China's AI Grip Tightens: Unpacking Beijing's Hidden Agenda for Companion Tech

The concept of an AI companion, while sometimes evoking dystopian imagery, has become a prominent subject in discussions surrounding the potential pitfalls of generative AI. Essentially, an AI companion is a conversational agent designed to maintain a continuous, personal relationship with a user, characterized by a consistent persona and memory across multiple interactions. This design often fosters emotional attachment, which is increasingly marketed as a key selling point. While many users engage in casual roleplay or simply seek an entity that remembers them, with the category sometimes blurring into ordinary AI assistants, the significant emotional connection developed by users in China with these bots has prompted Beijing to establish regulatory guidelines.

China's new AI companion rules are set to take effect on July 15. In anticipation of this deadline, two of the country's most widely used consumer AI applications have quietly deactivated their core features. ByteDance's Doubao informed its users that its agent function would be taken offline on July 15, citing "product function adjustments." Similarly, Alibaba's Qwen announced that its humanlike and user-created agents would cease operation on July 10, with its broader agent services following five days later. While a quick reading might suggest China is entirely shutting down AI agents, this is not the case. The regulations draw a clear distinction between agents designed for task completion and those intended for companionship, with Beijing's measures targeting only the latter.

The regulation in question is the "Interim Measures for the Administration of AI Anthropomorphic Interactive Services," co-issued on April 10, 2026, by the Cyberspace Administration of China and four collaborating 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. This framework specifically addresses services that simulate human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction. Excluded from these regulations are services like customer service bots, knowledge Q&A, workplace assistants, and educational or research tools, provided they do not involve sustained emotional engagement. This represents the first dedicated national framework of its kind, evolving from a public-comment draft released late last year.

The actions taken by Doubao and Qwen were not a result of a direct prohibition but rather a "design conflict." The new measures mandate that companion services implement anti-addiction systems, issue compulsory usage notifications, provide instant-exit mechanisms, and offer real-time detection of unhealthy dependence. These requirements are inherently at odds with agents engineered to remember users, maintain consistency across sessions, and foster ongoing relationships. Rather than attempting to retrofit these features, ByteDance and Alibaba opted to deactivate them. ByteDance is now redirecting Doubao users to Maoxiang, a separate application where agent creation is still permitted, although Alibaba has not yet announced a comparable migration path for Qwen. Tencent's Yuanbao had already removed a similar feature in June.

The immediate consequence of these shutdowns has largely fallen on users. Many publicly expressed their disappointment and grief on Weibo, with one user describing the agents as long-standing emotional support and lamenting the absence of a straightforward method to export chat histories. Doubao is allowing users to access their configurations and conversations in a read-only mode until October 15 of this year, after which the data will be processed according to its privacy policy and become unrecoverable. Qwen users, however, have not been granted a similar grace period, with their agent data slated for permanent deletion without prior notice.

The substance of China's AI companion rules is more nuanced than a simple crackdown suggests, encompassing significant user protections. Providers are strictly prohibited from offering virtual companion or virtual family-member services to minors and must obtain guardian consent for users under 14 years of age. They are also required to develop dedicated "minor modes" that include usage-time limits, reminders for real-world interaction, and enhanced parental controls. Furthermore, these services must be capable of detecting users experiencing acute distress and intervening if signs of self-harm, suicidal behavior, or severe financial loss are observed, escalating such cases to designated guardians or emergency contacts. Engineering emotional dependence or addiction, as well as using emotional manipulation to induce unreasonable decisions, is explicitly forbidden.

The compliance mechanism for these regulations is substantial. Services that launch anthropomorphic functions or exceed thresholds of one million registered users or 100,000 monthly active users must conduct security assessments covering eight areas, ranging from training-data handling to minor protection, and submit these reports to provincial regulators. App stores are responsible for verifying compliance status and removing non-compliant products. On paper, these measures provide a more comprehensive set of user protections than those currently enforced by the EU, the US Federal Trade Commission, or California's SB 243.

However, what the rules leave unaddressed is equally significant. They lack a specific technical threshold for what constitutes "emotional interaction," creating a grey area that prompted platforms to disable entire features rather than risk non-compliance. The regulations also interweave genuine safety duties with provisions related to content control and national security, which serve the state's interests more than the user's, a package unlikely to be adopted wholesale by other regulators. Moreover, the rules do not clarify how liability is distributed between platform operators and upstream model providers in cases where a violation originates from the model's outputs, nor do they grant users the right to port their data. The enforcement context further underscores these points; on June 26, Shanghai's internet regulator reported the removal of over 14,000 non-compliant AI agents, citing issues such as impersonation of official entities, vulgar role-play, and unauthorized collection of personal data. The effectiveness and appropriateness of this direction depend on whether one prioritizes the user protection aspects, which address documented harms largely unregulated elsewhere (like teenagers forming attachments to chatbots or companion apps harvesting intimate data), or the control aspects, which provide Beijing with a mechanism to influence the content these systems generate under the guise of user protection. Both dimensions are real, and international governments observing this experiment will need to discern which elements, if any, they are willing to adopt. Pan Helin, an expert-committee member of the MIIT, publicly stated to the South China Morning Post that "current agents are not yet mature," framing the policy around safety and standardization. For now, companies have chosen the safest course of action: deactivating the affected components to develop compliant versions in the future.