Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When Your AI Art Generator Steals Your Style
Transcript
- Lucas: Luna, have you ever had that eerie feeling when you scroll through ai generated art and suddenly see something that looks exactly like a real living artist's work?? Not inspired by — I mean straight-up mimicking their signature style? Luna: Constantly. I saw a piece last week that could've passed for a new Greg Rutkowski painting. And it turns out his name is one of the most common prompts on Stable Diffusion. He didn't license anything. Lucas: Exactly. And that's the legal battlefield we're talking about today. There's a major class-action lawsuit brewing — actually, multiple suits — from artists who claim their copyrighted work was scraped without permission to train these models. The lead plaintiff in one of the biggest is Kelly McKernan, a Nashville-based illustrator whose ethereal watercolor style is now being churned out by algorithms. Luna: So the core question is: when an AI generates an image in the 'style of Kelly McKernan,' is that copyright infringement? Or is it just a really advanced form of inspiration? Lucas: Right. And the defendants — Stability AI, Midjourney, DeviantArt — they're arguing fair use. Their models, they say, learn like a human artist studies a master. The difference is scale. A human looks at maybe a few hundred paintings. These models absorb billions of images scraped from the internet, including copyrighted ones. Luna: And here's the twist: the training dataset for Stable Diffusion, LAION-5B, includes images from Pinterest, personal blogs, and even medical records. The artists never consented. They can't opt out. Lucas: That's the gut-level unfairness. McKernan testified at a Senate hearing earlier this year that she's had clients back out of commissions because they could just generate something 'in her style' for free. That's lost income. Real harm. Luna: So what's the legal theory? I know there's a novel argument around 'style' not being copyrightable per se. You can't copyright a vibe. Lucas: Correct. Copyright protects specific expressions — a particular painting, a photograph — not the general style. That's why the plaintiffs are also alleging violation of the Digital Millennium Copyright Act, because the models allegedly strip metadata and don't allow takedowns. And there's a state-level claim under California's right of publicity for Midjourney's commercial use of artists' names. Luna: That's interesting. So they're building a case from multiple angles. Has the court ruled on any preliminary motions? Lucas: As of late May, a federal judge in San Francisco dismissed some claims but allowed the core copyright and DMCA claims to proceed. The judge said, quote, 'It is plausible that defendants possessed unlawful copies of plaintiffs' works during the training process.' That's a big deal. Luna: So the case survives, at least for now. But even if the plaintiffs win, would the remedy actually help? I mean, you can't un-train a model. Lucas: That's the practical nightmare. The plaintiffs are seeking damages and injunctive relief — basically an order to destroy the models trained on their work. But that's technically impossible without retraining from scratch. I think the more realistic outcome is a licensing framework. We're already seeing it with Getty Images suing Stability AI and settling later. And Shutterstock launched a contributor fund to compensate artists whose work is used. Luna: Speaking of which, the EU's AI Act, which came into force this year, has a transparency requirement: any foundation model used in the EU must disclose the sources of its training data. That could force companies to show exactly what they scraped. Lucas: That would be a game-changer. Right now, these datasets are essentially black boxes. We know LAION-5B contains over 5 billion images, but we don't know which ones. The Act doesn't require individual consent, but it does mandate public summaries of copyrighted works used. That's a start. Luna: I think the core tension here is between technological progress and individual creators' livelihoods. And it's not just visual artists — writers, musicians, voice actors are all facing similar battles. Lucas: Absolutely. And this is where a lot of our listeners might feel conflicted. These AI tools are genuinely useful. I use them for brainstorming. But I also want artists to be paid fairly for their work. Luna: Yeah, I think that's the honest middle ground. And speaking of keeping things going with integrity — you know, a big reason we can explore these nuanced angles without ad interruption is because of listeners who support the show directly. A couple of dollars a month genuinely makes a difference for us at buy me a coffee dot com slash fexingo. If these conversations have given you something — a new perspective, a concrete thing to think about — that's where you can help keep it ad-free. Lucas: Totally. It's not about a big ask. It's just that listener support directly funds the research and hosting. And we're grateful for every contribution, no matter the size. Luna: So back to the art world. You mentioned licensing frameworks — is there any startup or initiative that's actually trying to solve this in a fair way? Lucas: Yes, a company called Spawning AI launched a tool called 'Have I Been Trained?' where artists can check if their work is in LAION-5B and opt out. But opt-out is not the same as opt-in consent. Another startup, Bria AI, claims to be trained only on fully licensed content. They're positioning themselves as the ethical alternative. Luna: But those are niche. The big players — OpenAI with dall e, Google with Imagen, Meta with make a scene — they've been more cautious about training data. But they're still not compensating individual artists. Lucas: Right. And the hypocrisy stings. These companies are built on the idea that data is free, but they'll sue anyone who scrapes their own platforms. Midjourney's terms of service actually prohibit scraping their generated images. So they're protecting their own output while feeding on everyone else's. Luna: That's a powerful double standard. So what's the best-case scenario here? If you're a policymaker for a day, what do you do? Lucas: Short term: require transparency in training data. That's the EU's approach. Medium term: create a statutory licensing scheme, like the one for music sampling. You want to train on a copyrighted image? Pay into a collective rights organization that distributes to artists. Long term: we need a serious conversation about what 'fair use' means in the age of generative AI. The old four-factor test wasn't designed for models that memorize and replicate. Luna: And if the courts decide that training on copyrighted data without consent is not fair use, it could fundamentally change the business model of generative AI. Some of these startups would collapse. Lucas: They would. And maybe that's not a bad thing. We'd be forced to build AI that respects creators from the ground up, not as an afterthought. But I'm not naive — the lobbying power of Big Tech is immense. The EU AI Act took years to pass. This is going to be a long, messy fight. Luna: Well, we'll keep covering it as it unfolds. For now, if you're an artist, what can you actually do? Besides join the class action. Lucas: First, register your copyrights if you haven't. That's essential for statutory damages. Second, use tools like Glaze or Nightshade from the University of Chicago — they add tiny perturbations to your images that confuse AI models. Third, water mark your work and include metadata. And fourth, be vocal. The more artists speak out, the harder it is for companies to ignore them. Luna: Practical steps. I like that. So what's next for the McKernan case? Is there a trial date? Lucas: Not yet. The discovery phase is going to be massive. Plaintiffs are asking for the full training dataset, which the defendants argue is a trade secret. That fight alone could take months. I'd be surprised if we see a trial before mid-2027. Luna: So in the meantime, the technology keeps evolving, and more artists' works get absorbed. It's a bit of a race. Lucas: It is. But I think the law will eventually catch up. It always does, just slower than the tech. And as consumers, we can choose to support tools that are transparent and fair. That's a small but meaningful power we have. Luna: Definitely. Thanks for unpacking this, Lucas. It's a lot to digest, but really important. Lucas: My pleasure, Luna. And thanks to everyone listening. We'll be back next time with another angle on the ethics of AI.