A Chinese Court Just Wrote Evidentiary Rules for AI-Detection Reports. No US Court Has Done That Yet.

Quick answer

A Guangzhou Internet Court ruling this year is the clearest statement any court has made about what an AI-detection report can and can't prove. A copyright claimant's Dragon Boat Festival illustration scored 99.65% AI-generated across multiple tests. The defendant used that score to argue the image wasn't eligible for copyright at all. The court didn't treat the number as the answer. It treated the number as one piece of evidence that needed corroboration, named four specific factors for weighing a detection report, and then shifted the burden back onto the plaintiff to prove the image reflected real human creative choices. US courts have ruled on individual AI-detection disputes, but none has laid out a reusable standard this specific. This one is worth understanding even if you never touch a Chinese court, because the underlying question, what is a detection score actually worth as evidence, is the same one coming up everywhere detection results get used to decide something.

What was actually in dispute

The case centered on a Dragon Boat Festival–themed illustration. A company claimed it held the copyright, acquired from the original creator through an assignment agreement, and sued a university for using the image without permission in an article on its official WeChat account. The university's defense wasn't "we had a license." It was "this image was almost certainly AI-generated, and AI-generated work without sufficient human creative input doesn't qualify for copyright protection in the first place."

To back that up, the university submitted AI-detection reports, run multiple times, each coming back with a 99.65% AI-generation probability. It also submitted a set of comparison images: pictures made with mainstream AI tools using similar prompts, showing the same art style, composition, and character design as the disputed illustration, including matching anatomical errors like malformed or missing fingers that showed up in both the disputed image and the AI-generated comparisons.

The plaintiff's evidence for authorship, meanwhile, was thin. It had a copyright registration certificate, a screenshot of first publication, the assignment agreement, and payment records, plus screenshots showing the image's layers. What it didn't have was anything showing how the image was actually made: no original editable source files, no layer-modification history, no creation timestamps. The court asked for more. The plaintiff said it couldn't reach the original creator and had nothing further to submit.

The four-factor test the court actually used

The ruling didn't stop at "the detection report is credible" or "the detection report is unreliable." It set out specific criteria for weighing one at all: whether the tested image is clearly identified, whether the source of the tested material is reliable, whether the testing process was properly preserved (the university had submitted trusted timestamp evidence alongside its reports), and whether the detection platform used is reasonably neutral, meaning no undisclosed relationship to either party.

Even after a report clears that bar, the court held it still can't stand alone. Its stated reasoning is direct: AI-detection platforms generally don't disclose their detection standards, model architecture, training data, or error rates. A report can suggest a probability. It can't, on its own, establish how something was actually made. That's not a swipe at any particular detector. It's an acknowledgment that opacity is close to universal in this category of tool, including the ones courts are increasingly asked to take seriously.

What tipped this case wasn't the 99.65% figure by itself. It was the number plus the matching visual artifacts plus the plaintiff's inability to produce any record of the creative process. Corroboration is what did the work. The court explicitly framed it that way: once a detection report is paired with other evidence pointing the same direction, the burden shifts to whoever is claiming authorship to prove there was real human judgment behind the result. The plaintiff lost because it had nothing to offer once that burden landed on it, not because a percentage score convicted the image of being artificial.

How this compares to what US courts have done

Every AI-detection dispute that's reached a US court so far has turned on process, not evidence standards. In the Yale GPTZero lawsuit, the fight isn't over whether the detector's score was numerically accurate. It's over what the university did and didn't do once the flag came in. At the College of Charleston, the case that a recent CACM feature examined turned on the complete absence of any conversation between the student and the instructor before a grade was issued. These are real, important rulings, but they're fights about fairness of process. None of them produces a transferable rule for evaluating a detection report's evidentiary weight the way the Guangzhou decision does.

That gap matters because the underlying problem isn't unique to classrooms. Detection scores are starting to show up anywhere authorship is contested: copyright disputes, freelance-content fraud claims, insurance and compliance filings, even legal citations in court documents where the question isn't who wrote the sentence but whether the sentence describes a case that exists. Each of those contexts needs its own answer to the same question the Guangzhou court tackled directly: when is a score worth citing, and what has to accompany it before it means anything in a dispute.

What this means if you're the one relying on a score

The four-factor framework translates cleanly outside a courtroom, whether you're a platform moderating user-submitted content, an editor fact-checking a freelancer's authorship claim, or a business trying to document how a piece of creative work was produced.

Ask whether the tested material is unambiguously identified, not a screenshot or a re-upload that could have been altered since the original was made. Ask where the content actually came from and whether that chain is documented. Ask whether the testing itself is reproducible and timestamped, not a one-off result nobody can verify later. And ask whether the tool doing the testing has any stake in the outcome.

Then, regardless of how clean those four answers come back, treat the score as a starting point rather than a verdict, exactly the same standard this site holds itself to. A high score is a reason to look for corroborating evidence. It is not corroborating evidence on its own. If you're building a record of how something was created, whether that's an article, an image, or a piece of code, the court's advice to the losing plaintiff applies just as well outside China: keep the editable source files, the revision history, the timestamps, and the drafts. That paper trail is worth more than any single number a detector returns, and it's the only thing that actually survives a challenge.

If you want to see what a detection score looks like before anyone else does, run a draft through TheChecker.AI's free demo and read the sentence-level breakdown instead of trusting one aggregate percentage.

FAQ

Can an AI-detection report alone prove something was AI-generated in a legal dispute? Not under the standard this court applied. A report can be admitted as evidence, but on its own it only establishes a probability, not a fact. It needs corroborating evidence, like matching source material, documented process records, or independent testing, before a court will treat the underlying claim as established.

Does using AI tools automatically mean something can't be copyrighted? No. The court was explicit that AI assistance doesn't by itself disqualify a work from copyright protection. What matters is whether the final result reflects enough individualized human creative choice. A creator who can document that process, through source files, revision history, or similar records, has a real path to protection even with AI involved somewhere in the process.

Why won't courts just trust a high detection percentage? Because the court found that most AI-detection platforms don't publicly disclose their methodology, training data, or error rates. Without that transparency, a single number can indicate a possibility but can't independently verify how a specific piece of content was actually produced.