interoperability when no API is available for third party clients, mirroring of content to other platforms, and research applications

cathures.com•168 days ago

So much this. Sue the pants off of ANYONE using scraped datasets, enforce transparency on the part of any company using datasets period. Legislate and enforce it on the most powerful companies in the world. Speak the only language they understand - money.

exordelabs.com•168 days ago

Collecting Public Data will never be wrong.

Without it, how could we monitor information trends, misinformation, opinion shifts & clusters?
Information thrives when free, not silo'ed.

Not all data collection is for AI 👍 We do it for OSINT, Analytics. For good.
Check what we do at https://exordelabs.com

voidgroup.net•168 days ago

They scrape anything that is public, or even things that are only accessible to a few people. Although not all scraping is bad, the problem lies in what is done with the collected data (although doing it to the point it ruins platform performance is bad too).

voidgroup.net•168 days ago

SteamDB for example, is also done with tons of scraping of data, but it's done with community interests in mind for actual useful data.
Thing is, this data really shouldn't be allowed to be used in training data, although this is definitely not the platform's fault.

lina.yt•168 days ago

The only thing a platform can do for you is promise they won't do AI training themselves, like Bluesky has, and allow people to make it explicit they don't allow scraping (which is useless now but might be legally important if the law catches up), which they're also working on.

fugi.tech•168 days ago

I wonder how terrible it would be to automatically apply the various AI poisoning techniques to every uploaded image? I know some change the appearance so might not be suitable, but haven't read enough to know of other downsides.

href.random.sh•168 days ago

The methods require a lot of resources and given the benefits it's probably not worth it.

I also don't really like Nightshade style poisoning as it might hinder moderation and accessibility tools (CLIP isn't useful only for generating images but understanding them)

lina.yt•168 days ago

That would be an interesting feature! I'm not sure how practical that would be though, AIUI Glaze is itself one of those ML projects stuck on CUDA/Nvidia only... might be pretty expensive.

I wish more people used standards like OpenCL and Vulkan...

gbl08ma.com•168 days ago

IMO it would largely be a waste of computational resources as it's a cat-and-mouse game and it's safe to assume that those training on scraped data (especially the big corporations) will always have more resources to throw at it, than individual creators.

gbl08ma.com•168 days ago

Even if you had poison that was fully effective against the current training methods, the poisoning methods themselves can be used to train systems that aren't fooled by them. One can easily generate a vast training set by running images through the poisoner. The very same transformer models that

gbl08ma.com•168 days ago

are so good at generating images, are especially good at this. So for the "unscrupulous AI bros" it's just more of the same work they've been doing.

The "poisoned" images posted today might be unsuitable for training upon them now, but that won't be the case forever.

gbl08ma.com•168 days ago

If this concerns you, regulation is the only real answer, and even that will be ineffective just as long as the internet is global and there is at least one country where such regulations don't exist/aren't enforced.
It's exactly like with piracy.

houmus.org•168 days ago

Couldn't you actively fight them with fire by injecting garbage?

lina.yt•168 days ago

There's no way to inject garbage that wouldn't affect legitimate users of the network...

cardlemur.bsky.social•168 days ago

It's very unfortunate that this has become the norm but what was stopping people from doing this before was just being a good digital citizen and not anything actually enforceable and trying to enforce anything will be disastrous

cardlemur.bsky.social•168 days ago

Like overreacting to something like this becoming more widely available is how we got those Napster and LimeWire lawsuits back in the day so people should be very careful about what they advocate for in response

llopis.dev•168 days ago

Yep. Not only that, even for *private* data they at the very least kick up a fuzz (unless they don't care and use it anyway): Meta retaliated against the EU by not allowing the new Llama models just cause they were told to only train on data in an opt-in basis (they weren't allowed an opt-out model)

jerieljan.com•168 days ago

And it gets trickier considering that unauthorized scraping needs to be illegal everywhere too, and not just in select countries.

It's a tall order even for enforcement, imo.

Comments

Posting Rules

Reply