Studio Ghibli-Style AI Images: Copyright, Lawsuits and the Rules in 2026

Making a “Studio Ghibli-style” image with AI isn’t illegal in itself: copyright protects works and characters, not a style. The risk starts when an image copies a character (Totoro, Chihiro), a specific scene, or suggests a link to the studio, especially in commercial use. The bigger question, whether training AI models on copyrighted works without permission is lawful, is still open in 2026: no U.S. court has ruled on it for image generators.

We first published this story in April 2025, in the middle of the Ghibli-style image craze, and rewrote it on Sept. 25, 2026. It covers what OpenAI, Studio Ghibli and the Japanese government have said since, the lawsuits still pending, and the U.S. and EU rules, based on the court filings, statutes and official statements linked throughout. It isn’t legal advice: for a commercial project, talk to a lawyer.

What Has Changed Since March 2025

Sources for each row are linked in the sections below.

March 2025: Where the Ghibli-Style Trend Came From

On March 25, 2025, OpenAI replaced DALL·E with GPT-4o’s native image generation in ChatGPT. Within days, social feeds filled up with selfies, news photos and memes redrawn to look like Hayao Miyazaki’s films. OpenAI CEO Sam Altman posted that “our GPUs are melting” and announced temporary rate limits, including three generations a day for free users (Fortune, Mar. 28, 2025).

OpenAI’s rule at the time, spelled out in its GPT-4o image generation system card (Mar. 25, 2025): a refusal triggers when a user asks for an image “in the style of a living artist.” The company told reporters that “broader studio styles” such as Ghibli or Pixar were still allowed (TechCrunch, Mar. 28, 2025). Hence the paradox critics pointed out: Miyazaki is very much alive, but his studio doesn’t count as an “artist.”

Two facts that often get garbled:

• Miyazaki’s “insult to life itself” line dates from 2016. He was reacting to a machine-learning 3D animation demo (a crawling, zombie-like figure) shown by Dwango’s lab in an NHK documentary, not to image generators, which didn’t exist in their current form (Anime News Network, Dec. 9, 2016).
• The “cease and desist” letter credited to Studio Ghibli that went viral in late March 2025 was fake: the studio told NHK it hadn’t issued any warning (Screen Rant, Mar. 28, 2025).

To make images in ChatGPT today, see our ChatGPT Images tutorial (formerly DALL·E 3).

What Has OpenAI Changed Since?

OpenAI hasn’t restated its March 2025 style rule in any newer document: the ChatGPT Images 2.5 system card, published Sept. 8, 2026, doesn’t address artists’ styles. Its usage policies (effective Oct. 29, 2025, checked Sept. 25, 2026) do ban attempts to infringe other people’s intellectual property. The notable changes came on the video side:

• Sora 2 and characters: Sora 2 launched on Sept. 30, 2025 with an opt-out approach, where rightsholders had to ask to be excluded. Days later, Altman promised rightsholders “more granular control over generation of characters,” similar to an opt-in model, “including not at all” (Sam Altman’s blog, October 2025).
• The Disney deal: announced in December 2025, it would have brought more than 200 Disney, Marvel, Pixar and Star Wars characters to Sora, with Disney investing $1 billion in OpenAI (OpenAI announcement). Disney walked away when OpenAI announced Sora’s shutdown on Mar. 24, 2026 (Variety).
• Sora’s shutdown: the app closed on Apr. 26, 2026 and the API on Sept. 24, 2026 (OpenAI Help Center, checked Sept. 25, 2026). ChatGPT no longer makes video: see our ChatGPT guide.

In practice, nothing in OpenAI’s public documents guarantees that a Ghibli-style request will be accepted or refused: it comes down to whatever moderation applies when you ask.

Studio Ghibli and Japan: What They’ve Said

As of Sept. 25, 2026, Studio Ghibli hasn’t filed any publicly reported lawsuit against OpenAI. The studio’s position has come through three channels:

• Through CODA: Studio Ghibli is on the member list of CODA, Japan’s Content Overseas Distribution Association (anime, games, music, publishing). On Oct. 27, 2025, CODA asked OpenAI not to use its members’ content to train Sora 2 without permission and to respond to infringement claims over Sora 2 output. It also rejected the opt-out approach, arguing that Japanese law generally requires permission up front rather than objections after the fact (CODA statement). After Sora’s shutdown was announced, CODA called it a “milestone” and said it would investigate other generative AI services in fiscal 2026 (CODA, Mar. 27, 2026).
• Through the Japanese government: in October 2025, Minoru Kiuchi, the minister in charge of intellectual property strategy, said the government had formally asked OpenAI not to engage in actions that could constitute copyright infringement, calling anime and manga “irreplaceable treasures” (The Register, Oct. 15, 2025).
• Through Toshio Suzuki: in an interview with the Asahi Shimbun published in early September 2026, the studio’s co-founder and producer said he turned down interviews during the 2025 craze because he expected people to get bored quickly. Asked whether AI can imitate a Ghibli work, he said no, and added that Miyazaki doesn’t even know generative AI exists. He didn’t mention lawsuits or copyright (Anime News Network, Sept. 8, 2026).

Japanese law is fairly permissive about training: Article 30-4 of its Copyright Act allows analyzing works without permission when the purpose isn’t to “enjoy” them. But Japan’s Agency for Cultural Affairs drew a line in 2024 that speaks directly to the Ghibli case: imitating a creator’s style isn’t infringement as long as the style is just an idea, yet deliberately fine-tuning a model on a small set of one creator’s works to reproduce their shared creative expression falls outside the exception (Agency for Cultural Affairs, May 2024, PDF).

Is the Ghibli Style Protected by Copyright?

Generally, no. In the U.S., as in Europe and Japan, copyright protects the specific expression in a work, not ideas, techniques or a way of drawing. The U.S. Copyright Office says as much in its 2025 report, describing style as something “copyright does not protect,” while noting that some style imitations can capture protectable elements of an original work (Generative AI Training report, pp. 65-66).

That doesn’t make everything fair game. What separates a risky image from a safe one:

Trademark law adds a second layer for commercial use. The Lanham Act bars misleading claims about a product’s origin or endorsement; that’s one of the grounds in the artists’ lawsuit against Midjourney (see below).

Which Lawsuits Are Pending Over AI Images?

None involves Ghibli directly, but together they’ll shape the rules for image generators:

The studios’ case against Midjourney is the closest to the Ghibli situation: it’s about characters generated on request, not style. For what Midjourney’s terms say about your rights, see our Midjourney guide.

Fair Use: What U.S. Courts and the Copyright Office Say

In the U.S., training a model on copyrighted works is judged under fair use, a case-by-case defense. The first rulings involve books, not images:

• Bartz v. Anthropic (June 23, 2025): Judge William Alsup held that training on books that were bought and scanned was fair use, but building a library of pirated books was not. The case ended in a $1.5 billion settlement that received final approval on July 20, 2026 (Authors Guild).
• Kadrey v. Meta (June 25, 2025): Judge Vince Chhabria ruled for Meta because the authors hadn’t made the right case, while warning that flooding the market with competing content (“market dilution”) could tip other cases the other way (ruling).
• Thomson Reuters v. Ross: in February 2025, a judge rejected fair use for a non-generative legal research tool. The Third Circuit heard arguments on June 11, 2026; its ruling, the first from a federal appeals court on AI training, is still pending (LawSites).
• The government weighs in: on Sept. 1, 2026, the Justice Department filed a statement of interest in the consolidated copyright cases against OpenAI (including The New York Times’ suit), arguing that training AI models on copyrighted material “in and of itself” doesn’t violate copyright law. The filing isn’t binding on the judge (Sterne Kessler).

The Copyright Office released Part 3 of its AI report, on training, on May 9, 2025. It’s still a pre-publication version as of Sept. 25, 2026, according to its AI page. The Office concluded that commercial use of vast troves of works to produce content that competes with them, especially through illegal access, goes beyond established fair use boundaries, and that style imitation can hurt a creator’s market. It cites the Ghibli-style wave as an example that could undermine the studio’s licensing opportunities (report, pp. 65-66).

One more point that matters to users: an image made entirely by AI can’t be copyrighted in the U.S. The Copyright Office concluded in January 2025 that prompts alone don’t make you the author (Copyrightability report), and on Mar. 2, 2026 the Supreme Court declined to revisit the human authorship requirement (Thaler v. Perlmutter, SCOTUSblog).

EU Rules: Text and Data Mining, Opt-Outs and the AI Act

In the EU, training falls under the text and data mining exception in Article 4 of Directive 2019/790. It allows copying lawfully accessible works unless the rightsholder has reserved their rights “in an appropriate manner,” using machine-readable means for online content. A German appeals court applied this in the LAION case on Dec. 10, 2025: a photographer’s plain-language reservation in a website’s terms wasn’t valid because it couldn’t be interpreted by machines; a further appeal is still possible (DLA Piper analysis).

Article 53 of the AI Act adds two obligations for providers of general-purpose AI models, such as OpenAI’s or Google’s:

• A copyright compliance policy, which must identify and honor rightsholders’ opt-outs.
• A public summary of training content, following the Commission’s template of July 24, 2025: main datasets and top scraped domain names.

These obligations have applied since Aug. 2, 2025; the European Commission gained enforcement powers on Aug. 2, 2026, and models released before August 2025 have until Aug. 2, 2027 to comply (European Commission, checked Sept. 25, 2026). The July 10, 2025 Code of Practice spells out the copyright part (honoring robots.txt, excluding piracy sites, limiting infringing outputs, a contact point for rightsholders); OpenAI and Google are among the signatories (checked Sept. 25, 2026). Our AI regulation overview covers the rest of the law.

Two more European decisions to watch. In Germany, the Munich Regional Court ruled on Nov. 11, 2025 that ChatGPT’s memorization of song lyrics was infringing reproduction (Bird & Bird). And the EU’s top court is hearing its first generative AI case (Like Company v. Google, C-250/25): it will decide whether training counts as reproduction and whether the data mining exception covers it.

Using Ghibli-Style Images: Practical Precautions

This summarizes the sources above; it isn’t legal advice for your situation.

1. Personal use (profile picture, sharing with friends): low risk as long as no protected character appears.
2. Commercial use (merch, ads, book covers): no recognizable characters or settings, and no “Ghibli,” “Miyazaki” or film titles in product names or marketing. Describing the look rather than naming the studio also lowers the risk of implying an endorsement.
3. Ownership: OpenAI’s terms assign you the images you generate, but that says nothing about whether they’re protected. A purely AI-generated image can’t be copyrighted in the U.S., so others may be free to reuse it; details in our ChatGPT Images tutorial.
4. No targeted training: fine-tuning a model (LoRA or otherwise) on frames taken from Ghibli films is exactly what Japan’s Agency for Cultural Affairs places outside its exception, and it means copying protected works.
5. Transparency: ChatGPT images carry C2PA metadata and an invisible watermark; don’t present them as hand-drawn or as official artwork.

What About the Ethics?

The debate is less about legality than about the value of artists’ work. Two documented positions face off. On one side, creators and their organizations, whose concerns the Copyright Office echoes: mass-produced imitations compete directly with the artists whose work trained the models, without permission or pay. On the other, AI companies, and now the U.S. government, argue that learning from existing works is a legitimate transformation that innovation depends on.

The Ghibli case adds a twist: the look being imitated belongs to a studio famous for hand-drawn animation, whose leadership says AI can’t reproduce its work. Using the style is allowed; whether doing it at scale is a good idea is a personal call.

FAQ

Is it legal to make Ghibli-style images with AI?

Generally, yes: a style isn’t protected by copyright. An image becomes risky when it copies a character or a scene from a film, or when it’s sold in a way that suggests a connection to Studio Ghibli.

Is Studio Ghibli suing OpenAI?

No lawsuit had been made public as of Sept. 25, 2026. The studio disowned a fake cease-and-desist letter in March 2025, and it was CODA, a trade group it belongs to, that asked OpenAI in October 2025 to stop training Sora 2 on its members’ content.

Can I sell Ghibli-style images made with ChatGPT?

OpenAI’s terms assign you the images you generate and don’t bar commercial use. You’re still responsible for third-party rights: no characters, film titles or studio name, and no misleading presentation. For a significant commercial project, ask a lawyer.

Does ChatGPT still allow the Ghibli style?

In March 2025, OpenAI allowed studio styles and blocked the styles of living artists. No newer document clarifies the rule, and the Images 2.5 system card (September 2026) doesn’t mention it, so results depend on the moderation in place when you ask.

Can an artist’s style be protected?

Not by copyright as such. An artist can still act against copies of their works, against imitations that create confusion (false endorsement or trade dress claims under the Lanham Act) and, in the EU, opt out of AI training with a machine-readable reservation.

Does the EU AI Act ban training AI on copyrighted works?

No. It requires providers of general-purpose AI models to honor rightsholders’ opt-outs and publish a summary of their training data. The European Commission has been able to enforce these obligations since Aug. 2, 2026.