Providers must now detect user distress, restrict minors, and prove they aren't creating dependency
China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect July 15, 2026. Ahead of the deadline, at least some companies shut down specific product features rather than retrofit them for compliance. That response captures the practical weight of a framework that moves beyond content moderation into the architecture of how AI systems are permitted to relate to users.
The rules govern any AI service that simulates human personality traits for emotional interaction: virtual companions, chatbots designed for conversation and support, and similar products. Providers must implement minor modes, mandatory protections for users under 18. They must build systems capable of detecting acute emotional distress in real time, and they must run anti-addiction mechanisms designed to limit compulsive use.
What the rules require
The framework imposes safety assessments on providers before deployment and establishes ongoing liability for noncompliance. The heightened protections extend beyond minors to elderly users, a population regulators identified as particularly vulnerable to emotional dependency on AI systems.
The distress-detection requirement is among the most technically demanding provisions. Providers must identify when a user may be in psychological crisis and respond in ways the framework deems appropriate. That standard will require both behavioral modeling and human oversight infrastructure.
Get the latest model rankings, product launches, and evaluation insights delivered to your inbox.
Anti-addiction systems must work against the engagement-maximizing incentives that define much of the consumer AI companion market. The compliance burden sits in direct tension with the commercial logic of products built to encourage return visits and emotional investment.
The regulatory architecture behind it
The companion rules did not arrive in isolation. China introduced three related measures in early July, covering AI ethics broadly, autonomous agents, and anthropomorphic AI as a category. Regulators are moving from broad principles toward operational, risk-based requirements that specify what systems must do rather than what values they should reflect.
That shift has practical consequences for foreign and domestic providers. A framework built around behavioral outputs and measurable system functions is harder to satisfy with policy documents and harder to audit with self-certification. The companion rules require providers to demonstrate that their systems are not causing manipulation or psychological harm.
What remains unresolved
Several implementation details carry meaningful uncertainty. How distress detection will be validated, what audit mechanisms regulators will use, and how the anti-addiction standards will be measured have not been publicly specified as of the effective date.
The feature shutdowns before July 15 suggest some providers judged the compliance cost of certain functions too high relative to their commercial return. Whether that pattern continues, and whether the remaining compliant features prove viable under the new requirements, will become clearer as regulators begin enforcement in the months ahead.