Midjourney: Report Claims AI Model Is Hollywood's Worst Copyright Offender

Exclusive: Disney and NBCUniversal sued Midjourney last year, calling it a "bottomless pit of plagiarism." More than 18 months later, the company appears to have failed to clean up its conduct. VN.ai, a tech startup that tests generative AI tools for copyright infringement, has released its first quarterly report on how [...]

EXCLUSIVE: Disney and NBCUniversal accused Midjourney of being a "bottomless pit of plagiarism" in a lawsuit last year. More than 18 months on, the company has apparently failed to clean up its act.

, a tech startup that tests generative AI tools for copyright infringement, has released its first quarterly report examining how 15 popular AI models perform when it comes to ignoring studio IP.

Research conducted last month found Midjourney version 8.1 to be the worst offender when it comes to reproducing copyrighted character images. The AI model also outperformed the competition in so-called "brand safety violations," meaning characters were placed in "harmful contexts" such as scenes of violence or drug use. RELATED STORIES Disney execs offer $20 million for assets as startup files for Chapter 11 bankruptcy protection Everything we know so far about 'Percy Jackson and the Olympics' Season 3

According to a report shared with Deadline, Midjourney reproduced images of characters violating the brand's safety measures once in three text messages, or 37 percent of the time. This included a Midjourney v8.1 illusion that created an image of the Joker smoking weed while wearing an outfit that could have been plucked from Elsa's Frozen wardrobe. Watch on Deadline

Midjourney did not respond to requests for comment. Disney and NBCUniversal's lawsuit against the company continues. The startup has argued that training its AI models on images of copyrighted content is allowed under fair use.

The report found that each AI model generated at least one copyrightable character, even when a character's name was not referenced in a text message. This includes Google's Nano Banana and ByteDance's Seedance 2.5 model, although the latter has been significantly improved after receiving cease-and-desist letters from Hollywood studios.

As recently noted by Deadline, other Chinese AI models continue to pilfer IP, including the MiniMax H3, which produced videos that met the studio's brand safety standards 55 percent of the time when using an image in a notification.

When asked why AI models continue to ignore copyright, founder Nelson Chu said it's good for business. "It's great marketing. It works. It gets people to use your model and pay for it," he added.

But Chu foresees change. "I think at some point it's just not going to work out," she told Deadline. "And I think as more models build a track record of being good actors here, you're going to see more progress."

Get our breaking news alerts and keep your inbox happy.