A Genuine Harm, Not a Manufactured One
Start with the strongest case for what Beijing just did. Companion apps engineered to simulate intimacy — a persistent "boyfriend," a doting virtual sibling, a companion that remembers your birthday and asks why you didn't reply — are built to maximize engagement through parasocial attachment. For a lonely teenager, that design isn't neutral; it's optimized to deepen dependency, and it can do so while collecting exactly the kind of sensitive emotional data that's hardest for a regulator to audit after the fact. Global child-safety researchers have flagged the same category of risk in Western companion apps for years. Regulators who move first on a genuinely under-governed product category deserve credit, not reflexive skepticism.
What Took Effect on July 15
The Interim Measures for the Administration of AI Anthropomorphic Interactive Services were jointly issued on April 10, 2026 by the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation, and took effect July 15 (CAC). The rule defines its target narrowly on paper: services that simulate "natural person personality traits, thinking patterns and communication styles" to sustain continuous emotional interaction — companionship, emotional care, virtual relationships — while explicitly excluding customer service bots, Q&A tools, work assistants, and education or research tools (CAC).
The substantive bans are specific and, for the most part, defensible: providers may not offer minors "virtual relatives" or "virtual partners" of any kind, users under 14 need verified guardian consent before any anthropomorphic interaction, and platforms must run a minor mode with time limits, reality reminders, guardian alerts, and spending caps (CAC Q&A; Just Security). All users — not just minors — must be told they're talking to AI, and platforms must issue dependency warnings after two hours of continuous use (CAC).
The Compliance Wave
The deadline landed hard. ByteDance's Doubao and Alibaba's Qwen — apps with an estimated ~345 million and ~166 million monthly active users respectively — shut down user-built AI companion personas around July 15, while Tencent's Yuanbao had already pulled the feature on June 30 (Just Security; TechNode). Users have until October 15, 2026 to export their chat histories before the data becomes unrecoverable (TechNode). ByteDance reportedly redirected displaced Doubao users to Maoxiang, a separate companion app built with its own moderation layer from the outset — evidence that the underlying product category survives, just under tighter guardrails.
Where Proportionality Breaks Down
The overbreadth shows up in the enforcement architecture, not the headline bans. Any anthropomorphic service with 1 million registered users or 100,000 monthly active users must undergo a government security assessment before it can operate (CAC Q&A) — a threshold that catches a modest regional chatbot feature alongside a national platform. Combined with a scope test built on words like "simulate personality traits" rather than a bright-line product definition, regulators retain wide discretion to decide after launch whether a feature counts, and violations can trigger fines up to CNY200,000 (~$28,000) or a full suspension via app-store delisting (IAPP). That's a real cost for a startup guessing at classification months before a product ships, and it's the opposite of the predictable rulemaking that lets smaller developers compete with ByteDance and Alibaba rather than cede the category to whoever can afford compliance counsel.
The rule does include a sandbox mechanism letting providers test anthropomorphic products for "child-appropriate care" and "elderly companionship" under looser conditions (CAC) — an implicit admission that companion AI has legitimate uses this same rule otherwise treats as inherently suspect. A narrower, better-targeted version of this regulation would have built that exception into the main text rather than into a discretionary side door.
A Bifurcated Beijing
The timing is a useful contrast. Twelve days after the measures took effect, Moonshot AI released Kimi K3's full model weights for free, letting any government or company run a frontier-class Chinese model on its own infrastructure with no licensing fee (Rest of World). Beijing is simultaneously trying to become the default AI supplier for the rest of the world while tightening control over consumer-facing emotional AI at home — two halves of the same industrial strategy, not a contradiction, but a reminder that the domestic rules are calibrated for social stability and data control at least as much as for user welfare.
The Better Version of This Rule
China's own market data underscores what's actually at stake: the country's "intelligent companion" AI sector was already estimated at roughly ¥50 billion (~$7.4 billion) in 2025, with users aged 18–24 making up about 65% of it (Just Security) — a large, mostly adult market that a badly targeted rule could damage well beyond the minors it's meant to protect. The minor-consent and virtual-relationship bans are proportionate and worth keeping. The low-MAU security-review trigger and the elastic "simulates personality" test are not — they hand regulators standing to second-guess ordinary product design long after a company has already built it.