Who Owns an AI Image? Two Courts Gave Opposite Answers to the Same Question

Two rulings on AI-generated images landed within weeks of each other, and read side by side they show how unsettled the answer still is. One court threw out a copyright claim because the picture was not a "work." The other upheld a copyright claim on an AI picture and awarded damages to the person who made it. The difference between them was not the tool. It was how much of a human was in the picture.

The question has been pushed into public view from an unexpected direction. On 27 September, an X user posting under the handle Mitssi__ argued that designers should simply claim AI-generated concepts as their own designs, on the reasoning that a machine cannot hold intellectual property and the output therefore belongs to whoever picks it up. The post encouraged appropriating AI art to test the limits of copyright. Legal observers pointed out that the argument gets the law half right and the consequences wrong. The U.S. Copyright Office does decline to register works that lack sufficient human authorship. That does not make such works free to pass off as your own, and misrepresenting the authorship of AI output can run into fraud or unfair competition rules.

So the field has two failure modes at once. Creators fear their work will be absorbed into a model without credit. And opportunists see AI output as unowned property waiting to be grabbed. Both worries trace back to the same unresolved line: what makes an AI-assisted image the product of a person rather than of a prompt.

The case that failed

In Changsha, a court in the Kaifu district heard a dispute that started with a sachet. A man identified as Shao used an AI model to generate an image on a Dragon Boat Festival theme, entering only general prompts about the subject and a Chinese style, setting the image ratio and resolution as basic parameters, and then using editing software to remove a duplicated, flawed line of text. He transferred the economic rights to another person, Ma, who registered the work. A kindergarten used the image in its WeChat account. Ma sued for infringement, asking for 10,000 yuan.

The court's reasoning is the part worth keeping. It applied a three-layer originality test and checked each stage of the creation. There was no initial human creative conception. In the generation-control stage, the prompts were generic descriptions of subject and style and did not specify composition, element design, or lighting; adjusting basic parameters was not creative work, since the image was produced mainly by the algorithm. In the post-processing stage, deleting a flawed line of text was a technical fix and did not change the core content, so it created no new originality. The court concluded the image lacked genuine human intellectual contribution, held that it was not a legally protected work, and dismissed the entire claim. The judgment has taken effect.

The message is narrow and clear. Typing a few words, nudging a slider, and tidying up after the model does not earn you a copyright.

The case that succeeded

A case in Qinhuangdao went the other way on a similar substrate. There, someone using an AI drawing model did not press a button and stop. They screened keywords repeatedly, adjusted technical parameters, and made choices about style until an image emerged with a particular composition, color and mood. The court found that the result carried independent intellectual input and individualized expression, and that it qualified for copyright protection. When a technology company then published that image in a commercial article on its WeChat account without permission, the court found an infringement of the right of information network dissemination and ordered compensation of 500 yuan, taking into account the level of originality, the method of distribution and the size of the business.

Five hundred yuan is not a large award. The precedent is the valuable part. It establishes that an AI image can be protected when a real person made the decisions that give it its character, and that using one in commercial content without permission carries a cost.

What actually decides the outcome

Line the two rulings up and the deciding factor is not whether AI was used. It is whether a human made specific, individualized choices, and whether there is evidence of them. Generic prompts describing a category lose. Deliberate choices about composition, element arrangement and lighting, made over iterations, win.

That has a practical consequence for anyone generating images professionally. The photographs of a workflow matter as much as the workflow. Prompt text, parameter changes, iteration history, saved versions. These are not paperwork. They are the evidence chain that turns an image from something the model produced into something you made. Creators who keep records have a case to bring. Creators who do not, do not.

There is an uncomfortable edge to this for the AI image industry, because a great deal of the marketing around image models promises the opposite. The pitch is that a single prompt produces a finished asset. The law, at least in these two Chinese rulings and in the U.S. registration practice, says that the single prompt is exactly the part that earns no protection. The more the tools automate, the less of the output is clearly the user's.

Why this is genuinely hard, not just unsettled

The difficulty is not that the courts are inconsistent by accident. Copyright was built for a world where making an image required a person to make thousands of small decisions. Every one of those decisions was potential evidence of authorship. AI compresses those decisions into a prompt, and the compression is the whole value of the tool. So the law is being asked to protect human authorship in a workflow whose selling point is that it removes the human authorship.

That is why the two rulings feel contradictory without being wrong. Both are applying the same test, and they land on opposite sides because the human input in each case was genuinely different. The test is stable. What is moving is where the line falls in practice, and it will keep moving as models absorb more of the decision-making that used to belong to the person holding the mouse.

There is a practical version of this for ordinary users, separate from the professional case. Someone who generates an image for a social post, likes it, and reuses it has almost certainly made no protectable contribution and has little to worry about on either side. Someone building a brand around generated assets is in the professional case whether they realize it or not, and should behave accordingly. The distinction is not about how good the image is. It is about whether a business depends on it.

The fight on both sides

Put the courts and the viral post together and the shape of the next few years shows up. On one side, human creators are pushing for recognition and protection when their work feeds a model or is imitated. On the other, people are testing whether AI output can be taken freely because it is unowned. Both pressures are aimed at the same gap, and the gap is widening as generation gets cheaper.

The U.S. Copyright Office position and these Chinese rulings point in a broadly similar direction, which is that human authorship is the dividing line and the amount of it is what gets litigated. What is still missing is consensus on how much is enough, and how you prove it. Until that settles, the safest approach for anyone using AI images in commercial work is to keep more of the human in the loop than the tools encourage, and to keep the receipts for every choice.