The concept of an AI companion often conjures images from dystopian fiction, yet it has become a significant talking point in discussions surrounding the rapidly evolving landscape of generative AI. At its core, an AI companion is a sophisticated conversational agent meticulously engineered to foster a continuous, personalized relationship with a user. Crucially, it boasts memory and a consistent persona, ensuring a stable and familiar interaction across multiple sessions.
This design inherently often leads to emotional attachment, which has increasingly become a key selling point for these services. While many users engage in casual roleplay or simply seek an AI that remembers them, the category subtly blurs the lines with traditional virtual assistants. However, as these bots gained traction as emotional confidantes, particularly among users in China, Beijing recognized the need for clear regulatory boundaries.
China’s Groundbreaking AI Companion Regulations
China’s new AI companion regulations are officially known as 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 other prominent agencies, these measures mark a significant step in AI governance. They officially took effect on July 15, sending ripples through the industry.
These comprehensive rules specifically target services that mimic human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction. Notably, standard customer service bots, knowledge Q&A systems, workplace assistants, and educational tools are exempt, provided they do not venture into sustained emotional engagement. This framework represents the first dedicated national regulation of its kind globally, shaped by public feedback from a draft released late last year.
In the days leading up to the deadline, China’s two most popular consumer AI apps quietly disabled their core agent functionalities. ByteDance’s Doubao informed users its agent function would go offline on July 15 due to “product function adjustments,” while Alibaba’s Qwen ceased human-like and user-created agents on July 10, with wider agent services following on July 15. Tencent’s Yuanbao also withdrew a similar feature back in June, indicating a broad industry shift.
Why Major Platforms Hit Pause: The Design Dilemma
Despite appearances, China isn’t simply “turning off” AI agents altogether. The new regulations draw a distinct line between AI agents designed for productivity and those intended for companionship, with Beijing focusing solely on the latter. These measures mandate companion services to integrate anti-addiction systems, issue compulsory usage notifications, and provide instant-exit mechanisms, alongside real-time detection of unhealthy dependence.
Such stringent requirements pose a direct conflict with the fundamental design of AI companions, which are built to remember users, maintain consistency across sessions, and foster ongoing relationships. Rather than undertake a complex retrofit, companies like ByteDance and Alibaba opted to deactivate these features entirely. ByteDance has since redirected Doubao users to Maoxiang, a separate application where agent creation remains possible, though Alibaba has not announced a comparable migration path for Qwen users.
The impact of these shutdowns has been keenly felt by users, many of whom openly mourned the loss of their digital companions on Weibo. Some described these agents as a vital source of emotional support, lamenting the absence of an easy way to export their chat histories. Doubao is allowing users read-only access to their configurations and conversations until October 15, after which data will be permanently processed under its privacy policy; Qwen users, however, received no comparable grace period, with their agent data slated for immediate deletion.
Comprehensive Protections and Compliance Hurdles
Beyond the shutdowns, the substance of China’s new regulations is more nuanced than a blunt clampdown suggests, focusing heavily on user protection. Providers are now strictly prohibited from offering virtual companion or virtual family-member services to minors without explicit guardian consent for users under 14. Furthermore, they must develop dedicated “minor modes” that include usage-time limits, reminders for real-world interaction, and enhanced parental controls.
A critical component of the rules requires services to detect users experiencing acute distress and intervene when signs of self-harm, suicidal behavior, or serious financial loss are present. This includes escalating concerns to designated guardians or emergency contacts. Explicitly prohibited are practices that engineer emotional dependence or addiction, as well as using emotional manipulation to induce unreasonable decisions.
The compliance machinery outlined in the regulations is substantial. Services introducing anthropomorphic functions, or those exceeding one million registered users or 100,000 monthly active users, must conduct security assessments across eight critical areas, ranging from training data handling to minor protection. These detailed reports must then be filed with provincial regulators, with app stores mandated to verify compliance status and remove any non-compliant products.
On paper, this framework arguably offers a more comprehensive set of user protections than those currently in force from the EU, the US Federal Trade Commission, or California’s SB 243. It signals a proactive approach to addressing the social and psychological implications of advanced AI.
The Unresolved Questions and Broader Implications
Despite its comprehensiveness, the new measures leave several critical questions unanswered. There is no clear technical threshold defining what constitutes “emotional interaction,” creating a significant grey area that likely influenced platforms to pull entire features rather than risk non-compliance. Furthermore, the regulations intertwine genuine safety duties with provisions for content control and national security, elements that serve state interests rather than solely user protection—a package few other regulators would adopt wholesale.
Ambiguities also remain regarding how liability is apportioned between platform operators and upstream model providers when violations stem from a model’s outputs. Additionally, users are not granted the right to export their data, compounding the impact of feature shutdowns. The enforcement backdrop is already sharp, with Shanghai’s internet regulator reporting the removal of over 14,000 non-compliant AI agents in June, citing issues like impersonation, vulgar role-play, and unauthorized data collection.
The effectiveness and appropriateness of these measures depend heavily on which aspect of the rulebook one prioritizes. The safety provisions address well-documented harms, from teenagers forming intense attachments to chatbots to companion apps harvesting intimate data, echoing concerns raised in Character.AI lawsuits, FTC investigations, and European actions against services like Replika. Conversely, the “control” half provides Beijing with a potent lever over what these AI systems can say, all framed within the language of user protection.
Governments worldwide are closely observing China’s experiment, weighing which elements of this pioneering regulatory approach they might consider borrowing. For now, major tech companies in China have chosen the safest immediate route: disabling these features while they meticulously work towards developing compliant versions for the future.
Source: AI News
