VN.ai, a startup that stress-tests generative AI tools for copyright problems, has published its first quarterly report on 15 popular models. Midjourney v8.1 came out worst at reproducing copyrighted characters, according to Deadline, which had the report first, and Yahoo's copy of it.
“It's great marketing. It works. It gets people to use your model and pay for it.”
Nelson Chu, founder of VN.ai, to Yahoo Entertainment · yahoo.com ↗
Midjourney also led on brand-safety violations, meaning protected characters placed in harmful settings such as violence or drug use: roughly one prompt in three (37%). One example showed the Joker smoking cannabis in a dress resembling Elsa's from Frozen. Every model produced at least one protected character, even from prompts naming none, though Seedance 2.5 has improved since studio cease-and-desist letters.
Midjourney did not comment and has argued that training on copyrighted images is fair use. VN.ai's founder Nelson Chu expects norms to shift as some developers set a good-actor example.
Disney and NBCUniversal sued Midjourney in June 2025, and Warner Bros Discovery has sued too; both cases are pending. Plagiarism Today reads the report as showing little change in about 18 months.