Does Copying Someone Else's AI Prompt Count as Infringement? A Ruling Keeps Prompts Outside Copyright

A company employee typed a passage into an input box, setting the art style, subject elements, and material details, pressed generate, and got a rather nice-looking dreamy illustration, then posted it to an online platform. A few months later, paintings published by Zhu and Sheng appeared online; their visual style was highly similar to this work, and they were even included in an art catalog. The key point was that the prompts the two used were exactly identical to the original.

The company believed this was plagiarism and sued the two, demanding that they stop the infringement and compensate for losses. The court did not support the claim.

The Xi'an Xincheng District People's Court recreated this case in an illustrated legal education comic and gave the reasons for its ruling. I'll break it down, because the issue this ruling touches on has actually been encountered by everyone who has used an image-generation model: does that prompt I spent so much effort writing count as my work?

Why the Court Said the Prompt Was Not a 'Work'

The object protected by copyright law is called a 'work.' To qualify, it must satisfy two conditions at the same time: it is independently created and reflects a personalized intellectual investment. In other words, it protects an expression with originality, not the idea hidden behind the expression.

The court examined this at three levels.

From the perspective of content, the prompt at issue was merely a set of abstract creative ideas and instructions; its core was listing image elements, art style, and presentation form. These belong to the conception stage and are 'ideas.'

From the perspective of form, it was merely a simple pile-up of various elements, without forming a structured expression involving selection and omission.

From the perspective of originality, the art style, material details, and so on selected in it all belonged to conventional expressions in this field, and did not reflect the author's unique aesthetic perspective or artistic judgment.

With the three levels combined, the conclusion is clear: although this prompt reflected creative intent, it did not reflect the author's personalized intellectual investment at the level of expression, so it did not constitute a work. The plaintiff did not enjoy copyright over it, so naturally there could be no infringement. The court dismissed all claims at first instance.

There is another passage in the ruling worth highlighting separately. The court noted that if this kind of simple keyword combination were recognized as a work, it could restrict the free use of language and also hinder the AI innovation ecosystem, running counter to the legislative purpose of copyright law, which is to encourage creation and promote social and cultural development. That sentence carries no little weight; it is effectively saying: granting copyright to prompts does more harm than good.

Prompts Are Not Protected, But That Does Not Mean Generated Images Are Not Protected

This is where misunderstanding is most likely. A prompt having no copyright does not mean that the image produced from the prompt has no copyright.

Two levels need to be distinguished.

The prompt is a 'recipe'; it is process-based. It describes how to do something, not what was made. A recipe itself is usually not protected by copyright, unless it has become so complex that it resembles an article.

The generated image is the 'result,' the final picture presented. If the painting reflects sufficient human choices and aesthetic judgment in composition, color, light and shadow, and details, it has a chance to be recognized as a work of fine art. The key to the judgment is how many decisions bearing a personal imprint a person invested in the image; it has little to do with how long the prompt string is.

This boundary has become increasingly clear in recent domestic rulings. On September 7, the Supreme People's Court issued the 'Opinions on the Lawful Handling of Cases Involving Artificial Intelligence'—the first domestic rule-making document on AI-related adjudication issued by the highest judicial body—for the first time clarifying that using a natural person's voice as training corpus without consent and imitating their timbre and tone to generate a recognizable synthesized human voice constitutes an infringement of voice rights and interests. At the end of September, the second-instance ruling in Shanghai's first AI-synthesized speech infringement case was handed down, and the platform operator was ordered to pay 50,000 yuan; a judicial appraisal showed that among 28 formant indicators, 24 of the audio involved had a deviation of less than 10% from the dubbing actor's own audio.

Also involving AI generation, the AI micro-drama copyright case in Hubei took a different path. A 47-episode AI micro-drama with a total runtime of about one hour was copied in its entirety. In finding that it constituted an audiovisual work, the court focused on the creative process: script planning, storyboard prompt setup, screening and selection of AI-generated clips, and matching of video editing with dubbing and subtitles. The creator made continuous, substantial intellectual investments and had full foreseeability and substantive control over the final film. On the amount of damages, the court also took into account the computing power costs consumed by tokens and the licensing costs of commercial tools, ultimately awarding 20,000 yuan.

Putting the three cases together, one can see the court repeatedly doing one thing: it does not look at the tool; it looks at the person. The more, more specific, and more foreseeable the human judgment invested, the more likely the output is to be protected.

So What Else Can Be Claimed Besides Copyright

Creators may feel a bit disappointed that prompts cannot obtain copyright protection. But copyright is not the only protection tool; there are actually several other paths.

One path is to focus on the result. Rather than agonizing over who owns that piece of text, it is better to spend effort on the generated image. Multiple rounds of image revision, local repainting, adjusting composition and lighting—these actions will genuinely increase the originality of the generated image and make it easier to show, when a dispute arises, that 'I made this image.'

Another path is contracts. Internally within a company, in outsourcing collaboration, and in client delivery, written agreements can clearly set out the ownership and scope of use of prompts, workflows, materials, and generated images. Problems that can be solved by agreement need not all be left to the courts.

Another path is trade secrets. For prompts to truly constitute a competitive advantage, it is often a whole workflow: model version, parameters, reference images, and the sequence of manual retouching. If this set of things is not public, has commercial value, and has confidentiality measures taken, it may fall within the scope of trade secret protection, and the evidentiary approach in litigation would be entirely different.

There is also technical traceability. Adding labels to generative content, recording generation time and model version, and preserving iteration records may in the short term look like mere process standards, but when a dispute actually arises, they become a chain of evidence proving one's investment.

Final Thoughts

The signal sent by this ruling is not pessimistic. It merely puts something overvalued back in its place: writing prompts is technical work and creative work, but it is more like a recipe than a painting.

What is truly worth spending time on is the dish the recipe produces, and the entire set of practices built around that dish. Anyone can copy a recipe, but the craft and process of cooking are not so easy to copy.