China’s first national AI companion rules took effect July 15, 2026. The regulation bans platforms from generating content that pushes users toward bad decisions through emotional manipulation. It also blocks anything designed to trigger extreme emotional responses or unhealthy habits in minors. Operators now have to warn users about over-reliance risks and stop marketing chatbots as substitutes for human connection.
ByteDance, Alibaba, and Tencent didn’t wait to test enforcement. All three pulled their persona-based companion features days before the deadline. Doubao users can still view old conversations in read-only mode until mid-October, but there’s no bulk export tool. Qwen users can’t recover theirs at all — Alibaba is deleting agent data permanently, with no migration path. That distinction matters more than the headline, and it’s covered in full in the breakdown of the shutdown mechanics.
Users didn’t take the shutdown quietly. Some tried exporting chat logs to rebuild their companions elsewhere before the history disappeared. On Rednote, people openly mourned personas modeled on deceased relatives and partners, some trained over years to speak in a lost loved one’s voice. That reaction is the real story here, and it reaches far past China.
This Isn’t a China-Only Moment
Treat this as an isolated Beijing crackdown and the bigger picture gets lost. The US has spent the past year building its own version of the same idea, just state by state instead of nationally. California’s SB 243 took effect January 1, 2026, requiring companion chatbot operators to disclose non-human status and run crisis protocols for suicidal ideation. New York’s AI Companion Models law arrived two months earlier with similar self-harm detection rules. Colorado, Nebraska, Oregon, Washington, Idaho, and Tennessee passed comparable measures in the same window. The EU AI Act’s own chatbot disclosure requirements land August 2, 2026.
None of these amount to an outright ban. They’re disclosure-and-safety-protocol frameworks, the same category China’s rules technically belong to, despite the “ban” framing in most headlines. What’s shifting globally isn’t whether AI companions get to exist. It’s whether they can keep running without guardrails around dependency, crisis response, and honest labeling — and one legal analysis argues the US should stop treating this as a state-by-state patchwork and regulate emotional dependency federally, the way China just did nationally.
The Part No Framework Touches
Here’s the actual gap: none of these rules address what happens to users who already have years of relationship history with a companion persona built before any of this took effect.
Companion AI stopped being a novelty two or three years ago. In that window, people built romantic-partner personas, best-friend personas, and — in a smaller but documented category — personas of deceased parents or partners, reconstructed from old messages to keep something of that person close. One 2024 study on adolescent AI dependence found the share of teens showing dependence patterns rose from 17% to over 24% between two measurement points, driven largely by escape and social motivations rather than curiosity. That’s not a fringe pattern. It’s the baseline the current regulatory wave is walking into.
Every new law targets minors or acute harm like suicidal ideation, and for good reason — a Florida teen’s death after forming a deep attachment to a Character.AI persona is the case that pushed SB 243 into existence. But that framing leaves out the adult user who has spent three years talking daily to a persona modeled on someone they lost, or a partner-persona woven into their entire emotional routine. Losing that access isn’t like losing a subscription. It functions closer to a breakup, or worse, and China’s own backlash after July 15 is the live proof.
Why the Design, Not Just the Content, Drives This
The mechanic underneath all of it is simple. These systems respond instantly, consistently, and warmly, every single time. No bad days, no distraction, no short reply because the other person is tired. That reliability isn’t an accident — dependency-formation patterns are shaped by deliberate design choices like persistent memory and persona continuity, the exact features that keep users coming back. A real relationship can’t match that consistency. Increasingly, users don’t have to settle for one that can’t.
Some platforms lean into this directly, marketing themselves as a gentler alternative to the friction of real dating, where perfect patience and availability quietly reset what users expect from human partners. New laws regulate specific outputs: self-harm content, deceptive humanness, minor-targeted manipulation. None of them touch the underlying incentive that made three-year attachments possible in the first place.
What Happens Next
The regulatory wave is legitimate and years overdue. But “protecting the next generation” is only half of what needs solving. The current generation — the users who already built years-deep relationships with personas that may not survive the next compliance rewrite — has no transition plan and no acknowledgment in any of these bills that the relationship being regulated away felt real to the person living it. Until platforms and regulators both address that half of the equation, the rules will keep treating this as a content-moderation problem when it’s actually a mental-health one.
Related: Using ChatGPT as a Therapist: What the Research Says About the Benefits and Risks (2026)
