On 7 September 2026, China's State Administration for Market Regulation (SAMR) published seven typical cases from its six-month campaign against false promotion in online food sales, which began in April 2026. Two of them, read together, say more about regulatory design than either says alone. In Beijing's Changping District, a seller used an AI virtual human in a livestream to claim that an ordinary food boosted immunity, and was fined 5,000 yuan. In Shenzhen's Luohu district, a company that marketed to elderly consumers through conference-style sales events was fined 250,000 yuan. Xinhua's report of the batch describes both.
The case for the rule
The strongest argument for strict rules on synthetic hosts is straightforward. A virtual human has no face to recognise, no reputation to lose and no accountability. It can run around the clock, be cloned across hundreds of stores, and deliver a health claim with a confidence no human pharmacist would risk. If a regulator let sellers treat the avatar as a separate speaker, every false claim could be disowned. Consumers cannot cross-examine a rendering.
China's rules close that loophole. The Live-Streaming E-Commerce Supervision Measures (SAMR and Cyberspace Administration Order No. 117, issued 18 December 2025, in force 1 February 2026) address AI-generated people directly. Article 37 requires operators using AI-generated human images or video to label them and to keep reminding consumers that the figure is AI-generated. It also places legal responsibility on the livestream-room operator who manages or uses the figure. Article 34 separately bars using AI to fabricate or spread false commercial information.
What the Changping case actually establishes
The case adds little new law. It confirms that the avatar is not a legal shield. Reports on the case say the regulator opened it in April 2026 after receiving a lead and ordered the seller to stop and pay 5,000 yuan. The Beijing Daily report frames it under the Live-Streaming Measures. Xinhua's summary lists the Anti-Unfair Competition Law as the basis. Both can be true, because Article 62 of the Measures routes false-claim penalties through that law and the Advertising Law. The batch summaries do not give the full chain of provisions, and a published penalty decision would settle it.
That is the right doctrine. Liability should attach to the human or company that deploys the tool, not to the tool. The same principle covers a script read by a paid actor, a ghostwritten testimonial or a generated voice. A regime that asks "who decided to say this?" ages better than one written around each new medium.
The proportion problem
The 5,000-yuan fine is also modest. The Measures set a separate 10,000 to 100,000 yuan band for failing to label AI-generated hosts, in Article 59. A seller who both deceives and, as here, may have had a labeling problem could face a larger bill for the disclosure lapse than for the deception. We cannot say from the published summaries whether labeling was reviewed in this case, and that is the point: the public record does not show how regulators weigh the two duties.
Contrast the Luohu case. According to Xinhua's account, the company claimed an ordinary food could cure "three highs" (hypertension, high blood sugar and high cholesterol). Some elderly buyers stopped taking blood-pressure medication, and some suffered cerebral hemorrhage. The fine was 250,000 yuan, fifty times the Changping penalty. This case involved no livestream and no AI. It is a reminder that the harm in online food marketing comes from the claim and its audience, not the production technology.
Proportionate regulation should track that harm. An immunity boost claim on an ordinary food is deceptive and should be penalised. A claim that a food replaces hypertension treatment can kill. A regime that enforces on both with visible, graded penalties earns credibility. One that spends headlines on the novelty of AI hosts risks teaching sellers the wrong lesson, namely that the avatar is the offence.
The campaign is broader than the headline
SAMR's campaign reaches platforms, sellers, manufacturers, marketing and advertising agencies, and livestream operators, as Kelley Drye-style law firm analysis of the announcement notes. Local bureaus are also publishing their own batches. Shenzhen's market regulator posted four cases on 30 July 2026, including an ordinary juice marketed with constipation-relief claims and herbal tea labelled as disease-preventing. Full-chain coverage is defensible when the harm is health misinformation aimed at vulnerable buyers.
The risk is sweep. Campaign-style enforcement tends to produce counts of cases, and the incentive is to find more of them. Honest sellers of herbal products, who cannot easily tell whether a traditional-use description counts as an implied health claim, face the same regime as fraudsters. Clear safe-harbour wording, published penalty reasoning and a distinction between negligent and deliberate claims would reduce that uncertainty without weakening the deterrent.
What to watch
Three things will show whether this enforcement is well calibrated. First, whether SAMR publishes the penalty reasoning, including the provisions applied, so sellers can predict outcomes. Second, whether platforms, which carry a duty under Article 17 to prevent AI-driven false claims, are penalised alongside sellers. Third, whether fines scale with consumer harm, as Luohu suggests, or with how new the technology looks. Liability for the deployer is the right rule. Proportionate penalties, applied predictably, are what will make it work.