China Drafts New AI & Mental Health Laws That Draw Interest from Around the World

A significant change in global AI regulation has occurred with the release of a draft framework titled “Provisional Measures on the Administration of Human-like Interactive Artificial Intelligence Services” for public comment by China's Cyberspace Administration (CAC).

Remarkably, this proposal focuses on the relationship layer of AI—systems that mimic personality, emotion, and human-like interaction—as opposed to previous laws that concentrated on content filtering or model-level safety.

According to the concept, the main safety hazards that require explicit design and operational controls are psychological injury, emotional dependency, and behavioural manipulation.

Scope & Regulatory Positioning

Any company or individual in mainland China that offers AI systems that imitate human personality traits, thought processes, or communication styles and that communicate emotionally through text, audio, visuals, or video is subject to the regulations.

The CAC distinguishes these systems from general-purpose foundation models and generic generative AI rules by defining “human-like, emotionally interactive” AI as a unique regulatory category. Both sector-specific regulations and this new framework must be followed by providers in regulated industries like healthcare, banking, and law.

Prohibited Behaviour

The document adds new limitations specific to emotionally interactive AI, but it also includes normal Chinese limits on anything endangering public order or national security. Among them are:

• False promises that have a substantial impact on user behaviour or harm social connections.
• Promoting, suggesting, or elevating self-harm or suicide.
• Damaging mental health by manipulating emotions or verbally abusing someone.
• Using computational manipulation, false information, or “emotional traps” to induce irrational conclusions.

Even when the underlying material is not expressly banned, these restrictions enable regulators to treat some intimacy-orientated design patterns as illegal.

Essential Obligations for Providers

The draft imposes stringent design and operating requirements:

1. Lifecycle Security Responsibilities: Providers must incorporate safety measures into design, deployment, upgrades, and termination. This comprises monitoring, risk assessment, error correction, and log storage.
2. Emotional State & Dependency Assessment: Systems must be able to detect users' emotional states and levels of dependency while maintaining privacy. If severe emotions or addiction are recognised, providers must intervene.
3. Crisis Response & Manual Takeover: When high-risk behaviours emerge, AI must send supportive messages and refer users to professional aid. If a user expresses suicide or self-harming intent, providers must manually intervene and contact guardians or emergency contacts. This is positioned as a necessary capacity, not a best-effort commitment.
4. Safety Precautions for Vulnerable Users: Minors and the elderly must provide emergency contact details. Providers must assist older users in setting up contacts and warn guardians when risks occur.
5. Minors Mode with Guardian Controls: A specialised minor's mode must incorporate reality checks, time constraints, and guardian approval for emotional support services. Guardians require access to usage summaries, risk alerts, blocking tools, expenditure limitations, and duration controls.
6. Disclosement & Reality Reminders: Systems must explicitly indicate that the user is dealing with AI, with dynamic reminders on initial use, new logins, and dependence signals.
7. Session Duration Warnings: Users must be prompted to stop after two hours of continuous involvement.
8. Commercial Conduct Requirements: Emotional companionship services must provide a simple escape. Providers must maintain complaint channels, explain handling procedures, and provide feedback on results.
9. Training & User Interaction Data: Strict guidelines regulate training data quality, openness, and security. User interaction data, particularly chat logs, cannot be utilised for training without explicit authorisation. Annual audits are required for minimal data handling.

Platform Requirements & Security Evaluations

Providers are required to submit security assessments for new functions and technologies when reaching significant user thresholds or when risks to national security or individual rights emerge.

App stores and distribution platforms must verify these assessments, with non-compliant services potentially removed, thereby making regulatory compliance essential for market access.

Enforcement & Worldwide Context

China's enforcement mechanism handles AI companion violations by issuing warnings and making adjustments.

Globally, jurisdictions like New York, California, and Texas are implementing laws on AI companions, alongside the EU’s AI Act, which tackles emotional manipulation and consumer protection. There is a growing international acknowledgement of psychological safety as a critical risk factor.

In conclusion, China’s draft framework defines emotionally interactive AI as a regulated domain, incorporating psychological safety into mandatory product design. This framework serves as a model for other jurisdictions as emotionally capable AI systems become more widespread.

Read more news:

• Trump: The US & Venezuela Strike a $2 Billion Oil Deal
• The Most Incredible Technology to Be Revealed at CES 2026
• AI Boom 2025: A Half-Trillion-Dollar Surge Raises Questions about Wealth & Market Dynamics