AI companions may sound dystopian, but they are a common theme in broader conversations about the dangers of generative AI. What this refers to is essentially a conversational agent built to maintain an ongoing personal relationship with the user, with a memory and stable persona that remains consistent from session to session.
Their design often creates an emotional attachment, which is increasingly becoming a selling point. A lot of the uses are for casual roleplay, or when you just want something to remember yourself, and at the end the category changes to regular assistant. But as more and more people in China treat these bots as emotional friends of sorts, the Chinese government has now decided that this practice needs rules.
China’s AI Companion Regulation took effect on July 15, and just days before the deadline, two of the country’s most used consumer AI apps quietly turned off core features. ByteDance’s Doubao told users that agent functionality would be taken offline on July 15, citing “adjustments to product features,” while Alibaba’s Qwen said human-like user-created agents would go offline on July 10, and broader agent services would stop five days later.
A quick read shows that China is eliminating AI agents. it’s not. The rule draws a line between agents who do your work and agents who accompany you, and it is only the second type that the Chinese government objects to.
This regulation is the “Interim Measures on AI Anthropomorphic Interactive Service Management” jointly issued by the Cyberspace Administration of China and four partner agencies (National Development and Reform Commission, Ministry of Industry and Information Technology, Ministry of Public Security, and State Administration of Market Regulation) on April 10, 2026.
This includes services that simulate human personality traits, thought patterns, and communication styles to provide sustained emotional interactions. Customer service bots, knowledge Q&A, workplace assistants, and educational and research tools are excluded as long as they avoid sustained emotional involvement. This is the first dedicated national framework of its kind and was fleshed out following a draft public comment session late last year.
Design problem, not prohibition
Dubao and Kwen did not violate the ban. They ran into a design contradiction. This measure requires companion services to run anti-addiction systems, issue mandatory use notices, detect unhealthy addictions in real-time, and provide immediate termination mechanisms.
These demands are a nuisance for agents built to remember users, be consistent across sessions, and maintain ongoing relationships, and ByteDance chose to shut down this feature rather than renovate it. Alibaba appears to have made a similar call. ByteDance is currently directing Doubao users to Maoxiang. Maoxiang is another app that allows you to recreate agents. Alibaba announced that there is no equivalent migration path for Qwen. Tencent’s Yuanbao discontinued a similar feature in June.
That cost falls on the user. Many publicly lamented the closure on Weibo, with one post describing agents as long-time moral support and lamenting the lack of an easy way to export chat history. Doubao will allow users to view their configurations and conversations in read-only mode until October 15 of this year, until the data is processed in accordance with its privacy policy and becomes irrecoverable. Qwen users are not given a comparable grace period and their agent data is set to be permanently deleted.
What China’s AI Companion Regulations stipulate
This reality is more considered than the blatant crackdown would suggest. Providers are prohibited from providing virtual companion or virtual family services to minors and must obtain parental consent before providing services to users under 14 years of age. Providers should build a dedicated “minor mode” with screen time limits, reminders to return to real-world interactions, and enhanced parental controls.
We must also identify users in severe distress, intervene and escalate users who exhibit signs of self-harm, suicidal behavior, or significant financial loss to a designated parent or emergency contact. Manipulating emotional dependence or addiction, or manipulating emotions to induce irrational decisions, is expressly prohibited.
Compliance machines are heavy. Services that launch anthropomorphic features or exceed thresholds of 1 million registered users or 100,000 monthly actives must perform a security assessment that covers eight areas, from processing training data to minor protections, and submit a report to state regulators. App stores must review their status and remove non-compliant products.
In theory, this is a more complete set of user protections than what the EU, US Federal Trade Commission, or California’s SB 243 have not yet put into effect.
What the rules leave unresolved
It is equally important that the measures do not solve the problem. They haven’t set a technical threshold for what counts as an emotional interaction, and that gray area is exactly why the platform pulled the entire feature without risking going in the wrong direction. These are not packaged wholesale by other regulators, with real safety obligations built into content moderation and national security provisions that address nations rather than users.
It also leaves open how liability is divided between platform operators and upstream model providers when violations result from model outputs, and does not give users the right to exfiltrate their data. The context of enforcement makes the point clear. Shanghai’s internet regulator announced on June 26 that it had removed more than 14,000 non-compliant AI agents for impersonating public authorities, vulgar role-playing and illegally collecting personal data.
Whether this is the right direction depends on which half of the rulebook you read. Safety Half addresses harms that have been documented but are largely unregulated elsewhere, from teenagers forming attachments to chatbots to companion apps collecting intimate data. China’s own official interpretation seeks support abroad, citing the Character.AI lawsuit over psychological harm to youth, the FTC investigation into companion services, and the European lawsuit against Replika.
Regulators are partially handing the levers to the Chinese government over what these systems say in the same language of user protection. Both are real, and the government overseeing the experiment will have to decide which parts to borrow. Pan Helin, a member of the MIIT expert committee, articulated this official lawsuit to the South China Morning Post, stating that “the current staff is not yet mature” and has framed policies around safety and standardization.
For now, the companies are taking the safest course: powering down the components and figuring out what a compliant version might look like later.
SEE ALSO: Meta revamps AI chatbot policy amid child safety concerns
Want to learn more about AI and big data from industry leaders? Check out the AI
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
