On March 25, 2025, OpenAI announced that image generation was now built natively into GPT-4o, rather than being handled by a separate model like DALL-E 3. Because the same model that produces text and code now also produced images, it could follow detailed instructions, render readable text inside images, take an uploaded picture and transform it, and lean on the model’s broader knowledge to make pictures that “feel smarter.” OpenAI rolled it out to Plus, Pro, Team, and Free users in ChatGPT.
Almost immediately the internet found its favorite use: feeding in personal photos and asking for them back in the soft, hand-painted style of Studio Ghibli, Hayao Miyazaki’s famed animation house. Ghibli-fied selfies, pets, and memes flooded social feeds. Demand was so intense that Sam Altman publicly said the company’s GPUs were “melting” and OpenAI temporarily added rate limits to keep the system standing up. (Note: Altman’s running commentary appeared largely on X, which is paywalled to automated fetchers; this entry relies on OpenAI’s official launch post for verifiable claims and treats the GPU-demand detail as widely reported context rather than a quoted primary source.)
The wave was a genuine cultural moment and a genuine controversy at once. To many users it was delightful and harmless fun. To many artists it was the opposite: a frontier lab letting anyone reproduce the distinctive look of a living studio on demand raised pointed questions about whose work the model had learned from, whether mimicking a recognizable house style is homage or appropriation, and where the line between “inspired by” and “copied from” actually sits. Miyazaki’s own past remark that he was “utterly disgusted” by an AI animation demo recirculated as a counterpoint to the trend.
For business readers, the episode compressed three lessons into a week. First, distribution plus a fun default behavior can create demand that outstrips even OpenAI’s compute. Second, “style” is now trivially transferable, which is thrilling for marketers and threatening for the artists and studios whose styles are being transferred. Third, the copyright and consent questions that hang over generative media (see training-data) do not stay abstract; they arrive the moment a tool this capable meets a culture this online.