As the animation and manga industry gathers to celebrate creativity at events like Anime NYC, it’s an opportune moment to address a pressing issue: the unauthorized use of intellectual property (IP) that is inadvertently embedded within contemporary generative AI models. These advanced AI systems have been built on a foundational principle: to collect and learn from vast amounts of data available on the internet. While this has resulted in powerful AI tools, it has also led to significant concerns regarding the integrity of the data they were trained on, raising questions about ownership and originality.
The Overlapping Worlds of AI and Intellectual Property
Recent IP Risk Audits (IRAs) conducted by 2nd Set AI highlight the extent of the challenges posed by these generative AI models. Our analysis focused on a wide range of iconic characters, revealing a troubling intertwining of these AI systems with existing intellectual property. This overlap is particularly pronounced with the leading figures in shonen manga, where characters like Monkey D. Luffy from One Piece, Son Goku from Dragon Ball, and Naruto Uzumaki exhibit a “SIGNIFICANT” risk rating in our audits. These characters have established a substantial digital presence over the decades, making them integral to the visual library that AI relies upon during its training.
The concern extends beyond established characters; even new heroes that gained popularity in the age of AI, such as Izuku “Deku” Midoriya from My Hero Academia, Yuji Itadori from Jujutsu Kaisen, and Tanjiro Kamado from Demon Slayer, are replicated with striking accuracy. This raises critical issues for creators and rights holders who seek to protect their intellectual property.
A Genre-Agnostic Problem
The contamination of intellectual property is not limited to specific genres or demographics. Our audits reveal that classic shojo figures like Usagi Tsukino from Sailor Moon can be just as easily replicated as foundational seinen characters like Motoko Kusanagi from Ghost in the Shell. The models employed in AI training have effectively scraped data from dedicated fan communities as well as gritty cyberpunk narratives, leading to a wide-ranging issue of IP contamination across various media formats.
The complexity of the training data creates a maze of potential infringement issues. For example, David Martinez from Cyberpunk: Edgerunners—a character from a Japanese/Polish anime based on a Polish video game—was accurately recognized by these AI systems. Even original digital IP like the VTuber Mori Calliope is captured with high fidelity, illustrating a new realm of risk for creators and rights holders.
The Shortcomings of Generic AI Models for Professional Creators
For serious intellectual property owners, the ability of a generic model to generate characters like Naruto is a significant flaw rather than a beneficial feature. These one-size-fits-all AI models present two primary challenges for professional production.
The first challenge is what we term "Fan Art Contamination." Because these models are trained on a chaotic array of fan art, they inevitably absorb a multitude of fan interpretations, including alternate character designs and relationships. This leads to inconsistencies in branding and character portrayal that are unacceptable for professional studios. The result is a character representation that amalgamates various fan perspectives rather than adhering to the official narrative and design.
The second challenge involves the evolution of character designs. Characters are dynamic assets that often undergo changes throughout their appearances in different media, such as from manga to anime or between seasons. A public AI model trained on this inconsistent history tends to blend all versions of a character into a generic representation that lacks specificity. This makes the model unreliable for professional use, where accurate representations are essential.
The Path Forward for Generative AI in the Creative Industry
To address these challenges, a new approach to generative AI is essential. Intellectual property holders need platforms that are specifically trained on their own unique universe, devoid of external influences. At 2nd Set AI, we advocate for the development of generative systems that are trained exclusively on official brand guidelines, style manuals, and a studio’s canonical assets. This methodology ensures that creators have the control necessary to produce IP with unwavering consistency and reliability.
Moreover, such a specialized platform must integrate safeguards that are conscious of intellectual property rights. These systems should actively prevent the generation of protected characters from outside a studio’s library. For creative teams building their own worlds, this means there is no risk of unauthorized references or character likenesses appearing, thereby ensuring a clean and secure chain of title.
This tailored approach also paves the way for the creation of new intellectual property. Our audits indicate that newer characters, such as Maomao from The Apothecary Diaries and Mizu from Blue Eye Samurai, showed zero replication in our assessments. This is indicative of their novelty, as they are too recent to have been included in the training data of generic models. This reveals a crucial insight about public AI models: they primarily act as vast archives of existing content and struggle to generate genuinely new ideas or characters that they haven’t previously encountered.
A purpose-built generative platform, in contrast, can focus solely on learning from proprietary character sheets and key visual art. This allows creators to establish new franchises within a secure and creative environment. By fostering a partnership where the AI learns directly from the creator’s vision rather than from a chaotic public web, we can carve out a future for creative intellectual property that empowers creators to explore both existing narratives and new possibilities.
In conclusion, the ongoing evolution of generative AI presents both challenges and opportunities for the creative sector. By prioritizing dedicated, IP-conscious approaches and ensuring that generative models are trained on official, proprietary content, we can navigate the complex landscape of intellectual property and creativity with greater confidence. The future of storytelling and character development lies in embracing these innovative tools while safeguarding the integrity of original creations.
