A Wuhan Court Priced AI Compute Into a Copyright Ruling, and Set a New Kind of Precedent

A court in Wuhan has ordered a defendant to pay 20,000 RMB, roughly 2,900 US dollars, for copying an AI-generated short drama. The amount is small. The method behind it is not.

For the first time in a Chinese copyright case, the bench factored the plaintiff's compute spending into the damages calculation. Token consumption and API licensing fees were treated as concrete financial harm, alongside the traditional measures of runtime, distribution reach and how long the infringement lasted. The court leaned on operational records and software licences as baseline evidence rather than treating them as buried overhead.

That changes what an AI production company has to keep, and it changes what a legal claim looks like when somebody steals the output.

The facts were not subtle

The original work was a one-hour short drama produced in early 2026 with a combination of AI tools and published across platforms including WeChat. Within twenty-four hours of release, a competitor had copied the whole production, given it a new title, and started monetising it with in-stream advertising.

The court had little patience for the copying. Before it could award anything, it had to answer a prior question: is an AI-assisted short drama a protectable work at all?

It decided yes, classifying the drama as a fully protectable audiovisual work. The reasoning turned on human authorship at every stage. The bench found that people wrote the script, engineered the prompts, selected the outputs, and assembled the final cut. The model generated frames. Humans decided which frames existed.

Why the compute number matters

Once the work is protectable, the question becomes how to measure the loss. Copyright damages conventionally track the value of what was taken. For a film, that means production cost, market reach, licensing revenue that was displaced.

AI changes what "production cost" contains. A traditional short drama has a cost structure that a court can audit: cast, crew, locations, equipment, post-production. An AI production has prompts, iterations, and a long tail of failed generations that were paid for and thrown away. The spend that produced the work is spread across API invoices, GPU rental, and software subscriptions, and none of it appears in a budget line that looks like a film crew.

The Wuhan court treated those bills as evidence of what it cost to make the thing. That is a defensible move and a consequential one. It means the money a studio burns on bad generations is recoverable in principle, because it is part of the investment the infringer took for free.

It also creates a new evidentiary discipline. The court explicitly advised creators to retain project logs, raw scripts and prompt drafts. An AI studio that cannot show the iteration history has no way to prove what the work cost. The infrastructure bill just became the backbone of the case.

The counterpoint from Xi'an

The ruling lands in the same week as a decision pointing the other way on a related question.

A court in Xi'an's Xincheng district heard a case about whether copying somebody else's AI prompts infringes copyright, and concluded that a prompt is not a work. The reasoning is straightforward: a prompt is an instruction, and instructions are not the expressive content copyright protects. What matters is the output, and the output here was generated by a model rather than written by a person.

Read together, the two decisions outline a division that is starting to settle. The prompt is not the protected thing. The curated output is, and the cost of producing it is measurable.

That division is friendlier to AI studios than it first appears. It removes a protection they might have wanted, over their prompts, and hands them a stronger one over their finished work and the money they spent making it.

What a studio would have to keep

If compute expenses can become damages, then the record a studio maintains is no longer an internal hygiene issue.

The court's advice was specific: project logs, raw scripts and prompt drafts. In practice that means an AI production needs to be able to reconstruct, after the fact, what it generated, in what order, at what cost, and which outputs were kept.

Most generative pipelines do not preserve any of this. A prompt typed into a web interface disappears when the tab closes. A batch run that produced a hundred candidate clips leaves behind the files that were saved and nothing about the ninety-nine that were not. The cost of the discarded generations is real, and it is exactly the part that is hardest to document.

Building that record is unglamorous engineering. It means routing generations through an account that invoices, logging prompts and responses in a form that survives, and keeping the receipts. The payoff is that a copied work becomes a claim with a number attached rather than a complaint about unfairness.

There is also a signalling effect. A studio that can show its logs is a harder target, because the cost of infringement is legible before anyone goes to court.

What this does to the economics of AI content

The immediate effect is a rise in the value of record-keeping. Studios that treat their generation logs as disposable are accumulating a liability they cannot see. Studios that keep them are building an asset that shows up in court.

The second effect is subtler. Compute costs are variable and, for video generation, large. A one-hour AI short drama represents an enormous number of generation attempts, most of them discarded. Pricing that into damages means the plaintiff's claim scales with the effort they put in, which rewards deliberate production over lucky prompts.

Third, it gives AI studios a reason to formalise what they spend. If API invoices are recoverable evidence, then a company that runs its generations through a personal account on a consumer plan has a weaker claim than one that invoices through a business entity. The legal posture follows the accounting.

What is still unresolved

The 20,000 RMB figure is modest, and the reasoning has not been tested on appeal. A higher court could narrow the compute-cost element, particularly if it decides that spending money on failed attempts is not a loss the infringer caused.

There is also the question of apportionment. If a studio spends on compute for ten projects and one is copied, how much of the invoice belongs to the stolen work? The Wuhan court did not have to solve that, because the drama in question was a single production. A larger studio with a shared generation budget will have to.

And the precedent is domestic. Chinese courts have been more willing than most to find copyright in machine-assisted output, going back to a 2023 Beijing ruling that recognised copyright in AI-generated images. A US court facing the same facts would likely split authorship differently, and the compute invoice argument would have no obvious hook in a system that has not decided the underlying question.

Still, the direction is consistent with a pattern in Chinese AI jurisprudence over the past two years: decide the narrow case, keep the human in the authorship chain, and let the operational reality of running models become part of what the law recognises. That last part is the new development. The court asked who made the work, then asked what it cost to make, and accepted the model's meter as the answer.