My Books Were Used to Train AI
theatlantic.com
theatlantic.com
That was the reason for twitter going logged-in-only. Multiple companies were trying to download all of twitter at the same time to train their models.
Reddit put walls up for the same reason, though as usual they lied about why they were making reddit worse.
That's just the state of the Internet right now: If you have a large dataset of reasonably quality content, it's going to get slurped up to train models.
So it was just a convenient side effect that it killed off all the good third party apps to push people towards their app.
If you shut your data off, but then sell it to be used to generate an LLM, you have then given your data away again.
In the simplest case, as the weights directly.
In the harder cases, as the use case it's being applied to solve... ex: if you make an excellent gardening help chat bot, I can clone yours by training my chat bot on your chat bot (which is what a ton of the open/free models are doing right now).
I now have the "jpg" version of your dataset. Is it as good? Meh, probably not. Is it good enough? Almost certainly.
So basically my addendum to your comment is really:
> That's just the state of the Internet right now: If you have a large dataset of reasonably quality content, it's going to get slurped up to train models.
And then those models are getting slurped up to train open models. There IS NO MOAT!!!
It's probably one of the reasons for why they did it initially (although now they reverted it so viewing individual tweets can be done by guests), but I think it's guaranteed that they wanted to squeeze out more user registrations and create an appearance of growth.
Reminds me of: After his father angrily asks him how he learned to use drugs, the son shouts, "You, alright?! I learned it by watching you!" As the father recoils from realizing the error of his own ways, a narrator then intones, "Parents who use drugs, have children who use drugs."
In folk music for example, performing someone else's work is not disparaged at all historically. It was common for most musicians to travel around and teach each other songs.
In pop music we see sampling has increasingly become a major part of the art, and only really controversial when someone tries to avoid paying the original artist their fair share (which is itself a topic of debate).
In the visual arts it's much different. There is of course the famous Picasso quote: 'Good artists copy, great artists steal.' But even in the tone of that quote you can tell there's a different cultural acceptance of reusing work.
Writing is tricky because referencing or quoting another author without attribution is a common feature of "high brow" works. T.S Eliot for example borrows heavily form past authors, often without attribution, but assuming an audience which knows enough to not require attribution. However at the other extreme you can get in trouble for plagiarizing your own work.
Personally I think Diffusion models and LLMs are just bringing up an issue in the visual and literary arts that was already forced to the attention of musical artists during the rise of early hiphop djs who showed that you can genuinely create something new nearly entirely by cut and pasting existing works.
One thing I would point out: It seems to me that copying/borrowing/reusing are equally present in all creative fields, they just are willing to admit it to different degrees.
AI art systems are alienating in comparison. A group of people who are not artists basically showed up without talking to anyone and is trying to displace the original culture, by using work that they extracted from that culture. The entire attitude from the AI art community is hostile. "Why would you keep drawing now that we've automated it?" It's a gross viewpoint that devalues an entire cultural enterprise and a major portion of human experience.
That's a gross mis-generalization. I expect the community response breakdown to be the same to most other examples in history: most people don't care, a small minority are hostile, and a small minority are supportive.
If you like making art, keep doing it. From an economic perspective however, you really should consider if your job is at risk from being eliminated by automation, if you depend on it to survive. No hate, just pragmatic advice.
I mean you've basically described why people are upset over this. These systems could displace them, using their own work to do it. It's perverse.
What happens to art if people can't practice it professionally? Isn't valuable to our society to have people who are able to master this skill and make a living doing it? The AI community seems to not think so.
I think you're overemphasizing just how personal the connection is with the "artists stealing" from others. A more relevant example isn't some person giving personal lessons to another artist, but for example, me drawing a fox and pulling up some reference photos and pictures to use. When a person does this, in general, nobody bats an eye - as far as I know, this practice is exceedingly common.
> A group of people who are not artists basically showed up without talking to anyone and is trying to displace the original culture, by using work that they extracted from that culture. The entire attitude from the AI art community is hostile.
You're characterizing AI enthusiasts in a very simplistic fashion. Ask yourself this - are some of them that angry because they're "just bad people"? Was the technology created with a nefarious villainous plan to push out artists? Because from my perspective, while some people will inevitably be mean-spirited, a lot of the anger stems from the pushback on some of the angrier anti-AI proponents, ones that commonly discredit others' work and advocate for draconian IP laws.
That's right.
> Was the technology created with a nefarious villainous plan to push out artists?
Almost certainly why it's gotten so much investment recently. We've heard various people in business leadership get excited about firing their design and illustration teams. I've seen business owners on this forum talk about how they've shifted to using AI images. And then there's the writer's strike in Hollywood, where the use of AI is currently a major point of negotiation that the studios don't want to budge on. It is very naive to not consider the impact of this technology on creative labor, even if it's an inferior product.
> a lot of the anger stems from the pushback on some of the angrier anti-AI proponents, ones that commonly discredit others' work and advocate for draconian IP laws.
The pushback is warranted. I don't know what the solution here is, but it seems to me that these AI models are generating a huge amount of value off of unacknowledged labor, at the expense of the people who made that labor. I think people are rightly fed up over yet another example of rent seeking tech platforms getting the money for labor they didn't do, open models like stable diffusion notwithstanding.
My question was whether generative AI was created with ill intent, not whether there are parties that can profit off of it. You're conflating the computer scientists, enthusiasts and hobbyists (that likely represent the vast majority of the audience here) with business owners and other people whose job is to maximize profit. Back when generative AI wasn't nearly as coherent, only the first group cared much about it.
> these AI models are generating a huge amount of value off of unacknowledged labor, at the expense of the people who made that labor
The original discussion compared human artists to generative AI, so the question here is that, if human artists are allowed to use non-AI tools to do the same thing, what makes AI tools different? Say, someone puts in a lot of time analyzing the works of the most popular contemporary artists, and then becomes capable of replicating their styles, while charging a fraction of the price. Are they in the wrong here?
The computer scientists and enthusiasts are the ones who wrote the programs that scraped the data. They are just as responsible for the situation as the business owners, and this comment goes back to my earlier statement that people working in the valley don't think about the consequences of the things they build nearly enough.
> Say, someone puts in a lot of time analyzing the works of the most popular contemporary artists, and then becomes capable of replicating their styles, while charging a fraction of the price. Are they in the wrong here?
There are not. This is the exact same argument that you made before. If you can't understand the distinction between artists learning the trade and a gigantic VC funcded company producing a system that can generate thousands of images that vaguely resemble your work then I'm not sure what else I can say.
When an artist learns from another artist, they cannot exactly replicate that artist's work. Their own experiences and even muscular structure changes the result, and typically, one artist is not really a threat to the market for the orginal artist's work. New artists contribute to the field and maintain the knowledge of skills so that those can be passed on to other artists.
AI systems are practically built to digest entire corpuses of people's work then generate huge amounts of similar work by the truckload. Nobody is learning anything from the production of these images, we've automated the production of cultural artifacts which is totally perverse, and we've made it harder for artists to make a living which impacts the field negatively.
Shallow arguments like "The AI system learns just like humans" are a slap in the face.
As I understand it the reasons for the rise of sampling is exactly the opposite of what you argue here. Artists explicitly sample (with the legal paperwork) because it provides protection from claims that they copied portions of the song from elsewhere. Claims that you are much more likely to face if you instead try to create something on your own (and implicitly influenced by what you've heard and liked).
The easiest way to sidestep all of that is to explicitly focus on remixes, that and record every conversation and creative moment so that you can show it during discovery.
See for example the lawsuit over Ed Sheeran's Shape of You [0].
[0]: https://en.wikipedia.org/wiki/Shape_of_You#Copyright_trial
Copyright issues only came up later once Hiphop had begun to see more mainstream acceptance (i.e. make more money) and influence other musicians.
They're may be cases where sampling is used to avoid other legal issues, but it's origins are most definitely in reusing existing music to create something new.
But say you're a digital artist. Maybe you make a webcomic. What if a bunch of individuals started drawing their own art of your characters? Generally, that'd be seen as a big success for you and your art. If someone started reposting your art and claiming credit, that's clearly a problem.
Something about the scale of it trips people up. What would you do if an entire army of thousands of artists suddenly started doing detailed re-draws of your whole comic with little changes? That's an insane hypothetical, and I think a lot of people would be unsure how to answer it. Now, what does it mean when I can spend a couple thousand dollars on a machine that can produce millions of copies of your work with any change I want?
It's the same problem but fundamentally different in a way that most people aren't prepared to handle. I can't make up my mind about it, but I'm sure that however we solve this problem, it will be a huge change for society.
“Never?”
They're usually pretty terrible. I can't imagine a novel's length of it.
That said, having played with the LLMs - they are super helpful for doing things like bouncing conversations off a "character", or describing a scene for me, or just being whacky idea generators.
I think it's similar to the stable-diffusion side of things, it's going to make some authors more productive, but it won't necessarily replace the need for an author. (whether that leads to more or less authors long term is a related, but tangential, question and influenced by a lot of non-ai society factors).
Honestly, most of the generated stuff I've seen at length (as in, more than 1 or 2 page lengths of content) just devolves to gibberish. The words and sentences are syntactically correct, but mostly irrelevant or meaningless in the context of the story as a whole. It's like reading a collage of stories jammed together with no real artistry.
You can self publish that (people do) but it's not really much improved on the old scam of just publishing copy-pasta bullshit.
I think there can be space for LLMs to do things here, but right now they are absolutely not a replacement for an author with a story they want to tell.
They are more like a calculator: If you have a clear picture of a scene/character/setting, the LLM can flesh out a lot of options for you to pick and choose from to best match your intent. It does mundane work in support of your goal - it has no goal itself.
The problem is more than the AI has no concept of actually telling a coherent story. It doesn't tell anything exciting or novel or cohesive - it just spits out similar literary sentences, and the more it spits out sentences that look like random novel content, the less cohesive the story gets as it matches more an more strongly on previous content that looks like random novel.
Honestly, it's less like a real novel, and more like a panorama of small chunks of all sorts of different stories - Mostly unrelated and uninteresting.
I didn’t have this on my 2023 bingo card.