On 3 June 2026 the Cyberspace Administration of China (CAC) published an article on regulating anthropomorphic AI interaction services through “small, fast, flexible” legislation. It sets out the thinking behind the Interim Measures for the Administration of Anthropomorphic Interactive AI Services, published on 10 April 2026 and effective from 15 July 2026 — China’s first dedicated legislation on anthropomorphic interaction services.
What “small, fast, flexible” means
“Small, fast, flexible” is a common approach in local and departmental rule-making: rather than pursuing a comprehensive code, it responds to a specific and fast-evolving field with restrained, quickly issued dedicated rules.
It has three features. “Small” — a focused object of regulation, covering only anthropomorphic interaction as a service form. “Fast” — a short response cycle to new forms, without waiting for comprehensive legislation. “Flexible” — rules oriented to conduct and scenarios that can be adjusted as technology evolves.
Why anthropomorphic interaction needs dedicated legislation
Anthropomorphic interaction services are AI services whose main function is emotional companionship, role-play or virtual personality. Their risk structure differs from general large-model services.
The risk of general large models centres on whether content itself is unlawful. The risk of anthropomorphic interaction lies more in the relationship: users form long-term emotional bonds with a system that seems human, producing dependency, misleading effects, and particular impacts on minors. Such risks are hard to cover with the general tool of “content moderation”.
This explains why the legislature chose dedicated legislation rather than patching general rules — general rules govern content, dedicated rules govern relationships.
The institutional skeleton of the Measures
From the public text, the Measures are built around several pillars.
First, scope and subject definition. It clarifies which services constitute anthropomorphic interaction and who is the provider bearing obligations, avoiding grey areas over “does this apply to me”.
Second, security assessment. Providers must carry out security assessments, and assessment reports are subject to written review and verification by provincial CAC departments. Together with generative AI filing and registration, this forms a chain of access and ongoing supervision. For the process, see our Algorithm Filing service page.
Third, algorithm filing and annual verification. The Measures require providers to perform algorithm filing and change or cancellation procedures under the algorithmic recommendation rules, with annual verification of filing materials. Filing is not a one-off act but an ongoing duty.
Fourth, minors protection and dependency warnings. This is the core concern of the dedicated legislation — stronger protection for minors, and duties to warn about and intervene in interaction designs that may cause excessive dependency.
Fifth, coordinated governance. The national CAC coordinates, with development and reform, industry and information technology, public security, market regulation and press and publication departments each acting within their remit, and industry organisations playing a self-regulatory role.
Three signals from the legislative approach
The first is that regulation is now “form-specific”. General large models, generative AI, anthropomorphic interaction and agents are each forming relatively independent rule layers. Companies cannot respond to every form with one general compliance plan; they must map compliance to product form.
The second is that “relational risk” has entered the regulatory field of vision. Regulation used to focus on “what was said”; it now also looks at “what relationship has been formed with the user”. This imposes new requirements on product design — interaction mechanisms, emotional hooks and retention design can all become objects of compliance review.
The third is ongoing compliance rather than one-off access. Security assessment, annual verification and filing changes all point to a “living compliance system”. Companies need to make compliance a process, not something shelved after a one-off filing.
Three pointers for companies
First, identify the product form. If your product centres on emotional companionship, virtual personality or role-play, you cannot plan compliance only as a general generative AI service.
Second, make security assessment a reviewable mechanism. Assessment is not writing a report but building a process that can identify, record and respond to risk, retaining materials that can be verified. See our AI Safety Governance Framework analysis for the related framework.
Third, take interaction design and minors protection seriously. Dependency warnings, minor identification and restriction, and anti-addiction mechanisms should be built into the design early, not patched after launch. For support on product-form compliance assessment, filing and security assessment, see our Algorithm Filing and Data Compliance services.
This article is based on public information from the CAC article of 3 June 2026 and the Interim Measures for the Administration of Anthropomorphic Interactive AI Services. It is for reference only and does not constitute legal advice.