AI-generated advertising in agricultural livestreaming and the auditability of disclosure regulation: a three-layer audit of public Douyin data from China

Agricultural products are credence goods, and in livestreaming, the grower’s visible presence signals the price. China has required labeling of deep-synthesis content since January 2023 and, since 1 September 2025, of AI-generated synthetic content. Whether public platform data can verify compliance in a market where the authenticity premium is especially salient remains unclear, given that growers themselves sell produce. We audited three layers of public Chinese platform data: 189 policy snippets, 19 accounts retrieved by AI technical terms, and 455 unique authors retrieved by crop names. In the one policy coding field with complete intercoder agreement ( κ = 1.000), only 6 of 189 snippets name authenticity as a farmer asset. Retrieval by technical terms returns intermediaries: 13 of 19 accounts sell AI tools or tutorials, and none discloses using AI to sell its own produce. Among AI-announcing cases retrieved by crop-name queries, an expanded text-field detector validated at 94.7% recall flags 10 of 455 authors: seven supply-side or service-provider accounts, one incidental non-agricultural mention, and two product-query artifacts; none discloses grower own-use of AI livestreaming. Checks of API-exposed platform metadata for all 575 video records and cover images for 472 records found no platform-side AI label or visible AI/AIGC/statutory AI-generated notice. AI was announced primarily by accounts that sell, service, or discuss AI, rather than by produce-selling grower accounts. No grower own-use disclosure appears in either audited slice: 0 of 19 technical-keyword accounts and 0 of 455 transaction-query authors. The zero admits two readings that the audited surface cannot separate: smallholders do not use these tools, making the requirement inert for them, or they use them without disclosure. No label-reading instrument applied to the audited surface can decide between them, because such inspection observes announcement rather than use. Platforms and regulators may hold backend data unavailable to an external audit. If the theoretically derived authenticity-penalty mechanism holds, disclosure would impose an uneven cost, falling more heavily on growers than on intermediaries. The observed pattern is consistent with this hypothesis but does not establish it.