A model can write and draw because vast amounts of material were fed into it earlier. But ask exactly what went in and where it came from, and often few can say. Most of the time that is not a problem; once a dispute arises, being unable to say becomes a hurdle that cannot be stepped around.
What Goes Into the Model
Think of training as preparing ingredients and it is easy to follow: text, images, recordings and code are gathered into a batch of samples and fed into the model bit by bit, and only then does it learn what it can later do. Where the material came from and how it was collected stay behind the scenes, out of sight.
The Opinion on Properly Trying AI-Related Dispute Cases issued by the Supreme People's Court addresses exactly this. The document takes cases where generated content infringes copyright and separates out first which factors to consider, then divides responsibility among the parties accordingly.
The Opinion states that the people's court shall weigh the type of service, industry characteristics, sources of training data, each party's degree of involvement, necessary measures taken and profits, and shall reasonably determine the liability of AI developers, providers and users in accordance with the law.
In plain terms, when something goes wrong, a single line about it being machine-generated does not push it away; the chain has to be taken apart to see who did what. Who collected the material, who delivered the service and who used the content each correspond to their own set of books.
For the side developing the model, one concrete matter is keeping those books. How the model was trained and on what is invisible day to day, yet has to be laid open once a dispute starts, which pushes backstage work into view.
Developers, Providers and Users
The Opinion divides the people on the chain into three kinds: developers, providers and users. The names sound tangled, but the division is clear: one end concerns how the model is built, the other what the content is used for, and the middle is the link that hands the service to people.
Developers handle how the model is built and what it trains on; providers handle how the service reaches people; users handle what they do with it. Along the same generation chain, the three roles stand in different places and hold different things in their hands.
In its general requirements the Opinion also states that distinctions are drawn between general-purpose and specialized, open-source and closed-source models in technical principles, risk spillover and control capacity, and that legal liability is reasonably divided among developers, providers, users and other parties.
Why split it so finely? Because the evidence sits with different people. Training data is with the developer, prompts are with the user, and the entry point for generation is with the platform; whoever holds what is where liability follows down the line.
Next time you read news about generated content infringing something, it is worth asking first which link went wrong. Was it material collected during training that should not have been, or someone who knew the original work and generated a close copy anyway; different positions call for different accounts.
Pleading No Infringement Means Showing the Books
Another key line sits in the same article: where an AI developer pleads no infringement, that developer shall be ordered to provide the sources of training data, training process records, model operating mode and scientific theoretical basis as supporting evidence.
In plain terms, to say you did not copy, you have to lay open what you trained on and how. A bare conclusion is not enough; records from the process have to be produced so the other side and the court can follow the trail back.
The other way round, where a rights holder claims a provider used algorithmic techniques to infringe copyright, that holder shall provide relevant evidence; it is not enough to say so in the air. The Opinion stops at that requirement and does not spell out a list of documents; in a given case, the court weighs the type of service, the industry, how each party took part and other factors.
That is exactly what the title means: what fed the model has to be spelled out. The point is not who argues loudest, but whether the training stage can be laid open, with sources, process and basis set on the table one by one.
Among the materials the document lists, sources of training data are a checklist, training process records are things accumulated step by step over time, and model operating mode and scientific theoretical basis serve to show that the approach holds together.
Where Users and Platforms Stand
On the user side, the Opinion states that where an AI user knew or should have known an earlier work existed and used AI to generate a work substantially similar to it without reasonable grounds for defense, and the earlier work's rights holder asks that user to bear infringement liability, the people's court shall support the claim.
In plain terms, someone who knows an original exists and still generates a close copy cannot use a line about a machine having made it as a shield. The tool only produces the result; whether the person knew or should have known still rests with the one at the keyboard.
On the platform side, the Opinion states that where generated content infringes personality rights such as reputation or privacy, and the provider fails to take necessary measures such as stopping generation of infringing content in time after being notified by the rights holder, that provider shall bear infringement liability for the resulting harm in accordance with the law.
A notice shall include preliminary evidence of infringement and the rights holder's true identity information; where a network user maliciously induces the generation of infringing content by entering infringing prompts and causes harm to others, that network user shall bear infringement liability in accordance with the law.
Where necessary measures such as stopping generation of infringing content or blocking the relevant generation instructions are not taken in time, the rights holder may, under Article 1195 of the Civil Code, ask that network user and service provider to bear civil liability.
A Sheet of Paper in Court
The Opinion also turns the lens on the court itself. Where a party obtains false evidence and fabricates facts for a sham lawsuit by deleting or altering generative or synthetic content labels, entering specific instructions or selectively presenting results, the people's court shall reject that party's claims in accordance with the law.
Another layer concerns evidence itself: where a participant or other person uses AI to forge evidence and obstruct the people's court in hearing a case, it is handled under Article 114 of the Civil Procedure Law, and depending on the circumstances a fine or detention may also follow.
Where litigation documents, case search reports and similar materials are produced with generative tools, the truth and accuracy of the relevant laws, judicial interpretations and cases shall be carefully checked before they are submitted to the court.
When submitting to the court, the use of generative tools shall also be truthfully stated, and responsibility for the truth and accuracy of the content shall be borne in accordance with the law. In other words, a machine can help draft, but the person who signs and files it has to check the content personally.
From the sources of training data to a sheet of paper in court, the point is the same thing: what was used and how it was used have to be spelled out. A model can be complex; the books cannot be muddled. That holds for developers and for people filing materials alike.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.