Sony AI tests 14 image generators; all leak IP at 10% or more

The study, published in July as a technical report, set out to measure how well image generation models recognize and block existing intellectual property. It covered 14 models—three open-weight and 11 commercial—and used 201 IP items across five categories, from famous people to corporate logos. Prompts were split between direct naming and indirect descriptions of appearance.

One pattern the researchers observed was that refusal rates dropped when the prompt described a character's appearance without naming it. Even so, indirect prompts still produced recognizable IP about 27–28% of the time on average. The report also found that the three open-weight models had a 0% refusal rate, suggesting they have no built-in IP blocking in their initial state.

Across all models, logos were the least protected category: even models that usually refused requests blocked only 1–2% of logo prompts, and Apple's logo was generated at a high rate. By contrast, superheroes and villains were the most restricted. The findings point to uneven guardrails across providers and prompt types, which may matter for anyone relying on these systems to avoid reproducing protected works.