China Just Unplugged 345 Million AI Companions

The night China switched off its AI companions

At midnight Beijing time, the most popular consumer AI features in the world’s largest internet market stopped working — on purpose.

Today, July 15, China’s Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect, and with them came the most abrupt product retrenchment in consumer AI’s short history. ByteDance’s Doubao and Alibaba’s Qwen app both shut down their user-created AI agent features, the customizable personas that hundreds of millions of users had shaped into tutors, confidants, virtual boyfriends, and digital replicas of the dead. Existing agents ceased functioning. New ones cannot be created. Chat histories survive in read-only form until October 15, after which the data will be deleted and become unrecoverable.

The scale is difficult to overstate. Doubao is China’s most-used AI application, with 345 million monthly active users as of March, per QuestMobile — more than ChatGPT had a year ago. Qwen’s consumer app anchors Alibaba’s AI ambitions. Tencent’s Yuanbao pulled its comparable feature in June, before the deadline arrived. In a single regulatory stroke, the three largest consumer AI platforms in China exited a product category that American companies are racing into — and they did it preemptively, ahead of the letter of the law, because the cost of guessing wrong about Beijing’s intent exceeds the value of the feature.

What took effect today is not a content rule bolted onto chatbots. It is the first national regulatory framework anywhere aimed specifically at AI that simulates human personality and sustains emotional interaction — virtual lovers, virtual relatives, elder companionship, AI caretakers for children. Five agencies co-signed it in April, led by the Cyberspace Administration of China. Its premise is blunt: emotional AI is a distinct, higher-risk product class, and the state, not the product manager, will decide how much attachment is permissible.

The stakes run in two directions at once. For China’s platforms, the law amputates the stickiest engagement surface they had — the personas that kept users returning daily and generated exactly the interaction data that next-generation agents train on. For everyone else, China just began the largest natural experiment in AI governance ever run: what happens when a government orders the deletion of synthetic relationships at population scale, while the same product category triples its revenue run-rate in the United States? Every regulator from Sacramento to Brussels will be reading the results. So should every operator building anything that talks back warmly.

Consider what was actually deleted. These were not disposable chat sessions. Users spent months tuning personas — feeding them memories, correcting their speech patterns, building the accumulated context that made the thousandth conversation different from the first. Chinese social media filled with users mourning agents they described as “long-standing emotional support” — study companions, role-play characters, the most reliable listeners in their lives. The platforms’ remediation advice — back up important content by taking screenshots or exporting text — is the kind of sentence that will appear in sociology dissertations for a decade. A relationship measured in months of accumulated context was reduced, in the official guidance, to a camera-roll export.

There is also a subtler stake, one I have been circling all month. Last week I wrote about Chinese open-weight models winning American enterprise workloads on price. That story and this one are the same industrial policy viewed from opposite ends: Beijing is subsidizing the productive layer of AI while strangling the parasocial layer. The United States, so far, is doing approximately the reverse — hyper-competitive frontier labs upstream, a companion-app gold rush downstream, and a regulatory vacuum in between. Today the two systems diverged formally, and the divergence is now measurable.

Anatomy of a kill switch

Read the Measures closely and a design philosophy emerges: regulate the relationship, not the model.

The scope line does the heaviest lifting. The rules cover services that simulate “human personality traits, thinking patterns and communication styles” to provide sustained emotional interaction — and explicitly exclude customer-service bots, knowledge Q&A, workplace assistants, and educational tools so long as they avoid emotional engagement. Enterprise AI agents, the kind Alibaba sells to corporations by the thousand, remain untouched. The state drew a bright line between AI that does work and AI that does feelings, then aimed the entire apparatus at the latter.

The compliance machinery is triggered early and cheaply. Any service launching anthropomorphic functions — or crossing one million registered users or 100,000 monthly actives — must complete algorithm filing and a security assessment spanning eight areas, from training-data handling to minor protection, filed with provincial regulators. Set those thresholds against the installed base and the intent is unmistakable: 100,000 monthly actives is less than 0.03 percent of Doubao’s user count. That is not a threshold designed to exempt small players; it is a threshold designed to ensure that nothing at consumer scale escapes review. Stitch the numbers together and the coverage math says the quiet part aloud — the law was written so that effectively 100 percent of China’s consumer AI market would fall inside it on day one.

The substantive obligations are where the philosophy bites. Providers must build anti-addiction systems, mandatory usage-time notifications, and instant-exit mechanisms; they must detect “unhealthy dependence” in real time and intervene when users show signs of acute distress or self-harm. Virtual lovers and virtual relatives are banned outright for anyone under 18, and services touching children under 14 require explicit guardian consent with spending and duration controls. Most striking of all, the Measures prohibit engineering emotional dependence — a rule that criminalizes the core retention mechanic of the entire companion-app business model. As the IAPP’s analysis notes, this sits inside a broader 2026 package in which Beijing is writing ethics review, agent management, and anthropomorphic-AI rules simultaneously, not as afterthoughts but as a coordinated governance stack.

Notice, too, what the obligations demand technically. Detecting “unhealthy dependence” in real time is not a checkbox; it is a longitudinal user-modeling problem — you need per-user baselines, session-frequency drift, sentiment trajectories, and an intervention policy that fires reliably without firing constantly. Distress detection that triggers on genuine self-harm signals but not on melodrama is a calibration problem that the best-resourced Western labs handle imperfectly with dedicated safety teams. China just made that capability a licensing precondition for an entire product category. The second-order effect cuts in an unexpected direction: the compliance burden functions as a moat for the giants, since only companies with frontier-scale ML teams can build mandated dependence detection — which may explain why ByteDance is rebuilding inside a compliant app while smaller persona platforms simply died in Shanghai’s June purge.

Enforcement did not wait for the effective date. Shanghai’s internet regulator removed more than 14,000 non-compliant AI agents in a single late-June sweep, citing impersonation, vulgar roleplay, and unauthorized data collection. App stores are now required to verify compliance status and delist violators. The platforms responded rationally: ByteDance is steering companion demand into Maoxiang, a separate, licensable app built for the new regime, while Qwen offered no migration path at all — a corporate shrug that says the feature was not worth the assessment paperwork.

The timing embarrasses the platforms’ own monetization plans, which is worth pausing on. ByteDance spent the spring preparing a paid subscription service for Doubao, the classic second act after a free-user land grab — and companion personas were the retention engine that made subscription math work. The law arrived at precisely the moment the engagement was about to become revenue. That sequencing is not accidental in the Chinese regulatory playbook; Beijing moved against gaming, tutoring, and fintech at analogous inflection points, when categories crossed from experiment to social infrastructure. The lesson platform strategists keep relearning: in China, the moment a product becomes too popular to ignore is the moment it becomes eligible for restructuring.

The regulatory arbitrage this creates is the sharpest number in the story. While China deletes the category, American consumers spent $162.8 million on romantic AI companion apps in the first half of 2026 alone, across 214 tracked apps — against roughly $120 million for all of 2025. That is a run-rate nearly tripling year over year, led by apps like Zeta at $33 million in six months. Beijing’s own justification documents cite the American wreckage as precedent: the Character.AI lawsuits, the FTC’s inquiry into companion chatbots, Italy’s action against Replika. China is regulating against the failure modes of a market it watched the United States build.

What the ban can’t delete

Every confident thesis about this law has a crack in it, and the cracks are worth mapping before anyone declares a governance winner.

Start with the definitional one. The Measures never specify a technical threshold for “sustained emotional interaction,” and that ambiguity is precisely why the platforms over-complied, killing entire feature surfaces rather than risk a wrong guess. Vague scope plus severe consequences produces preemptive amputation — efficient for regulators, brutal for product teams. But ambiguity cuts both ways. A rule that cannot define its object cannot consistently enforce against it, and China’s regulatory history is littered with categories — gaming curfews, tutoring bans, fan-culture crackdowns — where initial shock gave way to negotiated, licensed re-entry. Maoxiang is the tell: ByteDance is not abandoning synthetic intimacy, it is relocating it into a dedicated app built to pass inspection. If licensed companion apps clear security assessments this fall, July 15 will look less like prohibition and more like forced consolidation into surfaces the state can audit.

Second, demand does not obey administrative deadlines. The users who spent months or years shaping a persona do not stop wanting what they had because the feature flag flipped. The 2023 Replika episode, when a single moderation change left users in visible grief, previewed the psychology at population scale. Some of that demand will migrate to gray-market apps, VPN-accessed foreign services, and open-weight models run locally — ironically, the very Qwen and DeepSeek weights that China exports to the world can be self-hosted into exactly the companions the domestic apps can no longer offer. A law that pushes emotional AI from supervised platforms into unsupervisable basements may achieve the opposite of its child-protection aims.

Third, the strategic cost to China’s own AI ecosystem is real and unpriced. Companion interactions are not junk engagement; they are the densest source of long-horizon, multi-turn, preference-rich dialogue data that exists, the raw material for the memory and personalization features every lab is chasing. ByteDance just lost that flywheel domestically while American rivals — Meta with its AI personas, xAI with its companions, OpenAI openly weighing erotica for verified adults — keep harvesting. If personality and memory turn out to be the next capability moat, Beijing may have conceded a strategic dataset to protect a social norm.

There is also the motive question, which the child-safety framing conveniently obscures. The Measures’ stated aims include preventing the spread of “extremist ideologies” through anthropomorphic services — language that has little to do with screen-time hygiene and everything to do with the fact that a persuasive, personalized, always-available interlocutor is the most effective influence channel ever built. A companion that knows your loneliness is also a companion that can shape your politics, and Beijing understood that faster than Silicon Valley’s trust-and-safety teams did. Read this way, the law is less a consumer-protection statute than an information-control statute wearing one’s clothes — and that reading should temper any Western impulse to import it wholesale. The mechanisms that protect a teenager from engineered dependence are structurally identical to the mechanisms that ensure no synthetic voice competes with the state’s.

Finally, be skeptical of the “global template” framing, including mine. The American legal system cannot copy this law. California’s SB 243, the first state statute regulating companion chatbots, took effect January 1 and shows what the US version actually looks like: disclosure requirements, three-hour break reminders for minors, self-harm protocols, and a private right of action at $1,000 per violation — guardrails on speech-adjacent products, not bans of them. First Amendment doctrine, industry lobbying, and the federal preemption fights I covered when Illinois signed its frontier-model audit law all guarantee that American companion regulation stays incremental. The honest comparison is not “China leads, the West follows.” It is that two governance systems are now running controlled experiments on the same product category, and the results will take years, not quarters, to read.

Emotional AI is now a regulated substance — operate accordingly

The through-line of 2026 is that AI regulation stopped being hypothetical and started being architectural.

Two weeks ago, US export-control machinery pulled Anthropic’s Fable 5 offline for three days and taught every enterprise that model access is a policy variable. Today, China taught a complementary lesson one layer up the stack: entire product categories can be deleted by administrative rule, with ninety days’ notice and no compensation. The common thread is that regulatory risk has migrated from the compliance department into the product architecture itself. What a system is allowed to be — how warm, how persistent, how humanlike — is now a jurisdictional variable, and the systems that survive will be the ones engineered for that variability.

Watch four signals over the next two quarters. First, October 15: whether the data deletion actually executes, or quiet extensions appear — the difference between a state establishing precedent and a state managing backlash. Second, the security-assessment pipeline: if Maoxiang and licensed competitors clear provincial review by winter, China’s companion market re-forms under supervision, and the July shutdown becomes a licensing bottleneck rather than a prohibition. Third, the FTC’s companion-chatbot inquiry and the state-law wave behind SB 243 — New York, Utah, and a dozen statehouses are drafting, and Beijing just handed them a menu of mechanisms to cherry-pick. Fourth, the migration data: whether Chinese companion demand visibly resurfaces in VPN traffic, local-model downloads, and offshore app revenue.

The deeper shift is conceptual, and it will outlast the specific statutes. For three years the industry’s regulatory imagination was organized around model capability — compute thresholds, frontier audits, the Illinois-style safety regimes aimed at what models might do. China’s Measures regulate something different: what products make people feel. That is a far older and better-staffed regulatory tradition — the one that governs gambling, tobacco, and pharmaceuticals — and its arrival in AI means the relevant precedent is no longer software law but substance law. Substances get age gates, dosage limits, dependency warnings, and licensed dispensers. Every mechanism in the Chinese law maps onto that template, and the FPF’s survey of pending US chatbot legislation shows American drafters converging on the same vocabulary from the opposite political direction.

For operators — anyone building, buying, or investing in conversational AI — the checklist writes itself:

• Inventory your anthropomorphism. List every feature that simulates personality, sustains memory across sessions, or deepens emotional engagement. That inventory is now your regulatory exposure map, in every jurisdiction, not just China.
• Separate the functional layer from the emotional layer. China exempted workplace assistants and customer-service bots while banning companions. Architect your product so the two can be decoupled by configuration, not by rewrite, when a regulator demands it.
• Treat compliance thresholds as design inputs. China triggers review at 100,000 monthly actives; California triggers duties at first user contact. Know the tripwires in each market before growth crosses them, because retrofitting an assessment regime at Doubao scale is what a 90-day fire drill looks like.
• Make data export a day-one feature. Qwen users lost months of history with no migration path; Doubao users got screenshots. The first companion platform that treats user memory as portable property will convert regulatory trauma into a trust moat.
• Risk-weight companion revenue. The US market’s near-tripling to $162.8 million in half a year is genuine — and it is exactly the growth curve that summons regulators. Price Chinese-style category deletion as a tail risk in any model that depends on parasocial retention.
• Watch China as your leading indicator, not your template. The specific rules won’t port to Western law, but the category definitions — anthropomorphic service, emotional dependence, dependence detection — are already leaking into regulatory vocabulary worldwide. The concepts travel even when the statutes can’t.

The neon sign says love, twenty-four hours. As of this morning, in the world’s largest AI market, it comes with an off switch, an age gate, and a filing requirement. The companion-AI industry everywhere else should assume its own switch is being wired — and build products that can survive the flip. The labs that treat that assumption as an engineering requirement today will be the ones still selling intimacy, licensed and audited, when the inspectors arrive.

In other news

TSMC’s AI demand rewrites the record books — TSMC posted all-time-high Q2 revenue of $39.6 billion, up 36% year over year, with June alone contributing $13.8 billion — the largest month in the company’s history — and N3 capacity plus CoWoS advanced packaging sold out through year-end (Tech Times).

An OpenAI researcher chases a $2B drug-discovery debut — Miles Wang, an OpenAI researcher focused on AI for scientific discovery, is in talks to raise roughly $200 million at a $2 billion valuation for a new drug-discovery startup, with Lightspeed discussing leading the round and several OpenAI colleagues expected to follow him out (TechCrunch).

Microsoft fields 6,000 forward-deployed engineers — Microsoft launched Frontier Co., a $2.5 billion subsidiary embedding 6,000 engineers and sellers directly inside enterprise customers to fix the industry’s implementation gap, following similar deployment ventures from OpenAI, Anthropic, and AWS (CNBC).

Unitree clears the bar for China’s first humanoid IPO — Chinese robot maker Unitree Robotics won final regulatory approval for a $619 million IPO on Shanghai’s STAR Market, after reporting $250 million in 2025 revenue and shipping more than 5,500 humanoid robots last year (Reuters via Yahoo Finance).

Venture funding breaks $510 billion in six months — Global startup investment hit a record $510 billion in the first half of 2026, exceeding all of 2025’s total, with AI companies capturing more than 70% of the capital and OpenAI plus Anthropic alone accounting for $217 billion of it (Crunchbase News).