News

A Report Says Midjourney Reproduced Protected Characters in 37% of Prompts: Why That Number Needs Its Method Before You Use It

26 October 2026 · RenderBob team

Deadline reported that research named Midjourney the worst offender for producing protected Hollywood characters. The 37 percent figure comes from a secondary summary.

Deadline reported on 14 October that a research group had named Midjourney the worst offender for producing protected Hollywood characters. We could not open Deadline's article, so the figures below come from a secondary summary.

What is reported

A Medium post dated 19 October, citing Deadline's exclusive, says research from VN.ai, shared with the outlet, found that Midjourney v8.1 reproduced characters in violation of brand-safety measures in about 37% of text prompts. The post calls that the worst result among the models tested. Deadline's headline, which we did see, says the report "claims" Midjourney is the "worst offender for flouting Hollywood copyright". Midjourney did not respond to Deadline's requests for comment, according to the summary.

What we do not know

• The method. The summary says the testing method was not seen. We do not know the prompts, how many were used, or how a violation was judged.
• The comparison. The other models are not named in the summary.
• The version. The tested model was v8.1. The summary says Midjourney's default is now V8.2, from release notes we have not read.
• The number itself. The summary calls 37% "about one in three", which is closer to 33%, so even the secondary source is not consistent.
• The author. We know little about VN.ai.

The legal backdrop

Disney and NBCUniversal sued Midjourney in 2025, in a case widely reported as calling it a "bottomless pit of plagiarism". We found no new ruling in the coverage we saw. A research report is not a court finding and does not change that case.

Reading a report like this

• Ask for the prompt set. Prompts that name a character differ from prompts that merely describe one.
• Ask what counts as a violation. An exact copy and a close resemblance are different.
• Check the version and date. Results for one release may not apply to the next.
• Look for the comparison. A rate means little without the models beside it.
• Note who commissioned it. We do not know that here.

What to do in your own work

• Do not prompt with the names of protected characters, and review outputs for lookalikes before release.
• Keep a record of prompts and references, so you can show what a shot was made from.
• Ask your client whether their contract limits the generators you may use.
• Read the vendor's commercial terms, which are a separate question from what its filters block.

What to watch

Deadline's full article and the report itself, a response from Midjourney, and whether similar tests are published for video models.

RenderBob 3 and this story

RenderBob 3 keeps references and prompt logs with the studio, and because the model is a per-clip choice, you can swap one clip's generator if a vendor's guardrails worry a client, without recutting the film.