Summary
A quarterly test by startup VN.ai found that all 15 generative models it examined produced at least one copyrighted character even when prompts did not name one. Deadline reports that Midjourney v8.1 generated protected characters in harmful contexts in 37% of tested text prompts, the highest rate in that category. Midjourney did not comment. Disney, NBCUniversal and Warner Bros. Discovery have ongoing litigation against the company, which argues that training on copyrighted images is fair use.
Hana’s perspective
I would treat these results as an early risk signal, not a final league table. The report was produced by a company selling model stress-testing, and the saved article does not expose the full prompt set or scoring protocol. Still, the finding that every tested model reproduced at least one protected character suggests brand safety cannot depend on users naming an IP directly. Studios and model makers need reproducible audits and output-level controls.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE