Post

We recently shared Bluesky’s stance on user data and AI training, which has not changed. Bluesky will not train generative AI on user data. bsky.app/profile/bsky...

Bluesky@bsky.app

A number of artists and creators have made their home on Bluesky, and we hear their concerns with other platforms training on their data. We do not use any of your content to train generative AI, and have no intention of doing so.

Bluesky is an open and public social network, much like websites on the Internet itself. Websites can specify whether they consent to outside companies crawling their data with a robots.txt file, and we’re investigating a similar practice here.

For example, this might look like a setting that allows Bluesky users to specify whether they consent to outside developers using their content in AI training datasets Bluesky won’t be able to enforce this consent outside of our systems. It will be up to outside developers to respect these settings

356 reposts6.2K likes

We’re having ongoing conversations with engineers & lawyers and we hope to have more updates to share on this shortly!

Option to completely disable AI? I am so sick of it being pushed on me in every single app! I don't need help to think of DMs to people! I'm a good writer and my brain still works. I feel like I'm being offered a wheelchair every time I get off the couch. I want AI to fuck off! Thank you! ❤️

Aren't there better things to spend engineer and legal hours on, than legitimizing a manufactured moral panic? Like actually encrypting DMs or fixing the regular netsplits between the PDSs?

Back when I was a software engineer at Blizzard, I kept hearing our players saying things like "why don't you spend time fixing the 3D models for this and that instead of what you're currently doing", and, like, I'm not an artist, I'm a software engineer...

Tbh you don’t need Ai here on Bluey. Users will post content and use the app like the old (and good) Twitter. We don’t need Ai writing our posts etc. Keep it all simple, and keep Ai away from the platform. Let the users create the content we will post, and Bluey will thrive.

While you're talking with lawyers, can you also discuss changing section (2 C) and (2 D i) of the terms of service to remove or change the phrase "future offerings"? There's concern that that's too broad, and it would let you be malicious with our data if control of the team changed.

If you can't offer assurances that our creative output won't be fed to the AI slop machine without our consent, creators will (and should) stop posting on this platform

The only ways they could stop this is requiring login to view posts or private accounts. The latter reduces visibility for artists entirely. The former guarantees nothing unless they also put a clause in the user agreement about not training AI (still hard to enforce though).

So if a billionaire come to pay millions to "outside developers " in order to create AI networks in bluesky, which will be trained by our accounts posts, is OK for you..... this is insane, men. Anyone of the billionaires will can do that to manipulate the safety of bluesky.

Hi there! As an artist, I'd like to give my input on this. And while I know that you're likely being flooded by responses, I hope that someone sees this and at least gets them thinking. First: I hope that you're not just working with lawyers and engineers. Involve artists in the conversation too.

Second: This may be what you're working with the engineers on, but there are plenty of ways to 'glaze' artwork to prevent AI scraping. Perhaps you could apply that to any images that are uploaded to Bluesky, except for the users who have opted in Third: PLEASE have it be opt-in and NOT opt-out.

Finally: It makes me very happy to see companies and organizations working to make AI training ethical. First SAG-AFTRA negotiating the regulations for voice actors being able to be paid for training, and now this. I can tell that y'all genuinely care and are trying to make it as best as possible.

would not recommend this, glaze itself is an AI model and it's inference uses up a lot of resources just for a single image. I could not imagine how more computing you'd need just for any time someone wants to upload an image

Glad to hear this! As I understand it (happy to be corrected, not an AT expert), you should have both technical and legal remedies at your disposal. You can collaborate across hosts to track IPs that are accessing inhumanly large amounts of data at inhuman speeds and block or rate limit them, no?

I have been trying to warm up to Blue Sky, but I find the character limitations and other limitations to be irritating at best. I just tried to post a couple of videos, and was informed that videos could not be longer than 60 seconds. Too many restrictions. #Bluesky

Please let me know how I might join such a conversation. Been making some #powershell #websocket experiments with the site, and want to ensure that any user can easily ban ai training / scraping.

Do what you have to grow, we live in 2024 and users need to understand we live in a different landscape than before. Google, Meta, and X all utilize these tools for their benefit. We can’t lag behind now.

The benefit on those platforms being? Driving people away? Making people stop interacting cause everything they post gets fed into a machine so someone else can shit out 14 copies of something that looks just different enough to sell while the actual artist doesn’t get any of the profit?

The benefit is short term financial gain. Companies pay for access to that data, which can be funnelled to bonus payments. Same as it ever was. The cost is public trust, which is kind of a big cost to pay, but it seems like we’re in a post-trust world now. We’re seeing slash and burn economics.

Eager to hear what your lawyers and engineers come up with. I'm a singer-songwriter. I've stopped using most social media platforms because of their gross lack of ethics when it comes to AI. I was truly hoping that this site would be different.

It's as different as it technically can be, that's the big thing. There literally is no way to prevent this scraping, especially on something with a public protocol like BlueSky has with AT. This truly isn't about BlueSky's ethics. This is entirely about the ethics of other companies.

They have the traction and demand to make the bold move of siding with artists. If they didn’t think something could be done they wouldn’t bother meeting with lawyers and engineers. I hope they think as big as the orgs and artists who have been making strides with litigation.

The getting lawyers and all makes me feel a tiny bit better but im still kinda eh… when it’s up to rich fucks to honor people’s consent if they opt out of something. The rich will do what’s best for their own interest.. and once somethings used to train an ai model.. it’s in the ai’s system :/

give me a shout please, I’d like to understand this better as we can share notes on how trusted publishers of news and entertainment are also approaching these signals. Thanks.

@bluesky Please hold the line against “outside developers” with a platform-wide prohibition against use, abuse, or sale of our content / data to anybody. Why are you even thinking about setting up an opt in/out consent structure? Just say no to “AI” bros. Thank you!

tl;dr everything you post on here becomes public domain as far as bluesky is concerned bsky.app/profile/pdj....

Paul Johnston@pdj.bsky.social

Fascinating conversations going on regarding data access, consent, and ability to use "bluesky user" data for research. The fascinating thing is some aspects of the technology are not clear to people and there is an expectation of... control? Waiting for "if you want free speech go to twitter"

How about embedding metadata about users AI scraping consent into all objects in the public firehose? This would at least give users legal recourse to challenge unlawful scraping.

Scrapers and AI companies have no ethics They won’t respect anything but profits Can you modify the AI protocol to ban access from them or add on some proprietary code to prevent it?

I said it before and I'll say it again: bsky.app/profile/anam...

Ana Matoso@anamatoso.bsky.social

An important distinction needs to be made here. It's great that @bsky.app says they won't use our data to train GenAI (not to say they won't ever, but that's for another day). But this doesn't mean that other companies won't. Don't be fooled: every post is for sure being scrapped by somebody else.

a better sounding version of what they're saying is "we wont use your data for AI but we may give you a way to allow others to do it if you choose to" They'd absolutely need to make it opt-out by default otherwise they'd piss off a LOT of people, but i'd be surprised if they didnt

One key point about this - this setting *may not be respected by scrapers*, because robots.txt and all the other things around it are, at best, gentlemen's agreements.

Perhaps make it clear that everything you publish on Bluesky is the equivalent of publishing it on the public web. If you want to mitigate specific content being trained without consent, host it behind an auth/captcha gate and link to it. Perhaps Bluesky could even add a checkbox for posts/profiles.

You'd think it being an open platform would already be pretty obvious, given, you know, that it's advertised as an open platform where all posts are public.

Having a profile option between public and private where you need to do some kind of manual operation would be good. Maybe an option to exclude your posts from the firehose?

This feels like such a betrayal, tbh Lure artists to your platform with a promise that you won't use their art to train AI, which you KNEW most artists would interpret as "your art is safe here", then turn around and serve up their contributions as a tempting treat for third-party AI developers

Which is your higher priority here? Keeping everything "public" and "open", or protecting the interests of the users who create all the content? You're going to have to make a choice.

If so, then users must have the option to control what they consent to, to what extent and it needs to be transparent. Not saying no, just saying be open because if not it will fail, this needs to be done with GREAT care

This seems to be where more can be done. A simple pre-filter on the firehose API removing posts from individuals who opt-out seems like a pretty simple solution.

Right now, the only thing we have is these "gentlemens agreements" like robots.txt. Like robots.txt, there's no enforcement mechanism, and in order for it to work, the scrapers would have to respect it. They largely haven't. They don't have control over what outside companies do.

You can… impede, but not stop, scraping. Scraping is just web requests, same as your app or browser make. Everything you can do to stop or slow them has a workaround that a company with will and capital can implement quickly.

The AT protocol is open to everybody, so malicious actors will still use the data. But if the default user setting is to allow the use of the data it might work since most users will likely use the default setting, allowing developers consent for accessing most of the data.

Could it be done more quickly by using existing Bluesky concepts such as lists and blocks? - i’ve blocked @external-scraping-allowed.bsky.app -> I don’t want my posts being used - I’ve added @specific-scraper-bot.scrapingsite.com to my “robots.allowed” list ➡️ I’ve given consent to that bot

Cause it's easy to say and extremely hard to implement. As any other website, the data can just be gathered from the same HTML and CSS used for the browser to show the page. The web works on open protocols that can be read by everyone

That's the thing about decentralization. The platfo- er I mean protocol(s) is not entirely controlled by one entity. If Bsky was able to enforce consent, then this is centralized, as it would have to ban other instances from connecting and storing content.

If you can't enforce the consent of your own users, then what, exactly, is the point in the consent in the first place? Leaving something up to outside actors (good or bad) is like telling a child to not have any of the fresh cookies left on the table but you won't punish them if they eat them all.

Like, at this point, I'm thoroughly confused as to how you plan on protecting your users who refuse to train any AI you do or don't endorse. One of the reasons I left Twitter in the first place is because of the shoving of AI down users' throats without consent in a recent update.

Are you trying to tell me, and us by extension, that we ultimately have no real and genuine say in the matter? I'm not trying to incite discourse or discord. I want some definitive answers rather than a "we're looking into it, chums!" kind of thing. What are your intentions? How'll you protect us?

Yah.. usually robot.txt files only really work when the website that has them is big enough to be an actual threat to be sued from if the scraper was not compliant with them…

I hope the AI is better than the one Instagram is using. By chance I wrote 2 comments on 2 posts that were identical… “so adorable”… the AI accused me of doing SPAM. Also, gave me warnings when I praised people I follow, accusing me of looking for “likes”.

I would want a consent button but would it really make any difference? Like I tell my kids if you don't want it to be public knowledge or property then don't post it on the internet or social media.

No offense but AI is a major reason why a huge portion of X are leaving for this platform. A positive here would be to downplay AI in all forms until its attainable in a more controlled and noninvasive format. Not tell you how to run this site, just know amongst content creators AI is a dirty word.

This is pretty much the fullest extent of their capability, as frustrating as that is. They can't force other companies to follow their rules.

Posting on blue sky is in the public domain. I do not see any problem with other platforms disseminating the information. But there does need to be a license agreement that third party cannot sell or alter the information. Somewhat like an open source agreement.

They think they should be able to steal everything for free to then make products that they will sell for profit. AI as it is now and capitalism doesn't work for humanity. It just allows rich people and their bootlickers to steal from everyone.

The only thing here is though, the blueskys AT protocol component, the firehose, is public. Block or not, the firehose still sprays posts and other data. So they can still collect and train. I wonder if there's a way to tamper with their stuff though. That would be fun.

I’m a bit confused. Is it or is it not the case that anyone off BSky can use the APIs to scrape any user “site”. If yes, then the statement that BSky cannot enforce users’ preferences other than with their own developers is significant. Then opting in or out makes no difference for any outsider

You are correct, it doesn't matter at all what this consent thing defaults to. Blue Sky is an open platform that they as a company have no centralized control over the content. The *whole point* is to prevent centralized control. There's no way to stop it

the promise of technology like Blue Sky is, if bad actors are going to train on this data anyways, we should make it available to the public, so that the good actors can outcompete the bad. It's our best shot

I don't know how feasible it is to implement, but there is a thing called glazing that's basically a filter for images that makes it impossible to train AI on without diminishing the quality of the picture. Would be cool if you could auto-glaze images when posting them.

This is extremely problematic. If a user publishes content that they don't have the right to, that content can end up being used as training data. Where abuse can happen, it will. This isn't a matter for fence-sitting. Either commit wholeheartedly, or don't, and users can react accordingly.