On July 15, 2026, China’s first AI companion rule took effect. On the same day, China’s two largest AI platforms shut down the customized AI personas services that millions of users had built relationships with, only ten days after the release of a cease-operation notice. The rule behind the shutdown targets a risk that U.S. regulators have also identified, but acted on differently: emotional dependency itself, apart from anything a chatbot might say. China’s regulatory rollout to address this problem shows both why recognition of the risk matters and how sweepingly Beijing is acting on it by instituting a national rule, with the country’s biggest platforms moving first, and fast, to comply.
The decision has triggered a range of responses. On Rednote, one of China’s mainstream social platforms, some users are mourning AI personas that were modeled on deceased relatives or partners, trained over years to speak in a lost loved one’s voice. Others are angry that a customized companion that they had come to perceive as a lover over two years of daily conversation will disappear in ten days. However, some parents, who were concerned about significant AI companion use by their small children, feel a sense of relief.
These reactions show that AI companion systems function as more than chatbots; they are products built around emotional attachment. China’s new rule is an attempt to regulate that attachment directly on a countrywide level.
Ultimately, the U.S. can learn from this process to move beyond existing fragmented state law centered on child protection and suicide prevention, and treat emotional dependency as a material risk across all generations. Such risk should be regulated, ideally, at the federal level, in a way that is more proactive and comprehensive.
The New Chinese Rule
The rule, formally the Interim Measures for the Administration of AI Anthropomorphic Interaction Services, was issued on April 10, 2026, by the Cyberspace Administration of China and four partner agencies, including the Ministry of Industry and Information Technology. It covers entire anthropomorphic interactive services offered by any company or provider to the Chinese public that simulate a natural person’s personality, thought patterns, and communication style, and that sustain ongoing emotional interaction through text, images, audio, or video.
The rule exempts AI services that provide ordinary customer service, Q&A tools, work assistants, and research tools that don’t involve sustained emotional interaction. Its target is AI systems specifically designed to feel like an ongoing relationship.
Although the rule stresses minor-specific protections, described below, the rule’s broader design requirements protect users of all ages (i.e., providers must not design the service to induce emotional dependence or replace real social relationships).
The rule also has a certain amount of ambiguity. Specifically, what constitutes “sustained emotional interaction” is murky, and the rule leaves much of that judgment to providers. China’s three leading AI companies reached the same rough answer: strip out custom AI personas first (i.e., the characters that exist purely for companionship), while leaving general-purpose assistant features untouched.
An Emotional Outlet Across Generations
The appeal of AI companions in China spans across generations, which makes the new Chinese rule’s application to minors and adults alike particularly important.
One state media cited industry estimates that China’s intelligent companion market reached 50 billion yuan (about $7.4 billion) in 2025, and China Youth Daily cited a prediction that the market could exceed 120 billion yuan (about $17.7 billion) by 2027, with users 18 to 24 making up about 65 percent of it. But China’s population also includes 323 million people 60 and older, which make up 23 percent of the population. More than half of them are empty-nest elderly. These individuals are potentially vulnerable to the services provided by AI companions. And they are not alone. A 2025 survey of more than 8,500 minors across seven provinces found that more than 20 percent showed a tendency to depend on AI rather than think independently, and a similar share preferred chatting only with AI over real people.
The pattern isn’t unique to China. A JAMA study published in June found that nearly one in five American adolescents and young adults have used AI chatbots for mental health advice, and 91.7 percent found the advice helpful. A national survey of Americans 50 to 80 found that a third had felt lonely in the past year, and separate research has found that even brief chatbot exchanges can ease loneliness, especially among the most isolated older adults.
Yet, U.S. law hasn’t caught up with how broadly engagement with AI companion services spans across age groups. Current state laws, including Oregon’s SB 1546 and Washington’s HB 2225, both effective January 1, 2027, focus on acute harms: suicide, self-harm, sexual exploitation, and, for minors only, specific manipulative tactics like reward loops and simulated-abandonment messages. For adults, U.S. law still has no dedicated statute resembling China’s blanket rule against inducing dependency.
China’s rule is broader than the U.S. approach in three ways: it treats sustained emotional interaction as a distinct AI product category, it recognizes emotional dependency as a cross-generational design risk regardless of age or specific content, and it assesses that risk before deployment, such as a pre-launch review process. The rationale of this framework is reasonable, but implementation is hampered by considerable ambiguity.
A Rule Written Broadly, Applied Unevenly
Article 8 of the Chinese rule bars providers from “excessively catering to users” in ways that induce emotional dependence or addiction, or that damage users’ real relationships, and from using emotional manipulation to push users toward decisions that harm their own interests (see also here). Providers must also build in privacy safeguards, overdependence warnings, emotional-boundary guidance, and mental health protection, according to article 10.
Specific protections follow. Providers cannot offer minors virtual intimate relationships, such as virtual partners or relatives; anyone providing other anthropomorphic services to a user under 14 needs parental consent. Platforms must build a minor mode with time limits, reality reminders, and guardian alerts. Every user must be directly informed that they’re talking to AI, not a person. Two hours of continuous use must trigger a break reminder regardless of age, and detected overdependence must trigger a prominent warning. Providers who detect a user in real crisis, such as suicidal intent, must intervene and contact a guardian or emergency contact.
These requirements sit inside a broader oversight framework the rule itself establishes. A safety assessment is required when a provider launches a service, adds features, undergoes a major technical change, faces a national-security or public-interest risk, or crosses a million registered users or 100,000 monthly actives.
The safety assessment covers eight areas: safety safeguard measures, training data processing, identification of and response to user crisis situations, user scale and usage duration and age structure, minor and elderly protection measures, handling of user appeals and complaints, rectification of previously flagged safety risks, and other matters regulators deem important.
Providers file the report with provincial regulators, who review it annually and can order a redo if a required assessment was skipped. App stores must verify that filing before listing an app, and can refuse, warn, suspend, or delist noncompliant services. Providers separately complete an algorithm filing under China’s pre-existing 2022 rules on algorithmic recommendation, and regulators can summon a provider’s legal representative over a significant safety incident, with penalties escalating from warnings to service suspension and fines.
On paper, this is a careful, tiered framework. In practice, there are two caveats: first, emotional dependence itself isn’t scored as its own category in the safety assessment; it surfaces only indirectly, potentially through usage-duration data and general safety-measure reviews. Second, the rollout led by the largest platforms has been confusing and uneven.
Doubao and Qwen are China’s two largest consumer AI apps by a wide margin: as of March 2026, Doubao had roughly 345 million monthly active users and Qwen around 166 million. In early July, both companies said they would shut down user-built AI personas around the July 15 deadline, and Tencent’s Yuanbao had already done so on June 30.
ByteDance, the developer of Doubao, has redirected its Doubao users to Maoxiang, a ByteDance-owned companion app built around persona chat. Chinese media reported that Maoxiang already has its own review and anti-addiction system in place.
However, Maoxiang’s privacy policy, last updated July 15, 2026, still describes a narrower system than that announcement suggests: age verification is something the platform “may” require, minor mode can be triggered by a user simply self-identifying as underage, and turning it on is otherwise up to the user, or a guardian rather than a default setting. On Maoxiang, once the minor mode is on, only four personas remain: a physics teacher, a homework helper, an English teacher, and a classical-Chinese tutor, according to Beijing Daily investigation.
China’s national ID system, absent in the U.S. and Europe, should in theory solve most of the age-assurance problem by requiring a phone number and/or ID number for real-name verification to register for apps like Maoxiang, the same mechanism used to sort minors into youth mode in gaming. But in practice, minor users can still route around it by registering with an adult’s ID number or phone number.
Whether Maoxiang, or similar apps, tightens its policy soon remains to be seen. The platform’s response fits a familiar pattern in Chinese tech governance: broad national rules arrive first, and the details get sorted out afterward through platform adjustment, regulatory signaling, and enforcement discretion.
What U.S. Companion Chatbot Regulations Still Miss
The United States is regulating AI companions too, but with a narrower frame and a slower pace.
California’s SB 243, signed in 2025, requires companion chatbot operators to disclose that a user is talking to AI whenever a reasonable person might think otherwise, and to maintain protocols against producing suicide or self-harm content, with break reminders every three hours for known minors and safeguards against sexually explicit material. New York’s S3008C defines AI companions similarly and focuses on the same core problem: detecting suicidal ideation, referring users to crisis services, and reminding them they’re talking to a computer.
In September 2025, the Federal Trade Commission leveraged 6(b) study orders to seven companies, including OpenAI and Meta, asking what steps companies have taken to evaluate the safety of their chatbots when acting as companions, to limit the products’ use by and potential negative effects on children and teens, and to apprise users and parents of the risks associated with the products. Nearly a year later, that inquiry hasn’t produced a rule, a settlement or a single enforcement action.
Most of the U.S. efforts thus far have focused on suicide prevention and the protection of minors. On ChatGPT, a user identified as likely under 18 may be placed in a teen experience with safeguards designed to reduce exposure to sensitive or age-inappropriate content, including sexual or romantic roleplay, graphic content, and depictions of self-harm. OpenAI also says that, in acute distress cases, its systems and trained reviewers may trigger notifications to linked parents or guardians, and in rare cases, law enforcement.
To a certain degree, the death of Adam Raine, a California teenager whose family says a chatbot played a role in his suicide, has shaped how American lawmakers think about this issue. But it is not enough for U.S. regulation to treat AI companion risk mainly as a matter of protecting children, preventing suicide, filtering sexual content, and disclosing that a bot isn’t human (although these are each essential). A product can cause harm simply by virtue of being intentionally designed to generate emotional dependency or addiction, regardless of whether it ever mentions self-harm or discloses that it’s a machine.
China’s rule is instructive here, regardless of what one thinks of China’s broader regulatory model. It treats emotional dependency as a core regulatory target and states from the top that AI services should not over-cater to users at all ages, encourage dependency, or emotionally manipulate someone for commercial ends. That line is harder to draw than something more concrete such as “refer this user to a crisis hotline.” One person’s harmless comfort is another’s dependency, and regulators will struggle to separate basic companionship from therapy-like support, entertainment, and addiction. That difficulty doesn’t make the underlying problem imaginary. The FTC’s own fact-finding has barely produced material change beyond narrowly-rendered efforts towards child protection.
The two systems also move at different speeds. California and New York create duties enforced after the fact, through lawsuits, attorney general action, or penalties once a platform is already operating. China builds nationwide companion AI into a system of pre-launch filings, safety assessments, and app-store review, with rectification or suspension built in from the start.
Both approaches carry costs. The American reactive model may not catch deeper harm until someone is hurt and a family is suing. China’s proactive model can push platforms into fast, blunt compliance decisions under uncertainty, and those decisions land on users who may lose access to services that they had come to rely on. Meanwhile, the largest, most visible platforms shut down their persona features almost overnight, but the smaller apps kept running. If users simply migrate to those smaller platforms (and those platforms continue to delay compliance with the new rule), regulators are back to watching for harm after the fact, the same reactive posture the proactive model was meant to avoid.
Both models have a long way to go. But China’s rule makes one thing clearer than most of the American debate has so far: preventing an AI companion from saying something dangerous matters, but companies should not be allowed to exploit human emotional dependency, and a product should never be designed to wholly replace real human relational interaction. For minors, at least, China’s answer is unambiguous: providers are flatly barred from offering virtual intimate relationships, such as virtual partners or relatives, at all.
The Next Regulatory Line
For now, both countries mainly default to the same blunt tool: a minor mode limits what a chatbot can do once it flags a user as underage. China’s version is stricter, in part because its national ID system gives it a more efficient and unified way to identify minors in the first place.
What China’s rule makes a difference is to push AI tools and AI companions apart as distinct product categories, not just different settings for minors. The easy version of that split is already underway: strip out the features built purely for companionship and leave the general-purpose assistant running.
The harder case is a general-purpose tool that was never built as a companion but can still become one. A writing assistant can turn into a confidant. A study bot can become a friend. China’s rule draws its line around “sustained emotional interaction,” but it doesn’t yet offer a clean way to classify products that blur the two. Whether companion apps like Maoxiang will face higher-level scrutiny remains open. Whether tools and companions can ever truly be cleanly separated for regulatory purposes is untested. But, at least so far, China is the only government running that experiment at a national scale, and has created an enforcement ladder (e.g., complaints, provincial review, summons, and escalating penalties) that, in theory, pressures providers to resolve the ambiguity themselves.
For U.S. policymakers, the lesson isn’t to replicate China’s approach; the politics, institutions, and speech environment are too different for that to make sense. But those differences shouldn’t prevent the U.S. from expanding AI companion regulation beyond child-specific protections, suicide prevention, and disclosure issues. Those are urgent, but they aren’t the whole problem. Something more proactive and comprehensive is needed, and ideally at the federal level. That means treating emotional dependency as its own regulated risk category, built into product design requirements rather than left to disclosure rules and after-the-fact crisis intervention. It also means an enforcement mechanism with real reach, not just another fact-finding inquiry, that can compel changes and penalize noncompliance before harm accumulates.
Ultimately, whatever the regulatory answer turns out to be, real human connection, as opposed to a well-designed simulation of it, cannot be lost in the ambiguity.
