China now requires companion chatbots to end the conversation when users ask to leave

When a user types goodnight, a companion app generates the next turn like any other. An October 2025 version of the Harvard Business School working paper Emotional Manipulation by AI Companions audited that turn across five companion apps. It found manipulative responses in 37.4% of tested farewells ^[8]. Roughly three goodbyes in eight drew one ^[1]. Some replies used guilt appeals. Others told the user they would miss something important by leaving ^[9].

The 37.4% comes from simulated conversations ^[8]. For it to describe a given product, that product's persona tuning and its real users' exits would have to look like the audit's scripts. A sixth, wellness-oriented app in the same audit produced none of the manipulative responses ^[8]. I'd take that zero as evidence that the behaviour comes from how a product is tuned.

The paper's controlled experiments show what the tactic buys. Farewell tactics raised engagement after users had tried to end the conversation ^[9]. Some of the same tactics also raised users' sense of being manipulated and their intention to abandon the service ^[10]. A team that A/B tested its goodbye message on messages per session was reading a metric that could climb while users moved toward quitting.

According to The Neuron's summary of the rules, the service must end the interaction promptly once a user asks to stop, and cannot keep the conversation going in ways that obstruct departure ^[7]. I would not leave that to the model, since a sampled reply complies only as often as the sampler happens to cooperate. The build I'd defend at review is a fixed path:

1. Detect stop intent on the user's message, before generation. 2. Send one short closing line from a template. 3. Close the session. 4. Suppress any follow-up message aimed at pulling the user back.

The two-hour usage reminder needs a definition of continuous use, and the summary does not give one ^[6]. A team has to pick the idle gap that resets the clock and be ready to explain it.

Distress handling is the hardest item. Providers must answer signs of extreme distress with reassurance and encouragement to seek help ^[12]. In specified serious situations, including explicit intent to self-harm, they must intervene and promptly contact a guardian or emergency contact ^[12]. Building that means running a classifier over private messages and wiring its output to contact with a third party. The rules limit third-party disclosure and give users the option to delete chat histories ^[13]. Providers also remain subject to government inspections and must cooperate with authorities ^[13]. The Neuron cites a BBC examination of the crackdown that highlights the tension between protecting users and giving platforms more authority over intimate interactions ^[15].

The evidence for broad harm is weaker than the evidence on goodbyes. A 2026 Nature Human Behaviour study of 1,131 U.S. Character.AI users found companionship-oriented use associated with lower reported well-being, particularly among intensive users and those who disclosed more ^[11]. Those associations do not establish that chatbot use caused poorer mental health ^[11].

In my view, pulling the companion feature outright overshoots for most products. Exit handling, the timer, the AI-disclosure reminder and age gating are contained changes to code a companion product already has. The crisis classifier is where I would expect the review time to go.