What Beijing Is Truly Banning

The emergence of AI companions, once a concept bordering on the dystopian, has become a central theme in discussions surrounding the ethical implications of generative AI. These conversational agents are designed to foster ongoing, personalized relationships with users, leveraging memory and a consistent persona to ensure continuity across interactions.

The cultivation of emotional attachment is increasingly becoming a key differentiator and selling point for these AI companions. While much of their current use involves casual role-playing or the desire for an entity that “remembers” them, the category can also blur into that of ordinary digital assistants. However, as a growing number of individuals in China have begun to view these bots as a form of emotional confidant, Beijing has now stepped in to implement regulations governing their use.

China’s new regulations for AI companions are set to take effect on July 15th. In the lead-up to this deadline, two of the most popular consumer AI applications in the country have proactively deactivated features at the core of their companion offerings. ByteDance’s Doubao informed its users that its agent function would be discontinued on July 15th, citing “product function adjustments.” Similarly, Alibaba’s Qwen announced that its humanlike and user-created agents would cease operations on July 10th, with its broader agent services following suit five days later.

At first glance, this might appear to be an outright ban on AI agents in China. However, the reality is more nuanced. The new rules distinguish between AI agents designed for task completion and those intended for companionship, with the latter being the focus of Beijing’s regulatory action.

The governing framework is the “Interim Measures for the Administration of AI Anthropomorphic Interactive Services,” jointly issued on April 10, 2026, by the Cyberspace Administration of China and four other government bodies: 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. This regulation specifically targets services that simulate human personality traits, cognitive patterns, and communication styles to facilitate sustained emotional interaction. Excluded from its scope are applications such as customer service bots, knowledge Q&A systems, workplace assistants, and educational tools, provided they do not engage in prolonged emotional interaction. This marks the first dedicated national regulatory framework of its kind, shaped by public feedback on a draft released late last year.

A Design Challenge, Not a Prohibition

The proactive disabling of features by Doubao and Qwen is not a consequence of violating a direct prohibition, but rather a response to a design conflict with the new regulations. The measures mandate that companion services implement anti-addiction systems, issue mandatory usage notifications, and provide instant exit mechanisms, alongside real-time monitoring for unhealthy dependence.

These requirements pose significant challenges for AI agents designed for long-term user engagement, session continuity, and ongoing relationship building. Rather than attempting to retrofit these agents to comply, ByteDance opted to disable the feature. Alibaba appears to have made a similar strategic decision. ByteDance is now redirecting Doubao users to Maoxiang, a separate application where they can once again create agents. Alibaba has not yet announced a comparable migration path for Qwen. Tencent’s Yuanbao had also withdrawn a similar feature in June.

The impact has fallen upon the users. Many users expressed their disappointment and attachment to these agents on Weibo, with some lamenting the loss of emotional support and the absence of a straightforward method to export their chat histories. Doubao is allowing users to access their configurations and conversations in a read-only format until October 15th of this year, after which the data will be handled according to its privacy policy and become unrecoverable. Qwen users have not been afforded a similar grace period, with agent data slated for permanent deletion.

Key Provisions of China’s AI Companion Regulations

The regulations go beyond a simple crackdown, introducing a comprehensive set of user protections. Providers are prohibited from offering virtual companion or virtual family member services to minors and must obtain guardian consent for users under the age of 14. Dedicated “minor modes” are required, incorporating usage time limits, reminders to engage in real-world activities, and enhanced parental controls.

Furthermore, these services must identify users experiencing acute distress and intervene in cases of self-harm, suicidal behavior, or significant financial loss, escalating such situations to designated guardians or emergency contacts. The deliberate engineering of emotional dependence or addiction, as well as the use of emotional manipulation to induce unreasonable decisions, are explicitly forbidden.

The compliance framework is substantial. Services that introduce anthropomorphic functions or exceed thresholds of one million registered users or 100,000 monthly active users must undergo security assessments covering eight critical areas, from training data management to the protection of minors, and submit these reports to provincial regulators. App stores are mandated to verify compliance and remove non-conforming products.

On paper, these user protections appear more robust than those currently enforced by the EU, the U.S. Federal Trade Commission, or California’s SB 243.

Ambiguities and Future Implications

Equally important are the aspects that remain unsettled by these measures. The regulations do not establish a clear technical threshold for what constitutes emotional interaction, creating a grey area that has prompted platforms to disable entire features rather than risk non-compliance. The rules blend genuine safety obligations with content moderation and national security provisions that prioritize state interests over individual user rights, a combination that few other regulators would adopt wholesale.

Moreover, the regulations leave open questions regarding the division of liability between platform operators and upstream model providers when violations stem from the model’s outputs. They also do not grant users the right to port their data. The enforcement landscape further sharpens these points. Shanghai’s internet regulator reported on June 26th that it had removed over 14,000 non-compliant AI agents, citing impersonation of official entities, inappropriate role-playing, and unauthorized data collection.

Whether this regulatory direction is optimal depends on which facet of the rulebook is examined. The safety provisions address documented harms that are largely unregulated elsewhere, ranging from adolescent emotional attachments to chatbots to companion apps that collect sensitive personal data. China’s official interpretation draws parallels with international actions, referencing lawsuits against Character.AI concerning psychological harm to teenagers, FTC investigations into companionship services, and European actions against Replika.

Conversely, the control aspect provides Beijing with a mechanism to influence the content generated by these systems, framed within the language of user protection. Both aspects are significant, and governments worldwide observing this regulatory experiment will need to determine which elements they are willing to adopt. Pan Helin, a member of the MIIT expert committee, articulated the official rationale to the South China Morning Post, stating that “current agents are not yet mature” and framing the policy around safety and standardization.

For now, technology companies have chosen the most prudent path: deactivating the affected components and subsequently working to develop compliant alternatives.