287
Meta’s new AI image generator was trained on 1.1 billion Instagram and Facebook photos
(arstechnica.com)
Surprised no one
this post was submitted on 07 Dec 2023
287 points (97.4% liked)
Technology
59135 readers
6622 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
So I assume they added any necessary stuff to the TOS to allow this.
My question is if there's any legal mechanism to prevent this on other platforms? Pixelfed for example.
Companies will likely federate and pull images regardless, but can we go after them when they're caught? Nothing prevents them from taking the images for internal R&D, but at least we can stop them from selling products with that training data
So I assume they added any necessary stuff to the TOS to allow this.
Never read it but I assume it already was. Pretty much every platform has a clause that says something along the lines of "we own all the content you submit to our service".
Actually it's usually more "you own the content but by posting it grant is an irrevocable right for us and our partners to use it"
Basically allows them use without the responsibility for ownership of inappropriate content
Exactly. Instagram doesn’t claim ownership to any of your content, but Instagram's terms of use state that the user grants Instagram a non-exclusive, fully paid, and royalty-free, transferable, sub-licensable, worldwide license to use their content. Additionally, they can make money off your content without ever paying you a cut. Honestly, it’s pretty boiler plate at this point. No one should expect anything else from corporations.
My question is if there’s any legal mechanism to prevent this on other platforms? Pixelfed for example.
Good question!
I’ve been saying for a while that the fediverse is blind to this issue as everything here is completely scrapable through either the public web or by running federated servers. On top of that, being culturally inclined toward more “serious” conversation and providing content warnings and alt-text for images, we’re probably generating relatively valuable training data.
And yet everything is public as though it’s still 2012.
There are alternatives. BlueSky for instance is basically private to members only. They recently announced that content would be made public to the web and a number of users were upset.
Group chats and Discord servers are probably similar, and from what I can tell “new” popular places for social activity online.
A major issue the fediverse has, IMO, is that it’s kinda stuck trying to fight Twitter and Facebook circa 2012, when that battle was lost and we’re on to new battle fronts now.
Yea that's something that's been on my mind as well
There are benefits from that openness and verifiability in public spaces (ex. Lemmy communities), since now it's easier to determine if there's vote manipulation or astroturfing. But I think the fediverse needs a lot of work around privacy, and also education about what is/isn't private on these platforms.
There should also be more of a focus on setting up a legal requirement on what can be done with the information, but I'm not sure if that's a thing just yet. We developed GPLv3 to make sure FOSS products can't be incorporated for profit, but I'm not sure how it would work for data.
ex. It should be easy to save, record, and share posts on the fediverse, such as with embeds/screenshots/news stories
But also we want to prevent abuse, misuse, and AI training
Bluesky being only accessible by members doesn't completely prevent the content from being scraped by bots, though. Bots can be given user access in Bluesky too, and bots can read posts, create own posts and scrape posts and user profiles.
afaik activitypub/fediverse doesn’t have to be fully open… there’s private messages and followers only profiles on mastodon… sure, any server admins of your followed would be able to see anything you post (and thus in this case for threads for example, if you accept any follower from threads then meta can see your stuff) but this also doesn’t grant them a license to use the content
also, bluesky will eventually be the same: it only doesn’t have those issues now because they haven’t opened up their software… it’ll have federation in the future, which means it has to be somewhat programmatically open to others
I think in order to fight against these composite using our data for AI training we souks have to do something like watermarking our images explicitly stating that they are not for AI training. Or we create some type of counter measure that messes up the training.
You're never going to get rights over the training data your pictures that are freely available for anything to scan creates. By being on the internet your pictures basically have the right to be viewed by anyone or anything even an AI. You have never gotten to control who looks at your content after you post it.
You're trying to make the same argument the "don't copy my nft" bros tried to make.
Imagine going into court and saying you should get paid for all the stuff u gave away for free on the Internet willingly.
Well there's a difference between "don't look at my work without paying me, even if it's posted publicly" and "don't sell my work without paying me, even if it's posted publicly"
Like I said, there's nothing we can do about companies using all the data they can get their hands on for private R&D. It IS possible to protect against the second case, where companies can't sell an LLM product with copyrighted training data.
My question was about how that second case could be extended to stuff posted on the Fediverse, such as if an instance had a blanket "all rights belong to the user posting the content".
These laws exist, if companies can use them then so can we
That ought to satisfy all those who wanted "consent" for training data.
I wonder how they worked around user violations of copyright... Imagine all the content uploaded to Instagram/Facebook that the poster didn't create but simply uploaded their download/screenshot.
That shouldn't be an issue. If you look at an unauthorized image copy, you're not usually on the hook (unless you are intentionally pirating). It's unlikely that they needed to get explicit "consent" (ie license the images) in the first place.
Yeah but is it the same thing for a human to view data and an AI model to be trained on it? Not in my opinion as an AI doesn't understand the concept of intellectual property and just spits out the most likely next word whereas a person can recognize when they are copying something.
I understand. The idea would be to hold AI makers liable for contributory infringement, reminiscent of the Betamax case.
I don't think that would work in court. The argument is much weaker here than in the Betamax case, and even then it didn't convince. But yes, it's prudent to get the explicit permission, just in case of a case.
Doesn't really seem the similar to me at all. One is a thing that's actively making new content. Another is a machine with the purpose of time-shifting broadcasted content that's already been paid for.
It's reminiscent insofar as personal AI models on individual machines would go, but completely different as for corporate and monetizable usage.
Like if somebody sold you an AI box that you had to train yourself that would be reminiscent of the betamax case.
Yes, if it's new content, it's obviously no copy; so no copyvio (unless derivative, like fan fiction, etc.). I was thinking of memorized training data being regurgitated.
Yeah I just think that ingesting a bucnh of novels and rearranging their contents into a new piece of work (for example) is still copyright infringement. It doesn't need to be the Lord of the Rings or Star Wars word for word to get copyright stricken. Similar to how in the music sphere it doesn't need to be the same exact melody.
Edit: Glad you down voted instead of responding. Really shows the strength of your argument...
I didn't downvote you. (Just gave you an upvote, though.) You're reasonable and polite, so a downvote would be very inappropriate. Sorry for that.
Music is having ongoing problems with copyright litigation, like Ed Sheeran most recently. From what I have read, it's blamed on juries without the necessary musical background. As far as I know, higher courts usually strike down these cases, as with Sheeran. Hip hop was neutered, in a blow to (African-)American culture. While it was obviously wrong, not to find for fair use in that case, samples are copies.
It's not so bad outside of music. You can write books on "how to write a bestseller", or "how to draw comics" without needing permission. Of course, you would study many novels and images to get material. The purpose of books is that we learn from them. That we go on to use this to make our own thing is intended (in the US).
What you're proposing there would be a great change to copyright law and probably disastrous. Even if one could limit the immediate effect to new technologies, it would severely limit authors in adopting these technologies.
I'm arguing that AI and a human are doing different things when they 'learn'. A human learns. At the end of the day AI isn't doing anything near human intelligenc and therefore isn't critically thinking and applying that information to create new ideas, instead directly copying it based on what it thinks is most likely to come next.
Therefore a human is actually creating new material whereas AI can only rehash old material. It's the same problem of training AI on AI generated content. It makes any faults worse and worse over time because nothing 'new' is created.
At least with current AI tech
Well, that is a philosophical or religious argument. It's somewhat reminiscent of the claim that evolution can't add information. That can't be the basis for law.
In any case, it doesn't matter to copyright law as is, that you see it that way. The AI is the equivalent to that book on how to write bestsellers in my earlier reply. People extract information from copyrighted works to create new works, without needing permission. A closer example are programmers, who look into copyrighted references while they create.
Except that it's objectively different.
A closer example would be a programmer copying somebody else's code line for line but switching the order of some things around and calling it their own creation.
AI cannot think nor add to work. It cannot extract information in order to answer a question. It is spitting out an exact copy of what was ingested because that is the scenario the system decided was "correct".
If AI could parse information and actually create new intellectual property like a human, I'd find it reasonable, but as it stands it's just spitting out previous work.
Can we get back to this? I am confused why you believe that AIs like ChatGPT spit out "exact copies". That they spit out memorized training data is unusual in normal operation. Is there some misunderstanding here?
I don't think we're really talking to each other, but more past each other so I took a break.
To answer the question, it was an analogy and the ransomware part was to show the non-intelligence and creationary lack of AI more than be applied to the programming analogy. Sorry if that was confusing.
It was an ars technica (iirc) article I read in which the author made a working ransomware with GPT-4 by having it initially create a program to encrypt a file, then had it encrypt directories instead, then added flags and debugged it all of which he claims can be done by pretty much anyone malicious with access. Nowhere along the way did chat-gpt realize what it was doing though. A human would have.
Also ime at least I got completely copy and pasted paragphs from gpt 3.5 a few times dunno how much 4 has improved upon that.
I think my disagreement with you about AI copyright infringement is that you think that AI can create new things whereas I don't think that. I think the way I do because it can only ever rehash its training data. Our current AI systems can't actually create new thoughts. For example, with your 'how to write a book' author analogy, those people haven't just read people's advice and are now putting it on paper. Those people have also read tons and tons of novels. Taken classes on English and created and defended original ideas as part of that. If you trained an AI on English classes and novels it would have no idea how to write a "how to write a novel" type book while a person would. You have to have it copy something in order for it to perform, it's just the way that it works.
Furthermore it really wouldn't take a huge change to copyright law, just clear differences between the rules that apply to sentient vs non-sentient sources.
Well, that's simply not true.
You can say that without explaining but you just look like an idiot.
It's the same reason gpt4 will write you working ransomware without ever noticing that it's writing ranosomware. The AI doesn't understand what's going on. It just does what it does because of a virtual cookie based on a calculated score.
Ok, where did GPT-4 copy the ransomware code? You can't reshuffle lines of code much before the program breaks. Should be easy to find.
Is there an example of AI generated images that aren't hyper realistic or have perfect bokeh? I'm taking about an out of focus shot or the subject looks like a regular slob like you and I?
It's mostly bias in the training data. Most people aren't posting mediocre images of themselves online so models rarely see that. Most are also finetuned to specifically avoid outputting that kind of stuff because people don't want it.
Out of focus is easy for most base models but getting an average looking person is harder.
I would usually try to add things to the prompt you'd expect to find in a more casual scenario, like "smartphone" with half weight or something, or "video", or maybe like "Facebook". Just meta information you think attaches to more casual photos. Maybe even add "photo".
you can do that with stable diffusion and loras, yes
Loras are amazing. You can do anything or create anyone.
Create me.
On the focus part, I've seen some impressive results from people who input a specific camera, lens, and focal distance.
I assumed this was because it's making an average. Human attraction is highly sensitive to symmetry so this creates that symmerty by the way it works.
The models are deliberately engineered to create "good" images, just like cameras get autofocus, anti-shake and stuff. There are many tools that will auto-prettify people, not so many for the reverse.
There are enough imperfect images around for the model to know what that looks like.
All this AI photo generation is leading me to think that all imagery is going to be essentially meaningless. Is it real? Is it fake? Did a bot make it, or a human? As this tech continues to grow, i will be studying every image i come across while i ask myself those questions subconsciously.
I mean on one hand, you can "see" almost anything you can type out descriptively enough. Pretty neat! But now virtually anything can be "seen" which includes things that shouldn't be this easy to show. I'm thinking propaganda, deepfakes, blatantly making up fake news with imagery and video to back it all up. I guess we were always headed in this direction one way or another.
you can 'see' almost anything you can type out descriptively enough
A significant fraction of the population can't! https://en.m.wikipedia.org/wiki/Aphantasia
Even those people who have difficulties with imagining something visually can use AI image generators somehow. As long as they can write and understand what a sentence means, they can use any sentence as a prompt to get a calculated image. You don't need any artistic talent or phantasy to get started with creating basic artificial images. That's exactly why artists around the world feel their skills are now being devalued by AI generators.
The upside of generative AI better be no less than an end to toil because in the short run, they’re going to ruin the internet and prompt several genocides.
Considering most of the people on Instagram don’t even look like the photos of “themselves” that they post on Instagram, this might be an uncanny valley image generator.
Thing is they have the original data aswell to train on, so the machine knows what the average of someone looks like and the average to which they change it so they could in theory have a good grasp of the uncanny valley or at least nó the gap the scale back to am original look.
Deviantart has already done this since the AI image hype train first started. Every picture by default is selected as material for AI training; pictures have to be manually deselected by the user to be excluded. And of course it's a nightmare for those with tons of art submissions.
Facebook/Instagram may end up having to something like that in the future but I doubt it until someone higher up does something about it.
I wonder if they'd release the weights and training/inferencing code. They did it for LLaMA.
There's been a lot of open source alternatives to Stable Diffusion lately and it's great.
This is the best summary I could come up with:
Previously, Meta's version of this technology—using the same data—was only available in messaging and social networking apps such as Instagram.
Images include a small "Imagined with AI" watermark logo in the lower left-hand corner.
We put Meta's new AI image generator through a battery of low-stakes informal tests using our "Barbarian with a CRT" and "Cat with a beer" image synthesis protocol and found aesthetically novel results, as you can see above.
(As an aside, when generating images of people with Emu, we noticed many looked like typical Instagram fashion posts.)
The generator appears to filter out most violence, curse words, sexual topics, and the names of celebrities and historical figures (no Abraham Lincoln, sadly), but it allows commercial characters like Elmo (yes, even "with a knife") and Mickey Mouse (though not with a machine gun).
It doesn't seem to do text rendering well at all, and it handles different media outputs like watercolors, embroidery, and pen-and-ink with mixed results.
The original article contains 513 words, the summary contains 160 words. Saved 69%. I'm a bot and I'm open source!