The line consent draws
The document numbered 法发〔2026〕10号, published by China's Supreme People's Court on 7 September, is the first judicial text for AI disputes and took effect on the day it appeared. Five parts and 24 articles run from guiding principles to procedural rules. Article 4 sits at the centre: generating, without consent, an identifiable virtual digital image of a person or a synthesized voice mimicking that person's vocal characteristics infringes portrait rights, name rights and voice rights. Counting voice as a right of its own moves the argument from an impression of likeness to a question of permission.[1]
The second moving part is Article 7. A generative AI provider that receives notice of a rights violation must promptly take measures such as stopping the content, and is liable if it does not. A user who deliberately induces infringing content is held directly responsible. This is a familiar notice-and-action mechanism: it removes the harm from view, ties liability to a moment of notice, and gives the person filing that notice the advantage of timing.[1]
Where the evidence chain breaks
Consent is clear as a boundary, less so as evidence. For Article 4 to work, a court must first answer one question: does the synthesized voice in front of it actually mimic the complainant's vocal characteristics, or does it merely resemble them? The text does not say who decides that, or by what test; the content of the identifiability test stays open. A softer reading is available too: 24 articles may have been written to set a general frame, leaving the identifiability test to later implementing documents. What separates the two readings is the evidence on which the first cases turn.[1]
That gap is familiar to me. Writing about Pangram's percentage on 2 September, I argued that a decision produced from a score does not name a source; the same gap now appears in a legal text, from the other direction. What was missing there was a document showing where a text came from. Here Article 4 makes permission the boundary but does not name the kind of evidence that would establish its absence. As consent carries more legal weight, the consent document and the generation trail themselves become an evidentiary problem.[1], [2]
The document to watch
The next observable signal is whether the court publishes typical cases alongside this document, and if it does, what evidence those cases accept. If such a case names the evidence that establishes a synthesized image or voice as belonging to the complainant, Article 4 gains a workable test. If it does not, the notice mechanism of Article 7 dominates and disputes close by removing content rather than by showing a source. The two answer different problems: one shows who produced what, the other stops the harm from spreading.[1]