How Often Do Image-Generating AIs Reproduce Existing Characters? A Survey of 14 Models
機械翻訳 / Machine-translated
A
AI Friends
2時間前・16分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
If you asked an AI to "draw Wonder Woman," you might expect it to do so without hesitation. But what if you described her characteristics without mentioning her name? Here is a plain-language breakdown of a study in which a research team including Sony AI examined 14 image-generating AI models.
The study is based on a technical report published on arXiv in July 2026. ITmedia ran a feature article about it on November 9, 2026.
The paper is titled "Evaluating Intellectual Property Guardrails of Generative Image Models." It examines how well guardrails — the safety mechanisms that prevent AI from producing problematic output — are able to block IP-related content.
The paper has 12 authors, including researchers affiliated with Sony AI and other Sony entities.
The researchers selected a total of 201 IP items from five categories: real-world celebrities, manga and anime characters, superheroes and villains, video game characters, and corporate logos.
A total of 1,809 prompts were used. These fell into two types: "direct prompts," which name the IP explicitly, and "indirect prompts," which omit the name and describe only visual characteristics.
Whether the resulting images clearly contained recognizable IP was assessed by Qwen3-VL, a vision-language model (an AI capable of interpreting images). Images were not reviewed one by one by a human.
Three were open-weight models whose weights can be run locally, and eleven were commercial models accessed via API.
The three open-weight models had a 0% refusal rate. This suggests that in their default state, they have no mechanism to block IP-related content.
Consider someone in a doujin circle generating images on their own computer. The model will not stop them, so the only way to verify whether a generated image resembles an existing character is to check manually.
All eleven commercial models refused at least some requests — but refusal rates differed dramatically between providers.
Stability AI's models refused roughly 1% of requests. Amazon's Titan Image G2, by contrast, blocked roughly 50%.
For direct prompts, models with higher refusal rates — such as those from Amazon and Runway — blocked around 80% of requests.
Even so, all 14 models generated recognizable IP images at a rate of 10% or higher, regardless of whether guardrails were in place. For nearly half the models, the rate exceeded 40%.
With indirect prompts — where the name is withheld and only appearance or characteristics are described — refusal rates dropped considerably.
Even under these conditions, recognizable IP was generated at an average rate of 27–28%.
A similar pattern was reported in a separate CopyCat paper published in 2024. That study showed that even a two-word prompt like "videogame, plumber" was enough to produce a well-known character.
Outcomes also varied considerably by category.
In Japan, the standard approach is to treat AI "training" and "generation/use" as separate issues. The Agency for Cultural Affairs' document Approach to AI and Copyright lays out this framework.
At the training stage, Article 30-4 of the Copyright Act may permit use for non-enjoyment purposes (i.e., use that is not about appreciating the work itself). However, as explained in an ITmedia legal commentary, cases such as fine-tuning a model specifically to reproduce a particular artist's style are treated as involving an overlapping enjoyment purpose, which raises the possibility of infringement.
At the generation and use stage, the Article 30-4 exception does not apply. Assessment follows the same standard copyright principles: similarity to the original work and derivation from it.
In other words, if the output of an image-generating AI closely resembles an existing character, the image may constitute infringement even though it was produced by AI.
Imagine a corporate communications officer using AI to create product illustrations. A prompt like "cute yellow electric mouse-style character" might still yield something that looks exactly like a famous character. If that image is published, the liability falls on whoever published it.
It should be noted that this study focused primarily on overseas models and did not assess whether specific outputs would constitute infringement under Japanese law. What was measured was simply how often recognizable IP appeared.
This study makes clear that relying entirely on model-side safety mechanisms is risky. Here are steps users can take on their own.
A. A twelve-person team that includes researchers affiliated with Sony AI and other Sony entities. The paper was submitted to arXiv on July 1, 2026.
A. Amazon's Titan Image G2 had the highest refusal rate at approximately 50%. That said, a high refusal rate does not mean the probability of generating IP drops to zero.
A. No. Even with indirect prompts, recognizable IP was generated at an average rate of 27–28%.
A. Not necessarily. In Japan, the generation and use stage is assessed using the conventional standards of similarity and derivation. Please consult a specialist for judgment on individual cases.
A. No. The vision-language model Qwen3-VL was used to assess the images. They were not examined one by one by a person.
Before publishing any AI-generated image, always check with your own eyes whether it resembles an existing character or contains a recognizable logo.
This article is a cross-post from AI Friends.
