China is forcing consumer AI platforms to choose between emotional companion features and a new safety rulebook, after ByteDance and Alibaba switched off humanlike agent functions before a July deadline.

AI News reported that China's AI companion rules take effect on July 15.

ByteDance's Doubao told users its agent function would go offline on July 15, while Alibaba's Qwen said humanlike and user-created agents would stop working on July 10 and wider agent services five days later.

China AI Companion Rules Take Effect On July 15

The measures target services that simulate personality traits, thinking patterns and communication styles to sustain emotional interaction with users.

The April 10, 2026 framework names the Cyberspace Administration of China as one issuing agency and also lists four partner bodies covering development planning, industry and information technology, public security and market regulation.

The rulebook does not ban all AI agents.

It separates emotionally engaging companion services from customer-service bots, workplace assistants, education tools, research tools and knowledge Q&A services when those products avoid sustained emotional engagement.

That boundary explains the platform response.

AI News reported that Doubao and Qwen were affected because their features were built around persistent personas, user-created agents and ongoing relationships.

Tencent's Yuanbao pulled a comparable feature back in June, according to the same report.

Doubao And Qwen Users Lose Agent Features

ByteDance is directing Doubao users towards Maoxiang, a separate app where users can create agents again.

Alibaba has not announced an equivalent Qwen migration route.

AI News reported that Doubao will keep configurations and conversations visible in read-only mode until October 15.

After that point, Doubao plans to handle the records under its privacy policy and users will not be able to recover them.

The report did not identify a comparable Qwen grace period, and said Qwen agent data is set for permanent deletion.

The user-data difference gives the shutdown a practical consequence beyond compliance.

A platform can withdraw a humanlike agent feature quickly, but users who treated the agent as an ongoing relationship may lose the conversation history and persona settings attached to it.

Platforms Face Minor Protection And Security Assessment Duties

The rules bar providers from offering virtual companion or virtual family-member services to minors.

AI News reported that providers must obtain guardian consent before serving users under 14 and must build minor modes with time limits, reminders to return to real-world interaction and stronger parental controls.

The measures also require services to detect acute distress, self-harm signals, suicidal behaviour or serious financial loss and escalate cases to designated guardians or emergency contacts.

Providers are prohibited from engineering emotional dependence or using emotional manipulation to induce unreasonable decisions.

Scale brings another compliance trigger.

AI News reported that services launching anthropomorphic functions, or crossing one million registered users or 100,000 monthly active users, must complete security assessments across eight areas.

Provincial regulators receive those filings, and app stores are required to check compliance status before keeping products available.

Beijing Leaves The Emotional Interaction Threshold Unclear

The measures do not set a technical threshold for what counts as emotional interaction.

A memory-enabled assistant, a roleplay bot and a productivity agent can share similar product mechanics while being treated differently under the rulebook.

AI News also cited Shanghai's internet regulator as saying on June 26 that it had removed more than 14,000 non-compliant AI agents, including agents linked to impersonation of official entities, vulgar role-play and unauthorised collection of personal data.

Pan Helin, identified in the report as an MIIT expert-committee member, framed the official case around safety and standardisation and said current agents are not yet mature.

The report also said the measures leave unresolved how liability is split between platform operators and upstream model providers when a violation comes from model output.

The measures did not disclose a technical test for emotional interaction, a data portability right for users, or a clear liability split between AI app operators and upstream model providers.