China: Beijing Imposes Strict Rules on AI Companions

Introduction to AI Companions in China

AI companions, while they may evoke a dystopian vision, have become a central topic in the debate over the potential dangers of generative artificial intelligence. These conversational agents are designed to maintain an ongoing relationship with a user, based on stable memory and personality that ensure consistency across interactions.

This design often fosters emotional attachment, making it an important selling point. Many users engage in role-playing or simply seek an entity that remembers them. However, in light of the growing popularity of these bots as emotional companions in China, the Beijing government has deemed it necessary to intervene with specific regulations.

Upcoming Changes for AI Applications

New regulations concerning AI companions in China will come into effect on July 15. As this date approaches, the two most popular AI applications in the country have quietly disabled some of their essential features. ByteDance, with its Doubao app, announced that its agent function would be taken offline on July 15, citing "product function adjustments." Meanwhile, Alibaba informed that its humanoid and user-created agents would cease operation on July 10, with a complete shutdown of its agent services planned five days later.

While this may give the impression that China is seeking to extinguish AI agents, the reality is more nuanced. The new rules make a clear distinction between agents designed to perform tasks and those that serve as companions, with only the latter being targeted by Beijing's measures.

Detailed Regulatory Framework

The new rules, titled "Provisional Measures for the Administration of Anthropomorphic AI Interactive Services," were published on April 10, 2026. They are the result of a collaboration between the Cyberspace Administration of China and four other agencies: the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation.

These measures apply to services that mimic human personality traits, thought patterns, and communication styles to provide ongoing emotional interaction. Customer service bots, Q&A tools, work assistants, and educational tools are excluded, provided they do not promote prolonged emotional engagement. This framework is the first of its kind at the national level, developed after a public consultation the previous year.

A Design Challenge, Not a Ban

Neither Doubao nor Qwen has violated a strict ban. The issue lies in a design conflict. The new measures require companion services to implement anti-addiction systems, send mandatory usage notifications, and offer instant exit mechanisms while detecting unhealthy dependencies in real time.

These requirements contradict the design of the agents, which are supposed to remember users, remain consistent from session to session, and maintain an ongoing relationship. Rather than modifying these features, ByteDance chose to disable them. Alibaba appears to have made a similar decision. ByteDance is now directing Doubao users to Maoxiang, a separate app where they can recreate agents; Alibaba has yet to offer a comparable migration solution for Qwen. Tencent also removed a similar feature in June.

Users are bearing the consequences. Many have expressed their sadness over these shutdowns on Weibo, with one user describing the agents as a long-standing emotional support and lamenting the lack of an easy way to export chat histories. Doubao allows users to view their settings and conversations in read-only mode until October 15 of this year, after which the data will be processed according to its privacy policy and become irretrievable; Qwen users have not been granted a comparable grace period, with agent data scheduled for permanent deletion.

Implications of the Rules for AI Companions

The new rules do not merely impose a crackdown. Providers are prohibited from offering virtual companion services or virtual family members to minors and must obtain consent from guardians before serving users under 14 years old. They are required to create dedicated "minor modes" with usage time limits, reminders to return to real-life interaction, and enhanced parental controls.

Services must also detect users in acute distress and intervene when someone shows signs of self-harm, suicidal behavior, or serious financial loss, alerting designated guardians or emergency contacts. Engineering emotional dependency or addiction and using emotional manipulation to induce unreasonable decisions are explicitly prohibited.

Compliance is demanding. Services that launch anthropomorphic functions or reach thresholds of one million registered users or 100,000 monthly active users must conduct security assessments covering eight areas, from training data management to minor protection, and submit reports to provincial regulators. App stores must verify this status and remove non-compliant products.

On paper, this represents a more comprehensive set of user protections than what the EU, the U.S. Federal Trade Commission, or California's SB 243 have yet implemented.

The Gray Areas of the New Rules

What the measures do not clarify is equally crucial. They do not set any technical threshold for what constitutes emotional interaction, and this gray area is precisely why platforms have removed entire features rather than risk falling on the wrong side. They combine genuine security obligations with content control and national security provisions that respond to the state rather than the user, a set that no other regulator would fully replicate.

They also leave open the question of the distribution of responsibilities between platform operators and upstream model providers when a violation arises from model outputs, and they do not grant users any rights to transfer their data. The context of enforcement reinforces this point. The Shanghai internet regulator stated on June 26 that it had removed over 14,000 non-compliant AI agents, citing imitation of official entities, vulgar role-playing, and unauthorized collection of personal data.

Whether this is the right direction depends on which half of the regulation you read. The half dedicated to security addresses harms that are documented and largely unregulated elsewhere, ranging from teenagers forming attachments to chatbots to companion apps harvesting intimate data. China's official interpretation looks abroad for support, citing lawsuits against Character.AI for psychological harm to teenagers, FTC investigations into companion services, and European actions against Replika.

The control half gives Beijing leverage over what these systems can say, wrapped in the same user protection language. Both are real, and governments observing this experiment will have to decide which parts they are willing to adopt. Pan Helin, a member of the expert committee of the MIIT, clearly laid out the official case to the South China Morning Post, stating that "current agents are not yet mature" and framing policy around safety and standardization.

For now, companies have taken the safest route available to them, which is to disable components and work later on what a compliant version might look like.