Sony AI: 14 image models leak recognizable IP at least 10% of the time

The study, titled "Evaluating Intellectual Property Guardrails of Generative Image Models: A Technical Report," was published in July by researchers affiliated with Sony AI and other organizations. It set out to measure not only how often image generators reproduce existing intellectual property but also how well the safety measures meant to prevent that actually work.

The test design separated named prompts from indirect ones that describe appearance without naming the character or brand. That distinction matters: top-refusing models such as Amazon's and Runway's blocked about 80% of direct prompts, but refusal rates fell when names were withheld. The results also varied by IP category — commercial logos were rarely blocked, while superhero and villain characters drew stricter filtering. Separately, although real famous people had the second-lowest refusal rate, images that were actually recognizable as the person were the least likely to appear across all categories.

The findings suggest that current guardrails are not a uniform shield. The body does not say how the researchers propose to close these gaps, so it remains an open question how providers might respond.