A startup's copyright audit of 15 AI image models ranks Midjourney v8.1 worst: copyrighted characters in 37% of prompts
A startup's first copyright test of 15 AI image and video models ranks Midjourney v8.1 worst, and every model drew a protected character nobody named. The tester sells this testing to studios, and its method and full ranking are not public.
A startup that tests AI models for copyright problems has ranked 15 popular image and video models, and Midjourney v8.1 came out worst. VN.ai's first quarterly report, run in September and shared with Deadline, found that Midjourney put famous characters into harmful scenes, such as violence or drug taking, in 37% of text prompts, roughly one in three. Its example is the Joker smoking cannabis in a dress lifted from Frozen's Elsa.
No model came out clean. Every one of the 15 produced at least one copyrighted character even when the prompt never named it, including Google's Nano Banana and ByteDance's Seedance 2.5, which VN.ai says has improved a lot since Hollywood studios sent cease-and-desist letters. MiniMax H3 broke studio brand-safety rules in 55% of video tests that started from an image. Midjourney did not respond to Deadline, and the lawsuit Disney and NBCUniversal brought against it is still running.
Read the ranking with its maker in mind. VN.ai sells this exact testing to studios and rights holders, its prompt set and method are not public, and the full ranking of the other models has not been released. Its founder, Nelson Chu, told Deadline that copying characters is "good for business" for model makers, because "it gets people to use your model and pay for it."
In practice, which model made a shot is turning into a clearance question. If you deliver AI work to a studio or a brand, expect to be asked, and check every frame for a character you never asked for.
