As the industry celebrates original creation at events like Anime NYC, it’s a critical time to talk about the ghosts in the machine: the unauthorized fingerprints of intellectual property haunting today’s generative AI models.
Many of these models were built on a simple philosophy: scrape everything. They absorbed massive amounts of art and IP from across the web. The result is a powerful tool, but one haunted by the data it was trained on.
At 2nd Set AI, our recent IP Risk Audits (IRAs) show just how deep this digital haunting goes. We ran diagnostics on a wide range of iconic characters, and the results reveal that these AI systems are thoroughly interwoven with existing intellectual property.
Foundational Characters, Foundational Risk
The issue is most pronounced with the titans of shonen manga. Our IRA for Monkey D. Luffy (One Piece) returned a “SIGNIFICANT” risk rating, as did audits for Son Goku (Dragon Ball) and Naruto Uzumaki. With decades of content, their digital footprints are colossal, making them core pillars of the AI’s visual library. The same holds true for other long-running hits like Edward Elric (Fullmetal Alchemist) and Jotaro Kujo (JoJo’s Bizarre Adventure).
This is not just a legacy problem. The new guard of heroes who rose to prominence during the AI boom are just as exposed. Izuku “Deku” Midoriya (My Hero Academia), Yuji Itadori (Jujutsu Kaisen) and Tanjiro Kamado (Demon Slayer) are all summoned with near-perfect fidelity.
A Problem Across All Genres
This IP contamination is agnostic to genre or demographic. Our audits found that classic shojo royalty like Usagi Tsukino (Sailor Moon) is just as easily replicated as a foundational seinen icon like Motoko Kusanagi (Ghost in the Shell). The models scraped passionate fan communities and gritty cyberpunk aesthetics with equal efficiency.
The training data also creates a complex web of infringement across media formats. For instance, David Martinez from Cyberpunk: Edgerunners, a character from a Japanese/Polish-produced anime based on a Polish video game, was perfectly recognized. Even digitally-native IP like the VTuber Mori Calliope is captured with a high replication rate, signaling a new frontier of risk.
Why Off-the-Shelf AI Fails Professional Creators
For any serious IP holder, a model’s ability to generate Naruto is not a feature; it’s a bug caused by these ghosts. These general-purpose models are unworkable for professional production for two key reasons.
First is Fan Art Contamination. The models are trained on the chaotic and often non-canonical world of fan art, learning fan-created relationships and alternate costumes. This introduces a level of brand inconsistency that is unacceptable for a professional studio. The result is a character that reflects a thousand different fan interpretations instead of the official ground truth of the IP.
Second, Character Designs Evolve. A character is a living asset. Their design changes from the manga to the anime, or even from one season to the next. A public model, trained on this entire messy history, blends all versions together to produce a generic character accurate to no specific era, making it useless for production.
A Better Path for Generative AI in Publishing
The way forward requires a fundamentally different approach. IP holders need generative platforms that know everything about their own universe and nothing about anyone else’s.
At 2nd Set AI, we believe the future of professional production demands systems trained exclusively on official brand bibles, style guides and a studio’s own canonical assets. This approach provides the control needed to generate IP with perfect consistency. It also allows for critical safeguards. A professional-grade platform must have IP rights-aware guardrails built into its core. These systems actively prevent the model from generating protected characters from outside a studio’s library. For a team developing its own world, this means there is zero risk of a stray Saitama or Aang appearing in the background, ensuring a clean chain of title.
This specialized approach also unlocks the most important use case: developing new IP. Our audits found that newer characters like Maomao from The Apothecary Diaries and Mizu from Blue Eye Samurai showed 0% replication. They are too new to have been scraped. This reveals a core truth about public models: they are vast archives of the past, unable to genuinely create what they have not already seen.
A purpose-built platform, however, can be designed to learn a new, private vision from proprietary character sheets and key art. This allows creators to build new franchises within a secure, generative sandbox. It creates a partnership where the machine learns directly from the creator, not from the public web. The future of creative IP is about giving creators tools that are laser-focused on their worlds, both the ones that exist and the ones yet to be imagined.
Jeff Smith is Co-Founder & CEO of 2nd Set AI, a first-party generation platform for entertainment leaders which uses clients’ proprietary assets to create exacting visuals using a unique generative engine. Learn more at 2ndset.ai.
