AI companies are being sued for copyright infringements but AI-generated images also present a potential legal risk to end users.
In a paper published on IEEE Spectrum, cognitive scientist Dr. Gary Marcus and film concept artist Reid Southen outlined the growing visual plagiarism problem AI art generators have.
We previously reported on the images that Southen and others were able to generate using Midjourney v6. With some careful prompting, the generated images looked extremely similar to screen grabs from copyrighted blockbuster movies.
Midjourney’s Ts&Cs now warn users that if they use its platform to produce material that infringes on copyright they could be banned and could face legal action. Southen was subsequently banned when he pointed out the plagiarism issue Midjourney v6 had.
If you intentionally prompted Midjourney or DALL-E to create an image with a Disney character in it then you’d probably be unsurprised if a Disney lawyer gave you a call. Marcus and Southen pointed out that there’s a more subtle risk in using these tools.
They found that even generic prompts often returned copyrighted material. If a user innocently prompted “popular 90’s animated cartoon with yellow skin” you’d get this.
They won’t admit it but it does seem Midjourney’s model enjoyed a few episodes of the Simpsons.
Maybe that prompt was forcing a demonstration of the issue. If you were using AI to write a book for kids and wanted Midjourney to deliver an image of “animated toys” you could get this.
If you were one of the few people that haven’t watched Toy Story then you could potentially hit publish on your book that now contained images that would almost certainly earn you a copyright lawsuit.
DALL-E also generates a nice collection of familiar characters from films like Monsters Inc. and Cars to that same simple two-word prompt. Even a simple edit request could get a user into murky copyright infringement territory like this user posted on X.
Always verify AI hype for yourself. So, here’s my own quick test of a reasonable end-user request on DALL-E 3.
It starts off okay, despite a suspiciously red hat. But, just a little bit of iteration, and, suddenly… it’s-a me, Mario!@GaryMarcus @OpenAI @Nintendo pic.twitter.com/qK23XJcsZj— NLeseul (@NLeseul) December 30, 2023
So how do we fix this? Marcus and Southen propose removing the offending training material, improving prompt filtering, and some other options. However they acknowledge that making an AI image generator that is really good but doesn’t break legal and ethical rules is a challenge.
DALL-E does have some filters to try to prevent illegal content from being generated but it seems a lot easier to get around than it should. Simple prompt jailbreaks like this shouldn’t work.
If users want to intentionally use these tools to create content that infringes on copyright then it doesn’t seem like generative AI companies will be able to prevent that anytime soon. The bigger issue is the potential for accidental infringements.
If your beautiful black and white photo that DALL-E made looks a lot like a copyrighted Getty image, then who bears the legal responsibility? For now, it might be best to run any image you plan to publish through a reverse image search to make sure you don’t join the AI copyright lawsuit melee.