Midjourney v8.1 ranks worst of 15 in a VN.ai copyright audit
The startup says Midjourney put copyrighted characters into violent or drug-related scenes in 37% of prompts, and no model came out clean.
Key takeaways
- VN.ai's September audit of 15 popular image and video models ranked Midjourney v8.1 worst, with brand-safety violations in 37% of text prompts, roughly one in three.
- Every model in the audit drew at least one copyrighted character from a prompt that never named it, including Google's Nano Banana and ByteDance's Seedance 2.5.
- Deadline has separately reported that MiniMax H3 broke studio brand-safety rules in 55% of tests where an image was used in a prompt.
- VN.ai sells this copyright testing to studios and rights holders, and it has not published its prompt set, its method or the full ranking.
VN.ai ran its first quarterly copyright audit over 15 popular image and video models, and Midjourney v8.1 came out worst.
The startup stress-tests generative AI tools for copyright infringement, and its findings were published by Deadline's Jake Kanter on October 14. Read the ranking with the seller in mind: VN.ai sells this exact testing to studios and rights holders, it ran the study in September, and it has not published its prompt set, its method or the full ranking.
According to the report, Midjourney v8.1 placed copyrighted characters in a harmful context, such as violence or drug taking, in 37% of text prompts, roughly one in three. Deadline's account of the report gives one example: an image of The Joker smoking cannabis in a dress that could have come from Frozen's wardrobe.
No model in the audit was clean. Every one of the 15 drew at least one copyright-protected character from a prompt that never named it, the report says, including Google's Nano Banana and ByteDance's Seedance 2.5. Deadline has separately reported that MiniMax H3 generated clips that violated studio brand-safety standards in 55% of the tests where an image was used in a prompt.
VN.ai has tested these models before. In August, the company told Deadline that ByteDance's Seedance 2.5 blocked about 80% of prompts involving copyrighted material, far more consistent than earlier ByteDance models, though it still produced a faithful Shark Tale character without being named.
The audit lands in the middle of a live fight. Disney and Universal, Comcast's NBCUniversal studio, sued Midjourney on June 11, 2025, in federal court in Los Angeles, calling the image generator a "bottomless pit of plagiarism," and the case is ongoing. Warner Bros Discovery has also filed a copyright suit against Midjourney. Midjourney has argued that training its models on copyrighted images is permitted under fair use, and it did not respond to Deadline's request for comment.
VN.ai's founder, Nelson Chu, told Deadline that copying characters is "good for business" for model makers. "It's great marketing. It works. It gets people to use your model and pay for it," he said. Chu added that he expects improvement as more models set a precedent for good behavior.
For creators, the audit is a caution. If VN.ai's figures hold, an everyday prompt that never asks for a protected character can still return one, and the difference between a model that blocks it and one that does not may track the model's business model as much as its guardrails. The numbers here are VN.ai's, the underlying data is not public, and the company benefits when the problem looks large.
The full audit is not public. VN.ai's site routes studios, publishers and rights holders to request a forensic assessment at briefing@vn.ai, and Deadline's report on the findings is here.