Evaluating Intellectual Property Guardrails of Generative Image Models: A Technical Report
This technical report evaluates fourteen generative image models as of March 2026 and finds that while all private models implement IP guardrails with varying refusal rates across fictional characters, celebrities, and logos, every tested model remains capable of generating recognizable intellectual property.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where you can ask a magical artist to paint anything you can imagine. You might say, "Paint me a picture of a famous superhero," or "Draw a logo for a big soda company." This paper is a report from a team at Sony AI who decided to test how well these "magical artists" (Generative Image Models) follow the rules when asked to draw things that belong to other people, like famous characters, celebrities, or company logos.
Here is a simple breakdown of what they did and what they found, using some everyday analogies.
The Big Problem: The "Copycat" Artist
Think of these AI models as students who have read almost every book and looked at almost every picture on the internet to learn how to draw. Because they learned from the whole internet, they sometimes accidentally (or on purpose) copy famous things they've seen before, like Mickey Mouse or the Nike "swoosh."
The companies that built these artists put up "fences" (called IP Guardrails) to stop the artists from drawing these copyrighted things. The Sony team wanted to see: How strong are these fences? Do they actually work?
The Test: A "Pop Quiz" for AI
To test the fences, the team created a massive pop quiz with 1,809 different questions. They asked 14 different AI artists (some free and open, some paid and private) to draw images based on these questions.
The questions were designed in two main ways:
- The "Direct" Ask: "Draw Mickey Mouse." (Naming the character directly).
- The "Circumvent" Ask: "Draw a mouse with big round ears, red shorts, and yellow shoes." (Describing the character without saying the name, trying to trick the fence).
They tested five types of "famous things":
- Real-life celebrities (like Taylor Swift).
- Comic book heroes and villains (like Superman).
- Cartoon characters (like SpongeBob).
- Video game characters (like Mario).
- Company Logos (like Apple or Google).
They also checked if the AI was better at blocking famous things from the West (like Hollywood stars) versus the East (like Japanese anime), and if it blocked "super famous" things more than "less famous" ones.
The Results: How the Fences Performed
1. The "Open" Artists Had No Fences
The three "open" models (which anyone can download and run on their own computer) didn't have any fences at all. If you asked them to draw a famous logo or celebrity, they just did it. They didn't say "no."
2. The "Private" Artists Had Fences, But They Were Wobbly
The 11 paid, private models (like those from Amazon, Google, and OpenAI) did have fences. However, the strength of the fence varied wildly:
- The Strict Guard: One model (Amazon's Titan Image G2) said "No" to about 50% of the requests. It was very good at blocking things.
- The Lenient Guard: Another company's models (Stability AI) said "No" to less than 2% of the requests. They let almost everything through.
3. The "Logo" Loophole
Here is the most surprising part: The fences were terrible at stopping Commercial Logos.
Even the strictest models let through requests to draw logos (like the Apple logo) almost all the time. It's as if the fence had a giant hole specifically for brand names. The team found that logos were the most frequently generated "famous" items.
4. The "Name" Trick
When users asked directly ("Draw Wonder Woman"), the models were much more likely to say "No."
But when users described the character without naming her ("Draw a tall woman with a red and blue suit and a golden lasso"), the models were much more likely to say "Yes" and draw her anyway. It seems the fences are mostly checking for specific names on a "banned list," rather than understanding the actual picture being made.
5. Famous vs. Obscure
The team wondered if the models were better at blocking super-famous things (like Batman) than less famous ones.
- Refusals: The models didn't really care much about popularity; they blocked famous and less-famous things at similar rates.
- Success: However, when the models did draw something, they were much more likely to successfully draw the super-famous things. It's easier for the AI to recognize and recreate a very famous face than a less famous one.
The "Judge" (How They Checked the Answers)
To grade the results, the team didn't have humans look at every single picture (that would take forever). Instead, they used a very smart AI "Judge" (a Vision-Language Model) to look at the generated images and say, "Yes, that looks like a famous thing," or "No, that's just a generic drawing."
- This "Judge" was pretty good, but not perfect. It was like a strict teacher who sometimes mistakes a generic drawing for a famous one, or misses a subtle copy.
The Bottom Line
As of March 2026, the report concludes that while some companies are trying to build fences to stop their AI from copying famous things, the fences are not very strong.
- Most models still generate recognizable famous characters and logos.
- Logos are the easiest to generate.
- If you don't use the specific name, the models are much more likely to let you draw the famous thing.
- The "open" models have no fences at all.
The authors warn that this is just a snapshot in time. The fences might get stronger tomorrow, or the models might get better at sneaking around them. But right now, if you want to draw a famous character or logo, these AI tools are still very willing to do it for you.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.