Can I remove my personal data from GenAI training datasets?
knowingmachines.org
knowingmachines.org
https://c4-search.apps.allenai.org
Love the disclaimer: "The dataset is released under the terms of ODC-BY. By using this, you are also bound by the Common Crawl Terms of Use in respect of the content contained in the dataset."
- Supply copywritten keywords and approved responses to try and retrain the model.
- Train the copywritten content again with inverse weighting. I imagine this one wouldn't really work well and still keep the model performance up...
I'm okay with that. It's sad that you're not. If you're willing to start a business on such shady foundations, there's a really good chance your business will continue to make shady decisions in the future. It's better to find and remove the cancer early
For an image/photography example, I may have my photo taken at an event and sign a release or agree to the event's T&Cs that say they can put publicity photos on their social media or website, but I don't intend that my photo will be used by an unrelated party to generate images of my face (this may not actually be restricted, yet, but I'd prefer it to be and I suspect few would consider it a reasonable outcome).
Copying books can be fair use if it is sufficiently transformative, copyright doesn't block all use of copyrighted material. A major case on this was around Google building a search index for books:
https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....
Open AI is currently attempting to argue that training a model is more transformative than building a search index, whether they win that argument or not is still to be seen.
Until this is decided, likely by a long series of trials in high courts, or passage of law through the appropriate legislative bodies (and then likely after many challenges to it) will GDPR and CCPA be able to apply to these language models.
So no this is not currently a violation of any law.
An online shop in Germany once removed Cuban cigars because of US trade sanctions.
ClearView just declared to ignore the right to be forgotten, even for EU citizens. It will be interesting to se how this plays out. I guess it will take a few years…
No amount of legislation can outweigh the technical reality.
That's the reason you mostly see the pretty dumb guys getting caught. As long as you are smarter and more tech-savvy than 80% of the criminals, you are pretty much out of reach for the feds.
YouTube have decades of attempt at determining if uploaded content violates copyright, normally through fingerprinting content submitted by right holders. YouTube still fails at catching all copyright infringement. That's why they are generally protected from prosecution while they demonstrate reasonable attempt that preventing their services from being misused.
Imagine ALL IMAGES IN THE WORLD being submitted for fingerprinting. How about malicious or erroneous submitions that taint datasets?
Also: right to be forgotten is a horrible misleading name. At most you have right to request data to be removed from datasets. And only if you have some legal basis that demonstrate that your data is in that dataset. It is not a right to be "forgotten". Imagine you yourself being told to "forget an image, sound, etc", how is this enforced? You may even ask for proof that you know the content you are being told to forget. How do you do this?
It is illegal to possess, but despite the most aggressive enforcement on the planet it is still out there.
In case of personal data in this article, they have already identify which legal entity have it.
From the article I'm extremely skeptical people have actually identified anything, since LAION does not store content. LAION is a list of links to content hosted by others, publicly.
Maybe realistically any uploaded content can be found somewhere on the internet forever (I doubt that, tbh), but this doesn't mean we should throw our hands up and say "it's impossible to regulate".
We already have precedent for this in the form of copyright/DMCA takedown requests. The MPA/RIAA has a legal avenue to protect their content, why should it be any different for individuals?
Do you expect your online bank to secure your data? Your email provider? Your healthcare provider? So you do expect some privacy. I just expect some more. I believe that companies should not own my data just because they provide me with services off it. From that, consequences follow.
I have a better proposition: let's slap companies that don't respect privacy with fines until they too learn a lesson.
That was the expectation until Zuckerberg came along. People were putting tons of stuff on BBSs, intranets, and other stuff 20+ years beforehand because people had the idea of "consent" back then existed: using someones information for a purpose they didn't intend was (and still is) wrong.
Lots of art, science and technology can be considered as "using someone's information for a purpose they didn't intend." It is extremely normal for people to find new uses for things; information is not exempt from this.
Whoever treats others' information can only do so with a clear purpose and for a defined time period according to current EU laws.
If, however, you were referring to information that you published on the Internet for everyone's benefit then you would still need to consider intellectual property rights. In the open source software world we have the licenses that deal with this, and then there are copyright laws protecting content providers (not making a case here whether they are good or not).
I guess what makes a difference is if there is some business involved either in the production or in the consumption side of the equation and if we accept that "machine learning" is the same as "human learning".
EDIT: separated paragraphs, typo
There are also cases where loading copyrighted content is unavoidable in order to even view the license terms. For example, a website could deny the right to copy its content before you even see the licensing details. Temporary copies enabled through normal usage like this should be permissible.
Yet you would expect that people won't record your private conversations, and even less play it back to third parties.
Exchanges, of whatever information, whether it is personal stories or administrative data occurs in some kind of trust on what kind use you consent to. Respecting this consent is a social obligation on the receiver end.
How? It seems obviously correct. Suppose we trade nudes. You now sell my nudes online without my consent. You bet I have the right to sue you.
But RTFA, because what it is discussing is not public facebook/twitter posts. I'm not sure why the discussion keeps moving there as if that's the only case. Besides, a public facebook post is still only visible to those on facebook. It is as public as a private country club, even if it's easy to get access. What about what you write in copilot? What about your code on github? In a private repo? The article mentions this is already protected under current law as personal information.
The fundamental problem here is about consent. What would a reasonable person expect? 10 years ago certainly no one expected this. Mass gathering of data was only for the biggest of tech and governments. Now it's in the hands of anyone that wants to train an LLM. Things changed very fast and I think a lot of people are forgetting that.
I just don't think that's the right angle. I think if you're in the public sphere, the way that you anticipated your creations to be used just shouldn't matter at all. If you've given something to the public, the public should not be limited to whatever you thought they'd use it for.
And the conversation keeps coming back to that because people keep making that argument. I specifically agree that private Facebook posts are not covered under this argument! However, I also think that consent as a basis for analysis just doesn't apply to public posts.
As for consent, you may disagree, but that's how our entire legal framework works, and I for one am happy about this. Consent is the foundation of any democracy. A world without consent isn't figuratively tyranny, it is literally tyranny. Agree or not, the law does not give you the right to use any information (words spoken, pictures taken, bits, or whatever) that is publicly available/viewable in whatever way you want. There are rules. It is why Taylor Swift can't even post a picture of herself on Instagram and why you can't follow someone around all day taking pictures of them (aka. stalking). There are limitations because we live in a society and I for one think this is a good thing.
Yes, you are right when posting something public that people shouldn't limits on use for only what the poster explicitly intended. But that freedom is neither unlimited. That's the point I'm making. The line of public and private is getting blurrier by the day. These laws are always made by that of a reasonable expectation. Certainly no one that posted on Facebook 10 years ago had a reasonable expectation that their post would be used to train an LLM.
If we have to lose privacy and control in order to keep our participatory culture, I'm sorry, I'll take culture. If it's a multidimensional topic, I'm for tightening privacy in private spaces and loosening privacy in public spaces. If there's only one axis, I'll push the lever towards loose.
You'd be surprised because these industries have figured these problems out. What is allow without license, what needs royalties, what needs permission, and what can't be done. This was a huge conversation in the 90's and you can find many discussions about hip hop (around sampling) and copyright law.
Your culture is not in danger, even if you don't know its history. I'd suggest learning it though, because that's the best way to ensure your culture stays out of danger (a culture I am also part of fwiw).
Yeah, that's to the best of my knowledge true but IANAL. I'm with you about the fanfic scenario too. I'm also very open about sampling. But I think these issues are different than the data that we're talking about in this thread. If data can be recovered (an in some ways it is, others it isn't) then that's not really derivative. Derivative also needs some distance and not be too close. And importantly, these data are being used to create a product that is being sold. Where the processing of the data is the thing of value.
My point is that the environment has changed and there's a lot of gray area here. Turning this into a binary distinction is unhelpful. There's new nuances here and there were new nuances when sampling became popular in hip hop. We need to have open and honest discussions about these things, and I think a lot of discussions we have or observe are rather dismissive of these nuances (coming from both sides of the debate fwiw). I'm obviously very open to using data but we must also be aware of our data privacy, how it can be used, our social contracts, and what a reasonable level of a priori consent is. If we overly simplify these conversations then they aren't actually discussions. My points are to this, that there is gray and that there are very clear cases where you do not have unlimited access and usage to works that are publicly available. Private ownership is the root of capitalism afterall, and so it should be rather unsurprising that we have many laws and social contracts over ownership and the extent of what one may do with things they did not create. There certainly is a lot of anger and frustration in these conversations and I don't expect an artist making their living off of their art to understand all these nuances nor am I surprised that they are upset and possibly afraid. This is new territory and pretending it isn't is just as obtuse as calling generators fuzzy copy-paste machines.
But I want to be clear that we can have both goals. We can protect data rights, privacy, fairness AND have this sampling and creativity culture, for lack of better words. We just need to be careful, nuanced, and thoughtful to determine how to do this though. We won't be perfect and won't make everyone happy, but we can maximize social agreement conditioned under fairness and privacy. I just want to ensure that we are not approaching this conversation as that there are clear answers in what can be done with data and what can't be. Hell, we don't even have that answer for music, sampling, or fan fiction. We have answers as to what laws say, but even as you point out, that's ambiguous in many cases without even considering that the environment is not only changing, but changing rapidly. I think we all understand that there is a difference between using the Akira slide compared to the "Ice Ice Baby v Under Pressure" scenario. No one has the answers, and that's why we need to talk. And unfortunately "edge cases" are the norm in topics like these.
Note: I am an ML researcher. I use publicly available data to train models that have images of people, their art, their animals, their property, and such that I'm sure many do not know exists in these datasets. Similarly I do not even know all the data within some of these datasets. But I can still recognize that there is a gray area that exists here and personally I see it as my ethical duty to ensure we have these discussions in an open and honest way to determine the limits of what I should and shouldn't be able to do. It isn't up to me, it is up to our society to create a social contract.
Sure, it's a continuum. I'd say it's more that, viewed as a continuum, my opinion about the ends of the spectrum is binary, in that I want one end to go up and the other to go down. That doesn't mean I want to split it into two cases so much as that I want to consider it among other things on an axis between two points, "private" and "public", and my opinion is a function of placement between those two points.
Artists are frustrated about fanfic too! They feel genuine ownership of these characters. I've seen people have borderline breakdowns at fanfics. The idea that people are doing things that they didn't intend to "their" characters can be genuinely traumatic to a writer. If you make that a question about consent then fanfic becomes comparable to abuse, and I just don't agree that those should be the terms. Any work placed before the public is inherently participatory. Copyright splits the world into "creators" and "consumers" and I think that's problematic, because there's no such thing as passive consumption of a work. The listener or reader recreates the experience of the work in their head, based on their own assumptions, preconceptions and personality, and if they have an experience or viewpoint that diverges from that of the author, they must be able to share that in turn. So I have a genuine values disagreement with some of what we currently call "rightsholders" in the matter.
What's that got to do with deep learning? Very little: I just see where it leads if you phrase the topic about consent, and it's not a world where unwashed readers can get their dirty mitts on their characters.
(My personal favorite is... closed-source Minecraft (Java) mods. "Do not decompile against my wishes!", the description said. Minecraft mods, meanwhile, are only possible in the first place by decompiling Minecraft's source code against Mojang/Microsoft's wishes. You were this close to learning something!)
This isn't my recollection of a pre-facebook internet
The early days of the Internet were full of "information should be free" hackers, many of whom roam this very forum in their 40s, 50s, and 60s.
Well, information is free now. And this is what it looks like - consumed by LLMs. We got what we asked for.
I am still firmly in the Information wants to be free camp, and I agree with you that LLMs scanning it is a logical continuation of it all.
While that is somewhat true, there are multiple angles to it. The big one is expectations. It used to be that if you shared something on an obscure forum that had like 100 readers, most of those people expected that only 100 would see their posting. And it was true, so it reinforced the expectation for nontechnical people, which is basically everyone within a rounding error.
Then everything gets indexes and now fed to AI and suddenly billions of people have easy access to what they felt was very limited in distribution.
You could file this under "People don't threat model correctly", but also (more importantly) under the fact that technology should consider human nature before breaking expectations.
That isn't necessarily true. Someone else might have put the information on the Internet. That could be someone else uploading a photo of you, or records of your home purchase on a city's public records, or an obituary or wedding announcement that listed you as family, etc.etc.
Now, you’re probably thinking about this from a reasoning or technical perspective in which case it’ll appear to be a ridiculous concern… because it is a ridiculous concern. That’s not how our legal department sees it though. They see it as risk mitigation, and they actually take it rather serious.
One may not have been aware of the future existence of AI when writing on a public platform.
Or one simply may want his blog/thoughts to be intended for human consumption and not to train the latest commercial product which will reap from his content and not share any profit.
There is no expectation of privacy in public. That is pretty much the definition. It’s not victim blaming.
This is claimed a lot but absolutely not true. Stop repeating this very obviously false statement.
Do you believe you have the right not to have a satellite track or drone hover over you in public and share that feed, associated with your identity, with the whole world? I sure do. You should too.
I don't think enough people are internalizing how radically the world has shifted in the last 10-20 years. 10 years ago generative image models couldn't generate a black and white 32x32 image of a face. Now I can trivially transform a photo of my friend into a dog and make him bark. For most people this technology came out of nowhere because of the lag. The same goes true for other things like the internet, cellphones, social media, etc. Is it really that much of a wonder why there's so much craziness going on right now when so much has changed so fast?
And fwiw, speaking of satellites, there are now multiple companies (not governments. Companies) who's goal is to get a daily snapshot of the entire surface of the earth. They already have cameras where you can see clearly the color of cars from space. There are some where you can see people (albeit hard to see, but you only need to identify a dot to track it). We now should expect that resolution to increase, not only in number of pixels, but frequency. This technology can do a lot of good, just like ML, but at the same time it can be used for a log of destruction and oppression, just like ML. The question is if these technologies are going to go the way of atomics or the way of the internet. But even simply saying this, I, a dumb human who's barely more advanced than our chimp cousins and barely capable of communicating, have a hard time internalizing what this future world will be like, despite being only 10 or so years away.
If your expectation of privacy changes in any way with any of these options, then yes, you have an expectation of privacy in public. It is a spectrum, not a binary decision. But if you think it is a binary decision you're welcome to try to record people going to the bathroom in public restrooms and see what happens. I will not be covering your lawyer fees.
such a ridiculous example if you are trying to convince me that you care about these girls' safety as opposed to some abstract egg-headed right.
But if Disney puts information on the internet, and you misuse it, lawyers will chase you to the ends of the earth.
We both have same right and copyright, but somehow it's one rule for them and another rule for me.
Even for publicly shared data, there is legislation in many parts of the world (notably the entire EU) that enforces an individual's right to have their data deleted.
To that end, you should consider any file on your computer to be public as well if it's connected to the internet, as there are countless ways that data could be exfiltrated.
There's no difference between an AI being trained on public data, or a human being trained on public data. Likewise there should be no expectation to "unsee" something someone willingly posted in public.
Services are advertised as private. So why wouldn't people be acting like they are private. Oh, you're not tech literate and so don't know something "obvious?" Guess it's your fault for listening to your financial advisor and losing all your money too. We live in a specialized world. You don't want to go down this route and it is unreasonable because we can't have infinite knowledge and infinite time to gain that. There's absolutely no way you know in high detail: tech, medicine, economics, law, science, politics, or any other expertise. You don't even know one of these, but rather a small section of one.
Yes, they may be "public" but you can have an expectation of privacy in public. Also your lack of expectation in public is not infinite. The lack of privacy in public is an astonishingly new phenomena. Realistically in the last 10-15 years as we've seen cellphones become prolific. Things were way different just in the early 2000s, where there weren't nearly as many cameras either by CCT or just in peoples' pockets. Let alone microphones or high resolution images.
Most importantly, the environment has changed on the internet too. 10 years ago the only people scraping up the entire internet was big tech and govs. Now it's every startup and group of people that can. These are very different situations. 2015 image generation was still a joke. Goodfellow's GAN paper was posted in late 2014. That's not even 10 years ago that we're struggling to generate black and white faces at 32x32 pixels.
Things are VERY different and if you think you could have predicted this I'm going to call bullshit. Only Captain Hindsight could do that. This kind of thinking is simply just victim blaming and lording your particular expertise over others while ignoring your own ignorance because you let post hoc thinking rewrite your past.
The only times they care are when it can cost them money. SD didn't care about visual artists but could not do the same for generative music since the rights are managed by deep pockets.
And it’s not like ads will disappear; if anything, we’ll get AI ads.
Government can make the incentive — and has. Legislation can put the teeth in societal goods like this even if financial incentives don’t.
California has done exactly this with their right to be deleted law.
https://www.foley.com/en/insights/publications/2023/10/calif....
GDPR and many other laws are still applicable to what GenAI is.