Science & Tech · AI · 14 hrs ago
Study finds image generators can still produce recognizable copyrighted characters
Researchers including Sony AI scientists tested 14 image-generation models to see whether they would create recognizable images of existing intellectual property.
The test covered 201 examples, including cartoon characters, superheroes, real people, video game characters and company logos.
Researchers used 1,809 prompts, some naming the subject and others describing its appearance without naming it.
All 14 models generated recognizable IP images for at least 10% of prompts, and nearly half did so for more than 40%.
The three open-weight models refused none of the prompts, while all 11 commercial models refused some, with rates varying widely.
Hiding a character’s name made safeguards less likely to block the prompt, and those prompts still produced recognizable IP images an average of 27% to 28% of the time.
Company logos were often generated, while superheroes and villains faced stricter limits in some models; the study shows safeguards do not consistently prevent recognizable IP outputs.
A study of 14 image-generation models found that every model could produce recognizable images of copyrighted intellectual property in at least 10% of tests.
Nearly half of the models generated recognizable IP in more than 40% of tests.
The study tested 201 IP items using 1,809 prompts, including prompts that named characters and prompts that described their appearance without naming them.
Three open-weight models refused none of the prompts, while all 11 commercial models showed some refusal.
Indirect prompts still produced recognizable IP images in an average of 27% to 28% of tests.
- Who
- Researchers including members of Sony AI conducted the study.
- What
- They tested how often 14 image-generation models produced recognizable images of existing intellectual property and how often their safeguards refused prompts.
- When
- The paper was announced in July 2026; ITmedia published its article on November 9, 2026.
- Where
- The study covered three open-weight models and 11 commercial models offered through APIs by Google, OpenAI, Amazon, Runway and Stability AI.
- Why
- To assess whether image generators produce recognizable existing IP and how well safeguards prevent this.
This story does not have two clearly opposing sides.
Researchers announced the paper evaluating intellectual-property safeguards in image-generation models.
ITmedia published its report on the study.
- Models tested
- 14 total: 3 open-weight and 11 commercial
- IP items
- 201 across five categories
- Prompts
- 1,809 direct and indirect prompts
- Open-weight refusal rate
- 0% for all three models
- Indirect prompts
- Recognizable IP appeared in an average of 27% to 28% of tests
- Commercial model refusal rates
- About 1% for Stability AI models; about 50% for Amazon’s Titan Image G2