Infringement complaints over AI-generated content are moving from isolated cases to a routine process that platforms must handle. The Civil Code provides the notice and counter-notice framework, and the Interim Measures for the Administration of Generative Artificial Intelligence Services require service providers to establish a complaint mechanism. In 2024 the Guangzhou Ultraman case and the Beijing AI voice case brought two conflicts into public discussion — generated content against prior works, and voice against personality rights — offering reference points for how the mechanism applies to AI.
The basic structure of notice and takedown
An effective notice from a rights holder usually needs proof of ownership, location of the allegedly infringing content and preliminary evidence. On receipt, the platform should pass it on promptly and take necessary measures as appropriate. An incomplete or unclear notice leaves the platform unable to judge, and therefore unable to act within a reasonable time.
A counter-notice gives the accused party a chance to respond, explaining provenance, licensing or fair use. The platform makes a procedural judgment between notice and counter-notice; it does not replace a court’s substantive ruling. A notifier whose bad-faith notice causes loss may bear liability, which is one reason to keep complete records.
What is different about AI
Traditional takedown targets fixed content on a fixed page, where deletion stops distribution. Generative output is produced on demand and changes with the prompt, so deleting one item does not remove the capability. Necessary measures therefore extend upstream: prompt and keyword filtering, interception of specific outputs, model-level adjustment, and restrictions on accounts that repeatedly trigger infringement.
A platform cannot apply a single “receive notice, delete link” process to every complaint. Where complaints repeat around the same subject matter, the question is whether the output is incidental or systemic. The first is handled item by item; the second requires product and model changes.
The boundaries of keyword filtering
Over-broad filters harm ordinary creation and expression; narrow filters fail to stop repeat infringement. A layered approach works better: a high-risk rights list, explicit sensitive terms, and semantic similarity checks, each tied to a different response. The source of the list, its update history and hit logs should be retained so that the measure can be shown to be reasonable and proportionate.
Filtering is neither the only measure nor proof of exoneration. It reduces probability; it does not promise zero output.
Evidence retention determines the position in a dispute
The most direct evidence in complaint handling is the timing, content, conclusion and measures taken in respect of each notice and counter-notice. Generation logs, user prompts, model versions and filter hits should map to a specific complaint, rather than support only a general statement.
Retention also has to be readable. A complete set of log fields is of limited value if it cannot be located quickly during a defence or inspection. Index records by complaint number and keep technical records and decisions in one place.
Membership fees and profit
Paid membership, per-generation charging and advertising revenue will be used in disputes to discuss profit and the degree of care owed. Charging does not automatically mean a higher duty, but it affects how the business model, user expectations and infringement risk are assessed overall. Products that sell themselves on specific protected content need a licence chain reviewed in advance, not assembled after complaints arrive.
For complaint-response mechanisms, infringement risk assessment and licence-chain review, see our AI infringement service and IP protection service; for content labelling, see AIGC labelling service.
This article is based on the notice and counter-notice rules for network infringement in the Civil Code, the complaint-mechanism requirement in the Interim Measures for the Administration of Generative Artificial Intelligence Services, and public information on the 2024 Guangzhou Ultraman case and the Beijing AI voice case. It is for reference only and does not constitute legal advice.