China's Supreme People's Court has published four network rule-of-law typical cases, including a 2025 Beijing judgment holding a book seller jointly liable for unreviewed AI-voice promotion and a 2025 Shanghai judgment clearing a RAG search engine on stated conditions

What the SPC published on 16 September, and what it did not

The release is a news-bureau publication of four case summaries, each in three parts: 基本案情 (basic facts), 裁判结果 (judgment outcome) and 典型意义 (typical significance). The second paragraph says the SPC issued them on 16 September, during the National Cybersecurity Publicity Week of 14 to 20 September 2026, to support legal publicity for the week and to give play to the exemplary and guiding role of judicial decisions.

It is not a judgment. The SPC decided none of the four disputes. Case 2 was decided by Beijing Internet Court on 18 July 2025. Case 3 was decided by Shanghai Xuhui District People's Court on 1 July 2025 and affirmed by Shanghai Intellectual Property Court on 25 December 2025. What is new this week is the selection and the SPC's commentary, not the outcomes.

The release also is not entirely about AI. Case 1 is a personality-rights injunction against online harassment by two influencers with more than 20 million followers between them, granted by Changchun New District People's Court in Jilin for six months; the summary makes no finding that the harassment used AI. Case 4 is a criminal conviction, affirmed on 6 December 2023, for supplying software and server support to a Philippines-based cross-border gambling group. We cover the two AI cases and note the other two for scope.

Case 2: the commissioning company's failure to review made it a joint infringer

In 2024 a culture company selling parenting books engaged a sales influencer (带货主播) to make and post short promotional videos. Their agreement provided that if the videos infringed third-party rights, the commissioning company would bear responsibility. The videos used clips of a lecturer, Li Moujin, speaking and teaching in public, and paired them with an AI-synthesized audio track that the summary describes as highly similar to her voice in timbre, intonation and pronunciation style. A button beside the video linked straight to the company's online store. The company had not reviewed the source of the material or whether it was authorised. She sued the company, not the influencer, for an apology and RMB 250,000.

Beijing Internet Court found that using her likeness without consent, combined with an AI-synthesized voice that closely matched her voiceprint, infringed her portrait right and her voice interests. On the company's role the judgment outcome reads: 某文化公司未对视频素材的来源、视频素材是否获得李某瑾授权等事项进行审核，存在未尽合理审核义务的过失，构成共同侵权 (our translation: the culture company did not review the source of the video material or whether the material had Li Moujin's authorisation, was negligent in failing to discharge a reasonable review duty, and constituted a joint infringer). On 18 July 2025 the court ordered an apology and RMB 120,000 for economic loss and reasonable enforcement costs. Neither side appealed and the judgment took legal effect.

The SPC's typical-significance paragraph turns that into a stated rule, and the rule has conditions. In our translation: where a sales influencer commissioned by an e-commerce operator intentionally uses material that infringes another person's personality rights, the influencer infringes; where the commissioning e-commerce operator fails to discharge a review obligation prescribed by law or agreed by contract over the promotional content, and is at fault for the resulting harm, it is a joint infringer with the influencer and bears joint and several liability. The SPC then points to the facts that made this company's fault clear: the contract expressly gave it review responsibility and management authority over the content, and the videos made obvious and extensive use of a well-known person's likeness and a highly realistic synthetic voice, so the infringement risk was plain.

Two things the release does not say. It does not say every brand that commissions AI-voiced content is automatically liable; the rule as written requires a legal or contractual review obligation, fault, and a contribution to the harm. And it does not treat this as a voice-only dispute or announce any general prohibition on synthetic voices. Both portrait right and voice interests were infringed, and the wrong was the unauthorised use.

Case 3: the RAG search engine was cleared, on four conditions and with a heavier proof burden

A network technology company built an AI search engine on a large language model that had completed the Cyberspace Administration of China's filing for generative and synthetic algorithms, using retrieval-augmented generation, which the summary explains as supplementing search results with external knowledge sources. A culture company holding the network dissemination rights to two television series searched for each on 28 November 2024. The top result in each case was a third-party cloud-drive share link whose video matched the series. On 5 December 2024 the rights holder demanded removal; the operator deleted the results the same day. The rights holder sued for damages anyway.

Shanghai Xuhui District People's Court held that the share links did infringe the dissemination right, but that the search operator had not itself uploaded them and so was not a direct infringer. It then rejected aiding infringement on four grounds the summary lists together: the platform displayed content sourced from public web pages; existing technology at the present stage could not automatically identify infringing information; the operator had discharged its model algorithm filing obligation; and it acted promptly and effectively once it learned of the infringement. All claims were dismissed on 1 July 2025. Shanghai Intellectual Property Court dismissed the appeal on 25 December 2025.

The SPC's commentary is where the general statement sits. It starts from Civil Code article 1197, under which a network service provider that knows or should know that a user is using its service to infringe civil rights and fails to take necessary measures is jointly liable with the user, and from the SPC's Provisions on cases infringing the right of dissemination through information networks, which set the factors for knew-or-should-have-known. It then says that because retrieval-augmented search returns results that are more precise and more personalised than ordinary search, courts applying those provisions should consider the technical characteristics, the service model and how the results are generated.

The operative sentence is the one we quote. 搜索引擎服务提供者应当依法承担与其算法、数据优势程度相匹配的注意义务，对检索增强生成技术的算法合规负有更重的证明责任 (our translation: a search engine service provider should, in accordance with law, bear a duty of care commensurate with the degree of its algorithmic and data advantage, and bears a heavier burden of proving the compliance of its retrieval-augmented generation algorithm). The sentence continues with the no-fault formulation: 检索信息来源于互联网，搜索引擎服务提供者未主动上传、编辑、推荐该信息，不符合明知或者应知的主观状态，且在知悉侵权信息后及时作出有效处理的，不具有侵权过错，不构成侵权 (our translation: where the retrieved information comes from the internet, the search engine service provider did not actively upload, edit or recommend it, the provider does not meet the mental state of knowing or having reason to know, and it takes prompt and effective action after learning of the infringing information, the provider is not at fault and does not infringe).

So the conditions are four, cumulative: public-internet sourcing, no active uploading, editing or recommending, no actual or constructive knowledge, and prompt effective action on learning of the infringement. Algorithm filing appears in the Shanghai court's reasoning as one factor among several; it is not in the SPC's no-fault formulation as a substitute for the knowledge condition. And the same sentence loads a heavier burden of proof about the RAG algorithm's compliance onto the provider. Nothing in the release says filing clears outputs, addresses training data, or decides whether a generated answer, as distinct from a returned link, can infringe.

What weight a typical-case release carries

A typical-case release is the SPC's own selection of lower-court decisions, published with commentary to guide adjudication. The release describes that function in its own words: to give play to the exemplary and guiding role of judicial decisions. It does not amend the Civil Code, it does not create a new obligation with an effective date, and it does not say the cases have been designated formal guiding cases. We make no claim about how much weight a PRC court would give either commentary paragraph in a later dispute.

Keep the compilation's status separate from the judgments' status. The Case 2 judgment has taken legal effect and the Case 3 dismissal has survived appeal; both bind their parties. What the SPC has added is a general statement of the rule it reads each case as illustrating, and that statement, unlike a judgment, is not itself an adjudication.

The release sits beside two instruments already on this tracker. Case 1's commentary cites the SPC Opinions on adjudicating AI-related disputes for the injunction rule; that is the 7 September 2026 instrument covered separately, and this release is not it. The Case 2 fact pattern is also distinct from the Beijing Internet Court face-swap dispute covered in March 2026: different plaintiff, different defendants, and different material (a short drama there, commissioned promotional video here).

What we did not verify

What we opened: the full SPC release at court.gov.cn/zixun/xiangqing/512041.html as fetched to disk (27,068 bytes), read end to end in Chinese, including the header block with source and timestamp, the two framing paragraphs, the table of contents and all four case summaries with their facts, outcome and typical-significance paragraphs. Both Chinese passages quoted above were checked character for character against the fetched bytes.

What we did not open: the Beijing Internet Court judgment of 18 July 2025, the Shanghai Xuhui District People's Court judgment of 1 July 2025, the Shanghai Intellectual Property Court appeal judgment of 25 December 2025, any docket number for any of the four cases (the release gives none), the Civil Code or the cited SPC Provisions and Opinions as independent texts (the release itself quotes article 1197 and we rely on that quotation only), the CAC algorithm filing record for the model in Case 3, and any earlier publication of these judgments. Everything we report about the decisions is what the SPC's summary reports.

What we refuse to claim: we do not say the SPC decided anything on 16 September 2026, because it did not. We do not say a new nationwide duty took effect that day, because the release enacts nothing. We do not say algorithm filing or notice-and-takedown alone immunises a RAG or LLM provider, because the SPC's formulation also requires absence of actual or constructive knowledge and public-web sourcing, and adds a heavier proof burden. We do not say every commissioner of AI marketing content is jointly liable, because the stated rule requires a legal or contractual review obligation, fault and a contribution to the harm. We do not say these are the first such cases, or that they are formal guiding cases, or that they resolve training-data or generated-answer copyright questions, because the release says none of that. We give no figure for how many disputes or providers are affected, because the release gives none. Every obligation described above is reported as the SPC's summary states it, in labelled translation where we render Chinese, and none is stated as our own.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• SPC Opinions on adjudicating AI-related disputes (7 September 2026)
• Beijing Internet Court AI face-swap personality-rights case
• Asia-Pacific AI regulation tracker
• AI court rules and decisions tracker
• AI Regulation News hub