This is Hakase Shirai (X@o_ob), editor-in-chief of AICU. From April to May 2025, I traveled between France, San Francisco, and Los Angeles, investigating the "unseen online truths" of generative AI, particularly in the creative AI space.
This report delivers the reality I witnessed on the ground: the copyright interpretations of West Coast generative AI players shaken by ChatGPT and Ghibli-style imitation, the crucial perspectives and understanding of Japan in this context, and the dialogue happening at the forefront.
On March 26, 2025, Japan time, OpenAI released image generation capabilities for its foundational model, "GPT-4o."
The "Ghibli-style AI avatar" meme exploded on X. With GPT-4o's image generation, an era arrived where anyone could recreate the world of Studio Ghibli.
While OpenAI CEO Sam Altman didn't seem enthusiastic at first, the next day, March 27, he changed his own X profile picture to a "Ghibli-style" image.
be me
grind for a decade trying to help make superintelligence to cure cancer or whatever
mostly no one cares for first 7.5 years, then for 2.5 years everyone hates you for everything
wake up one day to hundreds of messages: "look i made you into a twink ghibli style haha"— Sam Altman (@sama) March 26, 2025
This is written in the typical ">be me" greentext meme format, expressing the irony and humor in the gap between a serious research career and suddenly being turned into a Ghibli-style meme online. In context, it's a self-deprecating joke, as if told by someone like Mark Chen.
The reply from Emad Mostaque, former co-founder of Stability AI, is particularly ironic.
My timeline is AGI
All
Ghibli
Images
— Emad (@EMostaque) March 26, 2025
I found the person at OpenAI who developed the Ghibli-fication, so I'll quote them here.
I'm Mark Chen (@markchen90), Chief Research Officer at OpenAI and I've uploaded my profile picture. Generate an image of a Tweet about the launch of GPT-4o native image generation - include a Studio Ghibli-style generation in the post. pic.twitter.com/IjhcSxYKik
— Mark Chen (@markchen90) March 27, 2025
This post can be seen as evidence that a chief research officer at OpenAI officially acknowledges and promotes "Ghibli-style image generation" as part of the PR for GPT-4o.
・Style imitation is mentioned at the level of an official OpenAI account and its research lead.
・The specific phrase "Ghibli-style" is used.
・There is intent to use it for "demo purposes."This indicates that Ghibli-style image generation is moving beyond personal use and beginning to spread into commercial and promotional applications. It is a highly suggestive piece of evidence in the context of debates around copyright, reliance, and style imitation.
Notably, at the end of March, Hayao Miyazaki's masterpiece "Princess Mononoke" was re-released in US IMAX theaters as a 4K remaster, grossing an estimated $4 million.
https://hotakasugi-jp.com/2025/03/31/anime-news-mononokehime-imax-boxoffice/
『もののけ姫』米国IMAX再上映で400万ドル突破！週末興行収入ランキング6位に - Film Goes with Net
宮崎駿監督の名作『もののけ姫』が、米国のIMAXシアターで4Kリマスター版として再上映され、推定400万ドルの興行収入を記
hotakasugi-jp.com
To put that "$4 million box office" into perspective, its original release on July 12, 1997, earned ¥20.18 billion, surpassing "E.T." (1982) at the time to become the highest-grossing film in Japanese history. In the United States, its distribution by the Disney subsidiary Miramax was a major topic. It earned a cumulative $194.3 million through theatrical releases and home media. The reality is that the weekend IMAX screenings alone more than doubled those initial earnings.
The beautifully remastered 4K IMAX version of "Princess Mononoke" was highly acclaimed. Its estimated weekend box office revenue of $4 million (¥570 million) exceeded expectations, landing it at number 6 in the North American box office rankings. With only 330 screens, its per-screen average was $12,134 (¥1.75 million), also in the top 10 nationwide.
Amid AI craze, Studio Ghibli's cult classic scores box office record - Entertainment News | The Financial Express
Studio Ghibli's 40th anniversary celebrations have been marre
https://www.financialexpress.com
For reference, the "Ghibli Park," which opened in November 2022 in Aichi Earth Expo Memorial Park, was projected to have an annual attendance of about 1.8 million people and an economic ripple effect of about ¥48 billion per year. The plan for Ghibli Park was agreed upon between Aichi Prefecture and Studio Ghibli in May 2017, developed over five years since the agreement and 17 years since the 2005 Aichi Expo. In reality, attendance is reportedly outpacing projections at around 2.8 million. Assuming an average ticket price of ¥3,800, the direct weekly revenue is ¥200 million. While the annual economic impact is said to exceed ¥48 billion thanks to lodging for out-of-town visitors, this economic effect is not direct revenue for Studio Ghibli.
At the very least, the ability to "imitate a style" means the ability to express a worldview. The perception that "Ghibli-style images can be easily made with AI" and the understanding of this on the Japanese side should not be swept away by memes and sentiment. It is not "free." We must properly grasp the numbers—the economic impact in terms of park attendance and actual remaster sales.
https://www.pref.aichi.jp/uploaded/life/438876_1978771_misc.pdf
https://www.pref.aichi.jp/uploaded/attachment/488369.pdf
By the way, there are definitely opinions in the U.S. like, "Stop messing around with Ghibli-style AI images and go see it in IMAX."
I think everyone should go see Princess Mononoke in IMAX to combat this stupid AI crap https://t.co/IDl11ew8KQ
— LEX 🔜 HOME (@_larxene) March 28, 2025
The ability to "imitate a style" means the ability to express a worldview. Rather than sticking to the single point that "style is not protected by copyright," we should recognize that the "worldview" created by Japanese IP and its economic effects should not be ignored.
This is the phenomenon that Ghibli-style AI image generation was causing in the United States.
In the US AI industry, researchers of Chinese descent are overwhelmingly leading the development of image generation AI. The lineage can be traced back to the anime-style image dataset "Danbooru2018," created by early Bitcoin researcher Gwern (@gwern), which made its way to China via the National University of Singapore (NUS). Since then, the paper trail has followed a path like this:
While one cannot generalize by ethnicity, the historical background shows that the technical precision of style imitation in vision technology and "ACG" (Anime, Comics, Games) related image generation algorithms is largely due to the "contributions as a hobby or research" by many who love Japanese anime and manga, a context linked to China's 2001 entry into the WTO. Meanwhile, a battle is being waged on the US and global stage, with West Coast Big Tech and China's Alibaba ("Wan") and Tencent ("Hunyuan") releasing new models and features almost weekly, a situation Japan cannot afford to ignore.
In Japan, the need to evaluate the increasingly prominent challenges of generative AI across countries and fields was recognized at the G7 Hiroshima Summit in May 2023. Since around 2023, the Agency for Cultural Affairs has been holding discussions in the Copyright Subdivision of the Council for Cultural Affairs on the theme "Views on AI and Copyright."
On March 15, 2024, the Subcommittee on Legal Systems made this clear in an "official document."
文化審議会著作権分科会法制度小委員会（第5回） | 文化庁文化審議会著作権分科会法制度小委員会（第5回）
https://www.bunka.go.jp/seisaku/bunkashingikai/chosakuken/hoseido/r05_05/
This was the first time Japan's Agency for Cultural Affairs had clearly communicated externally the two-step standard of "reliance" (ikyosei) and "similarity" (ruijisei). According to Gemini 2.5 Deep Research, the earliest official documents that can be identified where the Agency for Cultural Affairs clearly states that both similarity and reliance are necessary for copyright infringement to be established are the "Views on AI and Copyright (Draft)" dated July 2023 and the "Views on AI and Copyright" dated December of the same year (with an update on March 15, 2024). These relatively new documents are based on the premise that the principles of similarity and reliance themselves have "long been established" in case law.
※ However, this does not include specific documents, such as copyright manuals or circulars, that the Agency might have issued in the late Showa era (e.g., immediately after the 1978 Supreme Court ruling on reliance) that explained these two requirements in combination with similar clarity.
This is an important guideline for clarifying copyright infringement judgments in the age of generative AI and has influenced domestic discussions. On the other hand, this policy is only in Japanese and seems to be understood only by the diligent and knowledgeable people in Japan who follow copyright matters. AICU has also introduced public comment periods and other initiatives, but these have not led to discussions on a new copyright framework from a creator's perspective that can deal with global players, nor has it resulted in member-sponsored legislation.
https://www.bunka.go.jp/seisaku/bunkashingikai/chosakuken/pdf/94037901_01.pdf
A recent example is the law concerning the prevention of unauthorized recording of movies in theaters, which was passed by member-sponsored legislation during the 166th Diet session, promulgated as Law No. 65 of 2007 on May 30, 2007, and enacted on August 30, 2007. Copyright law is reviewed and revised with relative frequency as industry and media technology evolve.
So, how are AI and copyright defined in other countries? Recently, at the awards ceremony of the "Odyssey" international film festival for generative AI in San Francisco, I had the opportunity to discuss this with an executive from a startup whose image generation AI model is known for producing anime-style images. I'll keep the model's name confidential, but imagine an image generation AI where you can type prompts like "Dragon Ball," "Son Goku," or "Pikachu" and get an image of that character. He told me, "Only Japan cares about that copyright."
Project Odyssey - Season 2 データで見る、AI動画生成の現在。
https://corp.aicu.ai/ja/odyssey-season-2-20250512
Indeed, even among Japanese lawyers, there are slightly varied opinions on services that generate images based on trademarks or character names. Here is a legal summary of unauthorized AI training, reliance, and similarity in Japan, China, and the US.
Regarding AI training, Japan is "overwhelmingly lenient," while the US and China are stricter on similarity. This is one element behind the statement, "Only Japan insists on its own copyright." Furthermore, in the US, on top of fair use allowing for purposes like study, research, and public interest, everything is "common law," meaning case law is king. In this case, the judiciary may be more powerful than the legislature.
What's distinctive is that Japan's Agency for Cultural Affairs has been conducting discussions from 2024-2025 that clearly separate reliance and similarity. The argument that "training is illegal if there is reliance" tends to be more emphasized than in other countries, but it's hard to say this view has permeated internationally. This is the second point behind the "Only Japan cares about its own copyright" comment.
In both Japan and the US, "style"—such as the touch of an artwork, a particular artistic style, or compositional tendencies—is not protected by copyright law, as it's treated as an idea. For this reason, style imitation itself, like "Ghibli-style" or "Pixar-style," is not considered copyright infringement. If an AI-generated product is deemed "substantially similar," US case law questions whether the AI-generated image is "substantially similar" to an existing copyrighted work. Meanwhile, although there are issues like reproduction rights and public transmission rights when uploading for Image-to-Image (i2i), and rights of publicity, it is hard to say that Japan's Agency for Cultural Affairs is paying attention to the rights involved in uploading for i2i or LoRA, which precedes the unauthorized training in AI image generation.
Furthermore, in US copyright, the concept of "Fair Use" is very powerful. Copyright resolution methods like Creative Commons are also possible. In Japan, there is no clear definition for the concept of "commercial use."
While copyright is the most powerful right in this field, protection and solutions are also being sought through trademark law. It also relates to areas like "Trade Dress" and "False Designation of Origin." Recently, the Ministry of Economy, Trade and Industry (METI) announced its views on the unauthorized use of voice actors' voices from the perspective of the Unfair Competition Prevention Act, Design Act, and Trademark Act.
https://www.meti.go.jp/shingikai/sankoshin/chiteki_zaisan/fusei_kyoso/pdf/028_04_00.pdf
The model developer in question told me, "Since the Ghibli-fication of ChatGPT, our goal has been to surpass that quality," "We have many users from Japan, and we are keeping a close eye on that data," and "Only Japan cares about its own copyright." He's exactly right. The reproduction of nuances not just of "Ghibli-style," but of IPs like Dragon Ball and Gundam, has effective economic value.
Under Japan's copyright system, restrictions on unauthorized AI training are relatively loose, with the possibility of illegality only indicated in cases of "reliance." This was clarified in the Agency for Cultural Affairs' 2024 report and can be considered a pioneering global response.
On the other hand, the US often relies on flexible judgments based on fair use and common law, and style imitation tends to fall outside the scope of copyright. Lawsuits based on "Substantial Similarity" also tend to focus more on the output of the image itself rather than proving reliance.
The case of an OpenAI research lead publicly posting and utilizing a "Ghibli-style image" symbolizes the current situation where generative AI output is approaching cultural IP, blurring the boundaries of copyright, trademark, and publicity rights.
As shown by the sales of "Princess Mononoke" mentioned earlier, both reliance and economic impact are clear.
Of course, one could argue, "Looking at a timeline of Ghibli-fied images on ChatGPT made me want to go see it in IMAX," or conversely, "If you make a certain art style trendy with ChatGPT, it could become its own ad promotion." In fact, there was a moment when a "content policy" was set on ChatGPT, and Ghibli-style images could no longer be generated.
The San Francisco generative AI company in question was of the MAGA (Make America Great Again) persuasion. It was a time when mutual 25% tariffs had just been set, and their thinking was to destroy the old, flattened cooperative path by presenting anachronistic strong-arm tactics to gain new order and interests.
I told them about the "Manga-Mura" lawsuit, which once shook content holders all over Japan. In the damages lawsuit filed in 2022 and decided by the Tokyo District Court in April 2024, the three companies KADOKAWA, Shueisha, and Shogakukan were ordered to pay ¥1,736,642,277 to the former operator, Romi Hoshino. After reading the translated Wikipedia article, his only comment was, "So in the end, it was all settled with a damages payment." Moreover, his reaction was, "1.8 billion yen? That's a level we can solve with the funds we've raised." He added, "And even if we don't do it, other companies will. We've learned that Ghibli-fication is profitable." The reality is that it's not a fundamental "deterrent."
The sudden dismissal of Shira Perlmutter, the Director of the U.S. Copyright Office who had published research papers on copyright in the AI era, clearly illustrates this situation.
To use an analogy from "Princess Mononoke," I felt the presence of those who are seriously trying to "kill the gods" of copyright—people like Jiko-bō, the Jibashiri, and the Karakasaren, who possess the technology, funds, and organizational power to do so.
In the age of generative AI, merely expressing feelings or seeking legal solutions within Japan is not enough.
It may be that we have to restrain them through effective dialogue, specifically, through the price of litigation.
If, unbeknownst to me, a settlement has been reached between Studio Ghibli and ChatGPT, then that in itself is a "good story."
However, if this is not made open, more and more players will emerge to commit acts of piracy.
Pandora's box, its value now known, can no longer be closed.
Even within the same West Coast state of California, Los Angeles is different. As an economic sphere supported by film, music, art, and IP, it is extremely sensitive to style imitation and "knock-off" culture from generative AI.
LA, centered on Hollywood, is a region where protecting intellectual property (IP) as an "asset" is the key to survival, spanning the film industry, music industry, art galleries, and auctions. There is strong regulatory pressure from industry groups (like SAG-AFTRA for film actors and the RIAA for the recording industry) regarding AI synthesis of actors' voices, deepfakes in video, and AI music sampling, with some cases already in legal proceedings. This is a geopolitical situation that is the polar opposite of the "open-source culture" of things like Stable Diffusion.
The strictness of LA's IP culture is also reflected in the Digital Millennium Copyright Act (DMCA), enacted in 1998. Behind this bill's passage was powerful lobbying from the Walt Disney Company. The copyright protection period, which was supposed to expire after 70 years, was extended to such a degree that the law was derisively called the "Mickey Mouse Protection Act," becoming a symbol of corporate IP hoarding. The Copyright Term Extension Act of 1998 extended the copyright protection period for individuals from 50 to 70 years after death, and for corporate works from 75 to 95 years. This was ridiculed as being intended to extend the copyright of early Mickey Mouse works, typified by "Steamboat Willie" from 1928. In January 2024, the video portion of "Steamboat Willie" finally entered the public domain in the US. However, the design of the "modern Mickey," the theme park version, and trademark registrations are still protected. This could set a precedent for Japanese IP (Ghibli, Dragon Ball, Pokémon, etc.) in the future. Disney has also recently pivoted towards a more open community strategy in some respects. In fact, when I exhibited and sold generative AI art in collaboration with Japanese calligrapher GOYO at the Los Angeles art event "The Brewery Art Walk 2025," it was met with surprise and empathy.
Here, generative AI was evaluated not just as a tool, but as "context art" and "communication art."
Los Angeles Art Exhibition Day 1 - The day generative AI illustrations became communication art
https://note.com/o_ob/n/nac127fc7e03e
In other words, the current situation in America is a twisted phenomenon of technological freedom and cultural conservatism.
Meanwhile, walking through Little Tokyo in the same city of Los Angeles, you'll find the streets overflowing with knock-off T-shirts and pins, and cheap, unlicensed products are rampant. While Gashapon and figure culture are still alive, I hear comments like "official products are expensive." The Japanese IP industry is facing the dilemma of "only being able to produce expensive, adult-oriented figures."
This reality, combined with the "Ghibli-fication" by AI, suggests that the very "defensive strength" of Japanese copyright is being tested.
The current state of Japan's IP industry as seen through Little Tokyo 2025
https://note.com/o_ob/n/n3136914acc92
Unauthorized style learning by AI image generation, based on similarity and reliance, doesn't seem to be a simple case of "everyone's doing it, so it's fine." Here's a recent example I came across.
- Meme Poster
Prompt: 'Create a meme-style poster using IRASUTOYA illustration style. Theme: Monday morning mood, exhausted office worker.' pic.twitter.com/5vNsvNhGep
— Poonam Soni (@CodeByPoonam) May 12, 2025
Create a meme-style poster using IRASUTOYA illustration style. Theme: Monday morning mood, exhausted office worker.
This prompt generated an image like this:
If such images can be generated, it makes one feel that the efforts of AIPicasso Inc.'s "AI Irasutoya" and its operator, Takashi Mifune, were "all for nothing."
We work closely with creators. We have introduced a system for appropriate compensation for data provision and profit sharing for generated illustrations. One such initiative is AI Irasutoya. After about a year of consultation with Mr. Takashi Mifune, who runs Irasutoya, we were able to have him provide images and their descriptions for training. We strive to build such long-term relationships and earn the trust of creators.
In the world of generative AI, it's common for efforts to be trampled on by OpenAI. But if it were a matter of "standing on the shoulders of giants," like with open source or academic papers, or becoming nourishment like the fallen leaves in a forest, that would be one thing. But here, we're not just being beaten down by unauthorized scraping and an order-of-magnitude difference in computing infrastructure; we can see that "IRASUTOYA" is explicitly being used as a tag. This is not just a copyright issue. It should be possible to fight back using Japanese laws, such as the Act against Unjustifiable Premiums and Misleading Representations, for false advertising against AI style-learning services, where counterfeit works are passed off as genuine, infringing on trademark and identity rights.
There are many other trademarks, titles, and characters in Japan to defend, such as "Hatsune Miku," "Dragon Ball," "One Piece," "Gundam," and "Sailor Moon." Japan has numerous technological and legal foundations for defense. This is a field that AICU specializes in and one where we want to support the development of technology to compete on a global scale.
The copyright system doesn't exist just to "protect Japanese works in Japan." Rather, it is also a weapon to showcase Japan's creative and cultural power to the world and to engage in dialogue.
A look back at the history of ukiyo-e makes this crystal clear.
"Nishiki-e," which depicted the customs of the Edo period, achieved sophisticated color printing through a division of labor among the artist, the carver, and the printer. Their compositional power, color sense, and layout techniques had a tremendous influence on Impressionist and Naïve painters. Just as it was Alfred Jarry who recognized Picasso, it is always artists who recognize other artists.
And before many Japanese people realized their value, the works crossed the sea and were hung on the walls of the Museum of Fine Arts, Boston.
The Impressionist and Post-Impressionist painters who were influenced by and supported ukiyo-e include Claude Monet, Edgar Degas, Vincent van Gogh, and Henri de Toulouse-Lautrec. Monet, deeply influenced by the composition and color schemes of ukiyo-e (especially Hiroshige), had a Japanese garden and an ukiyo-e collection at his home. Van Gogh studied ukiyo-e and praised its innovation in his letters, promoting the idea of "Japonisme." Toulouse-Lautrec was directly influenced by ukiyo-e in his poster art and use of color planes, developing a style that connects to modern graphic design.
What we are facing now is a similar structure.
Those involved in creative fields understand its value.
That's why they study it and use it.
OpenAI has the financial capacity to respond to discussions on reliance and similarity, and the ability to maintain verifiable evidence. They pay the GPU costs, keep records, and generate with a design that ensures reproducibility.
That is precisely why we in Japan can assert our legitimacy on both "technological" and "cultural" fronts. The evidence is all there.
Manga, layout, color control with identity, speech bubbles—we, more than anyone, should know that all of these hold high value in the age of video generation AI.
From now on, Japanese IP and those who love it must continue to assert the following points:
・Fight with technology: Support the copyright protection mechanism itself with technology.
・Leave evidence: Clarify the accountability infrastructure for recording and proving what was used to train the AI and how it was generated. Document who first expressed an idea and its significance. If possible, do it in a way that can be understood by non-Japanese speakers.
・Establish region-specific rules: Flexibly build "local IP rules" like the EU and California to establish controllability and enable dialogue.
・Cooperate with pro-copyright allies.
・Propose new licensing systems: Standardize explicit permission systems that go beyond Creative Commons.
In conclusion, there is no illegality in the "style" that generative AI imitates. However, whether to "allow it or not" is left to the will of the stakeholders, and we have reached an era where the "unspoken norms" that have existed within Japan's doujin (fan-made) culture need to be redefined.
The current situation is probably one where the battle between Northern California, which is clearly trying to "kill the gods" of copyright, and the opposing forces in the south is "not clearly visible" to us looking on from across the sea via the internet.
But I have listened to claims like, "ChatGPT isn't 'out' just because it's similar; the fact that the rights holders are silent—that in itself is an answer," "Lots of people from Japan are using it (the reliance-based model) too," and "In the end, even if you stick to what's right, we're the ones who profit."
While listening to both sides impartially, how will Japan's copyright system and those who love Japanese works confront this reality? As a country, as artists, as citizens—we need to face the "future of culture" with intellect.
Ghibli producer Toshio Suzuki's offering of fair use images with the message "Please use them freely within the bounds of common sense" was a very clever method, but in the "expressing our feelings" phase of the AI era, the "bounds of common sense" have shifted.
新しく、スタジオジブリ5作品の場面写真を追加提供致します - スタジオジブリ｜STUDIO GHIBLI
https://www.ghibli.jp/info/013409/"Please feel free to use it within the bounds of common sense. Studio Ghibli" Toshio Suzuki (producer)
My own take, to put it in terms a child could understand, is:
・Who is being inconvenienced?
・We need to draw a clear line.
・We should propose a new copyright framework, not just copyright enforcement technology, but a solution for the AI era.
My company, AICU, a Japan-US AI media company with the vision of "creating the people who create," has updated its ethical guidelines, the "Generative AI Creator's Code of Conduct (v.2025)," in conjunction with the publication of our book, "Image and Video Generation AI ComfyUI Master Guide."
https://note.com/aicu/n/n6eb9982edb00
<iframe style="margin: 0px auto; display: block;" tabindex="-1" src="https://note.com/embed/notes/n6eb9982edb00" width="560" height="210" scrolling="no"></iframe>
Corporate Site: https://corp.aicu.ai/ja-jp/shigusa
Generative AI Creator's Code of Conduct (2025 Edition)
English Version: https://corp.aicu.ai/shigusa
Generative AI Creator's Code of Conduct (2025 Edition)
We cannot stand by and watch the "killing of the gods" of copyright. We intend to continue engaging with the world on this issue. Please send us your opinions. You can use the comment section of this article or contact us on X @AICUai.
This topic will also be covered in AICU Magazine Vol. 12 "AI Manga Revolution (tentative)" and Vol. 13 "The World of AI."
Thank you for reading to the end.
We look forward to your continued support of AICU.