Tuesday, 25 March 2025. OpenAI quietly ships an upgraded image generator inside GPT-4o — sharper text, better lighting, hands that finally have the right number of fingers. Nobody is talking about the hands.
By the weekend, a third of the posts on X are somebody's dog, somebody's wedding, somebody's passport photo, repainted with the same wide eyes and hand-brushed sky Studio Ghibli has used since My Neighbor Totoro. Sam Altman changes his own profile picture to match and asks, half-joking, for someone to make him a better one.
By Thursday he is posting that the company's GPUs are "melting" and capping free users at three images a day. He is only half-joking about that too. In Tokyo, the studio whose forty-year house style just got cloned by a few hundred million strangers in a single weekend still has not said a word.
Nobody at OpenAI built a Ghibli button. The company's own technical notes describe a "conservative approach" to style mimicry and a refusal that triggers if a user names a living artist by name — but "Studio Ghibli" is a company, not a person, and the model had clearly seen enough of its forty-year catalogue to reproduce the brushwork on request. The trend needed no filter. It needed a prompt box, a famous enough aesthetic, and a weekend. By the following Thursday the White House's own X account had posted a Ghibli-style image of an immigration arrest, which is roughly the moment a meme about cute watercolor skies turned into a genuine argument about what the style was being used to launder.
The backlash reached for one piece of footage with uncanny speed: a 2016 NHK documentary called The Never-Ending Man, in which Dwango AI Lab researcher Nobuo Kawakami showed Hayao Miyazaki an AI system that had taught itself to generate unnatural, boneless human movement — pitched, Kawakami explained, as something that might one day animate zombies in a horror game. Miyazaki watched in silence, told the room about a friend of his with a disability, said whoever built this had no idea what real pain looked like, and called it an insult to life itself. Kawakami pointed out it was only ever an experiment. Miyazaki said he understood that, and that he never wanted the technology anywhere near his own work. Nine years later, the clip did more work explaining the internet's unease than any article written that week.
Here is what almost none of that outrage mentioned: Japan had already answered the legal question, in Japan's own favor, seven years before anyone asked it about Ghibli. A 2018 amendment to the Copyright Act, Article 30-4, lets anyone — a researcher, a company, an AI lab — copy and analyze copyrighted works, including art, for machine learning, without the rights holder's permission, provided the point of the copying isn't to let a person directly enjoy the original work's expression. It is one of the most permissive AI-training exceptions written into any country's law. The Agency for Cultural Affairs revisited it in March 2024, kept the exception standing, and added only a warning against training on material pulled from pirate sites. Nothing in that guidance was written with Studio Ghibli in mind. It just happened to be sitting there, already law, the day the rest of the world needed to ask Japan a question Japan had answered in 2018.
So the silence from Ghibli's side was never really a choice between dignity and outrage. Toho and Studio Ghibli, as of this writing, have filed nothing and said nothing on the record, and under their own country's copyright law a model trained on their work — rather than a model asked to directly output a copy of a specific film — sits in genuinely contested, not clearly illegal, territory. American outlets framed the quiet as very Japanese restraint. It may equally have been a studio's lawyers reading the same statute everyone else eventually found.
Scale is the part that gets lost fastest in a trend this fast. The Ghibli Museum's own exhibition catalogue for Spirited Away puts the hand-drawn material behind that single film at roughly 250,000 pages of original drawings, produced over something like a year and a half to two years — and notes that one four-second kitchen scene alone needed more than a hundred separate drawings to animate a few seconds of steam and chopping. The internet produced a comparable volume of finished-looking images, about that exact style, in considerably less than a week.
Everyone assumed the fight, if it came, would be Japan versus the machine — the country that drew the pictures standing up to the country that scraped them. It is closer to the opposite. The nation most associated with hand-drawn animation is also the nation whose government wrote one of the most permissive AI-training laws on earth, in 2018, to make its own AI industry competitive — years before anyone imagined that exact law would end up shielding a weekend meme about its most famous export. Japan didn't lose this fight to Silicon Valley. Japan's own Diet had already decided how it would go.
- OpenAI's upgraded GPT-4o image generator shipped on 25 March 2025; within two days the Ghibli-style trend was global, and by 27 March Sam Altman was posting that the company's GPUs were "melting" while announcing a cap of three free-tier image generations per day.
- Miyazaki's "insult to life itself" remark is from the 2016 NHK documentary The Never-Ending Man, responding to an AI-generated movement demo built by Dwango AI Lab's Nobuo Kawakami, who said it had been pitched as a possible animation tool for zombie characters in a game — an early experiment, not a finished product.
- Japan's Copyright Act, Article 30-4 (added in a 2018 amendment), permits copying copyrighted works for AI training and data analysis without the rights holder's consent, as long as the use is not meant to let a person directly enjoy the work's original expression; the Agency for Cultural Affairs reaffirmed the exception in March 2024 while warning against training on pirated source material.