China's Top Court Rules AI Clones Violate Personality Rights

While the US and EU continue to debate deepfake and voice-cloning rules through a patchwork of state laws and disclosure mandates, China's judiciary moved with more force: on September 7, 2026, the Supreme People's Court issued its first national judicial Opinions on AI disputes — 24 articles explicitly classifying unauthorized identifiable AI clones of a person's voice, likeness, or persona as personality-rights violations, alongside rules for training data, provider notice duties, and emergency injunctions.

TL;DR

What "personality rights" covers, and why it's the right legal hook

Personality rights, as a legal category under China's Civil Code (in force since 2021), protect a person's control over their own name, image, likeness, voice, and reputation — the same underlying interest right-of-publicity law protects piecemeal across different US states. What's reportedly new here isn't the existence of personality rights protection itself, but the top court explicitly extending that framework to AI-generated clones — closing the ambiguity around whether an AI-synthesized voice or face, rather than a photograph or recording of the real thing, counts as a violation of the same right.

That distinction matters because it's exactly the gap voice-cloning and likeness-generation products have operated in globally: is a synthetic reproduction of someone's voice, generated by a model rather than recorded directly, legally equivalent to using their actual voice without permission? China's top court just answered that question with a clear yes, for its jurisdiction — not via new legislation, but through judicial guidance lower courts must follow.

What the Supreme People's Court actually issued

On September 7, 2026, the Supreme People's Court (SPC) released the Opinions on the Trial of Cases Involving Artificial Intelligence Disputes — 24 articles in five parts, framed as guidance for applying existing statutes (Civil Code, Cybersecurity Law, Copyright Law, Consumer Protection Law, Personal Information Protection Law, and others) to AI disputes. China has no dedicated AI statute yet; these Opinions are the highest national judicial rules on the topic to date.

Xinhua's coverage quoted Zhou Jiahai, head of the SPC research office: "We cannot expect every consumer to become an expert at spotting deception. The law must step in promptly to protect consumers' legitimate rights and interests." That framing explains why the document spans personality rights, algorithmic pricing, and provider liability in one release — courts needed a single reference as impersonation tools went mainstream.

The personality-rights section (Article IV) is the headline for clone products, but Articles VI–VIII matter equally for training pipelines and hosted inference.

Article IV — clones, voices, and "AI resurrection" in detail

The Opinions regulate what they call "AI face-swapping and voice-mimicking" and "AI resurrecting the deceased" under generative AI. The operative tests:

Likeness: Without consent, using AI to process a natural person's name or portrait to generate an identifiable virtual digital image, then using or publishing it, supports a personality-rights claim — specifically name and portrait rights.

Voice: Without consent, using a person's voice as training material to imitate timbre, tone, and pronunciation style, producing a synthetic voice identifiable as that person, supports a voice-rights claim. Note the liability hook: training use, not only publication of a recording.

Reputation: Manipulating a synthetic likeness or voice to spread false statements or improper conduct that lowers social evaluation constitutes reputation infringement.

Deceased persons: Unauthorized AI creation or use of a deceased person's virtual image that infringes name, image, or reputation can be sued by close relatives under Civil Code Article 994.

This codifies at national level what Beijing Internet Court cases had already signaled. In September 2025, the Beijing Internet Court's Typical Cases Involving Artificial Intelligence included Yin (voice rights extend to AI-synthesized voices) and He (personality rights extend to AI-generated virtual images). The SPC elevated that line of reasoning while declining to designate the Beijing court's copyright cases on AI-generated works as national typical cases — leaving who owns AI output unresolved nationally (more below).

Training data — Article VI's "reasonable scope" gray zone

Article VI addresses a question US and EU builders still fight about: Can you train on publicly posted personal information?

The Opinions say processing voluntarily disclosed or lawfully published personal information within a reasonable scope for model training, where the individual has not explicitly refused, is generally not a personal-information infringement. If processing significantly impacts the individual's rights, consent is required.

Courts weigh purpose, necessity, data sensitivity, foreseeable use, and disclosure context. That is not a bright-line "public internet = free training" rule — it is a fact-intensive test that will land on product teams shipping fine-tunes on creator content, voice samples scraped from social video, or celebrity likeness datasets.

For voice-cloning products specifically, Article IV's consent requirement for identifiable synthetic voice and Article VI's training analysis can stack: even "public" voice clips may fail if the output is identifiable and consent was never obtained.

Provider liability — notice, takedown, and malicious prompts

Article VII imports a familiar internet-law pattern to generative AI:

• If auto-generated content infringes reputation or privacy, and the provider fails to stop generation after notification, the provider bears tort liability for resulting damage.
• Notifications must include preliminary infringement evidence and the rights holder's real identity.
• If a user maliciously prompts the model to generate infringing content, the user is liable; if the provider fails to block relevant instructions after notice, the provider can share liability under Civil Code Article 1195 (China's notice-and-takedown analogue).

Builders shipping hosted inference should read this as: content filters and generation stops are legal infrastructure, not optional safety theater — especially for likeness and voice features where harm is immediate and public.

Article VIII adds personality-rights injunctions: where delay would cause irreparable harm (Xinhua's example — defamatory sexual claims using a digitally altered face), courts may order cessation of the act or require network/AI providers to stop providing the relevant service.

Beyond clones — algorithmic pricing and AI hallucinations

The Opinions are not only about faces and voices:

Article X — "Big data price discrimination": Businesses using algorithms to impose unreasonable differential pricing on the same product without justification bear tort liability if consumers are harmed. Courts examine whether treatment restricts fair trade, targets personal browsing or payment history, and whether stated justifications are non-discriminatory. E-commerce and travel platforms using dynamic pricing models need documented, auditable reasons — not opaque personalization.

Article VII (hallucination angle): Xinhua highlighted that providers notified of rights-infringing generated content must act timely or face liability — aligning judicial pressure with the same failure mode explainx.ai covered in AI hallucination legal cases where lawyers were sanctioned.

Article XIX — litigation abuse: Parties who use AI to forge evidence, strip synthetic-content identifiers, or file false litigation face fines, detention, or criminal charges. Legal-tech copilots serving Chinese courts or firms need provenance and verification workflows, not raw chat output in filings.

What the Opinions did not decide — AI copyright

The National Law Review and China IP Law Update both noted a deliberate silence: the Opinions do not rule whether AI-generated works qualify for copyright protection. Beijing Internet Court Case Li v. Liu (2025) held a Stable Diffusion output could be a copyrightable work with the user as author; the SPC did not endorse that as a national typical case and later stripped Internet Courts of jurisdiction over many online copyright disputes (effective November 2025).

Practical takeaway for builders: personality rights and copyright are on different tracks in China. You can infringe someone's likeness while still fighting over whether your model's output is protectable IP — see also is removing an AI watermark illegal for how disclosure and ownership interact elsewhere.

How this compares to the US and EU approach

The regulatory landscape here has been genuinely fragmented, and it's worth placing this ruling in that context:

• United States: No single federal AI likeness law exists. Protection comes from a patchwork of state right-of-publicity statutes, with the NO FAKES Act framework gaining adoption in some jurisdictions through 2026 — but coverage and enforcement vary significantly state to state.
• European Union: The EU AI Act requires disclosure when content is AI-generated (a deepfake must be labeled as such) but doesn't create the same kind of direct personal-ownership right over one's likeness that personality rights law does.
• China: A single SPC Opinions document extends personality rights to AI clones, sets training-data rules, provider notice duties, and injunctions — judiciary-led, nationwide, without waiting for an AI Act.

For regional context on China's AI industry — who builds what, and where products ship — see explainx.ai's top Chinese AI companies guide and American closed vs China open-weights strategy.

What this means for builders

If you're building or shipping any product involving voice synthesis, face/likeness generation, or persona cloning — and especially if that product serves users in or connected to the Chinese market — the Opinions are a concrete signal to treat consent and licensing as a legal requirement, not a product-design nicety. That's consistent with explainx.ai's coverage of AI hallucination legal cases where lawyers were sanctioned and AI and law for contracts and legal help: capability outruns legal review until courts publish rules like these.

Compliance checklist — voice and likeness products

Use this as a pre-ship audit for any feature that synthesizes a recognizable person:

1. Consent artifact — Signed license or recorded consent covering training, inference, and publication — not just "we have a clip."
2. Identifiability review — If output could be recognized by a reasonable listener/viewer, assume Article IV applies.
3. Deceased personas — Block "AI resurrection" flows unless close relatives authorized; Article IV + Civil Code 994.
4. Notice-and-takedown runbook — Provider-side ability to stop generation and block prompts within SLA after a valid rights-holder notice (Article VII).
5. Injunction readiness — Legal contact and technical kill switch for urgent reputation harm (Article VIII).
6. Training provenance — Document sources for voice/likeness training data; Article VI "reasonable scope" is not a safe harbor for celebrity scrapes.
7. Synthetic labeling — Align with disclosure norms in other markets (New York synthetic performer law, EU AI Act guide, C2PA on Android) even where China's Opinions focus on rights more than labels.

Open-weight voice stacks — VoxCPM2, Voicebox-style studios, Miso One real-time TTS — lower the technical bar to clone; they do not lower the legal bar. Self-hosted deployment does not exempt you if the product serves Chinese users or targets Chinese nationals' likenesses.

Product patterns that reduce risk

For agent products that speak on behalf of brands, also read Hugging Face speech-to-speech voice agent guide alongside this post — telephony agents amplify impersonation harm because callers assume human identity.

Criminal and autonomous-vehicle provisions — why builders should skim them

Parts IV–V include articles easy to miss if you only ship SaaS:

• Article XX — Criminal liability for AI-enabled fraud, defamation, illegal data extraction, obscene material distribution, and circumventing driver-assistance monitoring leading to accidents.
• Article XI — Product liability for autonomous and driver-assist vehicles, including court-ordered production of driving logs.

If your AI stack touches mobility, fintech impersonation, or adult content, the Opinions tie civil personality rules to criminal enforcement language — relevant for risk teams even when your core feature is voice cloning.

What's still uncertain or untested

The Opinions themselves are published (Chinese text via the SPC); English summaries from Xinhua and legal analyses confirm core articles. What remains live uncertainty:

• Damages and penalties — Articles state when courts shall support claims, not typical award sizes.
• "Reasonable scope" for training — Article VI requires case-by-case fact tests; no safe harbor list for dataset types.
• Cross-border enforcement — Opinions govern Chinese courts; effect on apps hosted overseas but used in China is untested in public case law explainx.ai could verify at publication time.
• Copyrightability — National silence persists; Beijing Internet Court copyright wins are not SPC-typical cases.

Anyone making a compliance decision should read the official Opinions text with counsel — not rely on summaries, including this one.

What people are asking (September 2026)

Does this ban all deepfakes in China?
No — it subjects identifiable clones without consent to civil liability. Consented, authorized uses remain viable with proper rights clearance.

Does training on a celebrity's YouTube clips violate the rules?
Likely high risk if the output voice or face is identifiable and consent was never obtained — Article IV targets training material explicitly for voice.

Are US companies liable if only Chinese users misuse the API?
Article VII's provider duties apply to generative AI service providers in the Chinese legal context. Cross-border enforcement and entity structure matter — get local counsel.

How is this different from the EU AI Act?
The EU Act emphasizes risk tiers, documentation, and disclosure; China's Opinions emphasize personality rights and tort liability under existing civil law. You need both lenses if you ship globally — see EU AI Act complete guide.

Does this affect open-source voice models?
Yes, if you deploy them as a service or product in scope. Open weights shift who holds liability; they do not eliminate consent requirements for identifiable clones.

Related on explainx.ai

• AI regulation: EU AI Act and US policy, complete guide
• AI hallucination legal cases: lawyers sanctioned
• AI and law: legal help and contracts guide
• Is removing an AI watermark illegal? DMCA and the EU AI Act
• Felony Bench: AI agent legal liability under the CFAA
• Top Chinese AI companies and startups guide
• New York AI video disclosure law — synthetic performers
• VoxCPM2 tokenizer-free TTS and voice cloning
• Voicebox open-source voice studio
• Hugging Face speech-to-speech voice agent guide
• American closed AI vs China open-weights strategy
• Felony Bench — AI agent legal liability under the CFAA

Sources

• Xinhua — China's top court sets rules for AI deepfakes and other disputes (September 7, 2026)
• Supreme People's Court, Opinions on the Trial of Cases Involving Artificial Intelligence Disputes (最高人民法院关于依法审理涉人工智能纠纷案件的意见), released September 7, 2026 — analyzed in English by The National Law Review and China IP Law Update
• Beijing Internet Court Typical Cases Involving Artificial Intelligence (September 2025) — Cases Yin (voice) and He (virtual image), cited in SPC-adjacent analysis

Opinions text and article numbering reflect SPC release as of September 7, 2026, and English legal summaries reviewed September 11, 2026. This is explainx.ai's practitioner reading, not legal advice — confirm with qualified counsel before shipping or blocking features.