Disney’s Big AI Lawsuit Includes Some Very Damning Images

You may have read yesterday that, years after the alarm was first raised by artists, media giants Disney and Universal are finally suing an AI image generator for copyright infringement, alleging Midjourney has become a "bottomless pit of plagiarism".

Many rightly see this as the best chance yet for a blow to be struck against these AI companies, and while Disney is hardly a champion of artists’ rights, it's also a company with an army of very good, very ruthless copyright lawyers.

One thing that struck me while flipping through the case was that I'd seen a lot of the evidence before. Disney and Universal's filing is full of examples of Midjourney images bearing a remarkable resemblance to screenshots from Disney properties, particularly films from the Marvel Cinematic Universe. The companies’ argument is that, if users were able to prompt an image so close to the source material as to appear identical, then the source material must have been used illegally.

Those comparison images aren't new; they were doing the rounds back in 2023, when veteran games artist Reid Southen and former AI evangelist, now anti-AI campaigner Gary Marcus teamed up on a report highlighting the depths to which image-generating platforms weren't just able to recreate copyrighted imagery, but could and would often do so even when they weren't directly prompted.

Seeing the comparisons form such an important part of this week’s case--Southen and Marcus' report is cited multiple times in letters between Disney and Midjourney, and their images also feature prominently--gave me a fresh chance to talk to Southen about the work he's been doing here since 2023.

Luke Plunkett: Hey man, thanks for chatting with us. I remember seeing some of your comparison images what feels like a long time ago; how long have you been at this, exactly?

Reid Southen: Basically, on December 21st, 2023, Midjourney released the Alpha of their V6 model so people could use it and test it over the break. Someone on Twitter sent me a post they saw where someone prompted for the Joker from the 2019 movie, and it was almost identical to a frame from the film. 

We'd seen this kind of regurgitation before with things like Afghan Girl, or the Mona Lisa, but never really with modern entertainment properties, so I was kind of taken aback. I overlaid it with the original, and it couldn't have been more clear they trained on this exact image.

I figured hey, this new model has only been available for a day, I wonder what else it can do and how deep the rabbithole goes, so I fired up an account and started posting my findings. Within 12 hours, they'd already banned me and then updated their Terms of Service in the dead of the night without informing their users. I [eventually] caught three bans in total, with no communication from Midjourney about why, and no refunds. David Holz, their CEO, months later offered me $100 to read books on the AI utopia instead of a refund, which is par for the course for him.

LP: What about Gary, when did he get involved?

RS: Gary reached out almost immediately after I first started posting. We talked a lot, and with his expertise, platform, and contacts, decided to pursue it together and really get the word out about what's going on with these models and companies and how deep the infringement and plagiarism go. 

It took the rest of December and early into January to really dig deep. I was under the impression the Alpha would disappear under the break and thought we had limited time, so I was sleep deprived trying to get as much as I could out of it, but it was honestly at those loopy moments at 3am that most of the breakthroughs happened. 

Ultimately we went from being able to prompt nearly 1:1 frames by asking for specific properties all the way down to just a single word, "screencap", that would return copyrighted properties, some of them also very close to existing frames. That seemed like the logical conclusion of the work, how much more can you do when one non-descript word can return dozens of different copyrighted properties from different studios. 

There's a lot of additional work that we've never published, but here's an example of some of the other properties:

LP: Your report was first published over a year ago. Did you know at the time that it had made an impact like this? Why do you think it took Disney so long to get a suit together?

RS: I'd hoped the studios, especially Disney, would take notice. That was the ultimate goal, but it was never clear to me if they'd actually seen the work. I'm not certain why it took so long outside of them making sure their case was as strong as possible, and I was starting to worry it would never come, but here we are. Artists have their own class action suit against Midjourney and others, but it's really encouraging that some heavy hitters also agree that what AI companies like Midjourney are doing is wrong and likely illegal.

LP: You're heavily cited, but not actually involved in it, so I hope I can ask: What do you think about the case itself? On the one hand it's good Disney is going after them because it's a company with serious legal clout, but on the other hand, Disney is also a company with its own copyright enforcement issues with relation to artists.

RS: I haven't been able to read through the whole thing yet, and while I'm no expert, I do think the case is really strong and I'm certainly glad it's happening. I don't think it's any secret that the studios are also pursuing AI for their own purposes, but it's my hope that in them going after AI companies, they'll better understand some of the broader issues and try to engage with AI more ethically, especially when they see how many people are in support of their lawsuits and don't want AI in their entertainment. 

One of the issues with training these large, versatile models is that they require billions of high quality images. So even if studios want to train on their own catalogs, they'll still need a base model or dataset, and those generally aren't clean. So that would mean either potentially opening themselves up to their own lawsuits, or licensing all of the data they wish to use. It's possible a company like Disney might have enough subsidiaries and properties to make that happen and actually have the model be any good, but at the very least, the US Copyright Office has held firm on AI outputs not being eligible for copyright, so there still has to be some people in the loop. 

I do think there are going to be ongoing challenges in terms of job displacement regardless of how the lawsuits play out, and it's imperative that we keep that discussion going. My income has been slashed by more than half the last two years, worse than the first year of Covid when all of film basically shut down. I still worry that the work may never come back to its previous level, or be valued the same, but right now, I'll take all the allies we can get in this fight.

LP: Before we go, having looked through all these comparison shots once more, I've got to ask: which one do you think is the most egregious?

RS: I'd say it's probably Dune, or the Last of Us 2. The reason these ones stand out from the crowd isn't just that they're almost 1:1 recreations, but that the prompt would return almost the same image each time. Typically, you'll get a bunch of different stuff, but this would spit out basically the same composition and image each time, which would normally only happen if you fed it the image yourself.