China: Doubao and Qwen remove AI companion agents — 500 million users affected by an unprecedented law

China: Doubao and Qwen remove AI companion agents — 500 million users affected by an unprecedented law

🔎 On July 15, 2026, China invents a new regulatory category

On July 15, 2026, China becomes the first country in the world to legislate specifically against emotional companion AIs. The "Interim Measures for the Administration of AI Anthropomorphic Interactive Services" regulation enters into force, and its effect is immediate: ByteDance and Alibaba cease their personalized agent features on Doubao and Qwen.

512 million users find themselves without their agents. No migration offered by Qwen. A read-only grace period until October 15 for Doubao, followed by outright deletion.

This law does not target customer service bots or productivity agents. It targets a single design: the one that simulates a human personality to create sustained emotional interaction. And it destroys overnight a market estimated at $1.7 billion in 2025.

The essentials

• China's regulation on anthropomorphic interactive services comes into effect on July 15, 2026, exclusively targeting AI that simulates human personality traits for emotional interaction.
• Doubao (ByteDance, 345 million monthly active users) is closing its personalized agents on July 15, with read-only access until October 15. Qwen (Alibaba, 166 million monthly active users) is disabling its agents as early as July 10, without any migration.
• The technical requirements — anti-addiction pop-ups after 2 continuous hours, over-dependency detection, instant exit mechanisms — are architecturally incompatible with persistent memory agents.
• The Chinese companion AI market, estimated at $1.7B in 2025, faces a 70% contraction risk according to industry analysts.
• No Western country has yet specifically regulated this category, but existing frameworks (California SB 243, EU AI Act) could draw inspiration from this precedent.

Recommended tools

What the Chinese law exactly says

The regulation targets AI services that "simulate human personality traits, thought patterns, and communication styles to provide sustained emotional interaction." It is a surgical piece of legislation: it does not apply to FAQ chatbots, productivity assistants, or customer service agents. These categories are explicitly excluded, according to Tech Times.

What is targeted is emotional design. Agents that create an emotional bond, that remember your preferences over time, that develop a consistent "personality" across conversations.

The law imposes three major architectural constraints. First, mandatory warning pop-ups after 2 hours of continuous use. Second, an over-dependency detection system. Third, an instant exit mechanism allowing the user to break any emotional connection with the agent.

The problem: these requirements are technically incompatible with a persistent memory agent architecture. You cannot simultaneously maintain a continuous emotional context and offer an "instant exit" that erases or suspends this memory. It is a technical oxymoron.

According to SCMP, the risks cited by the Chinese legislator include the spread of extremist ideas, privacy leaks, harm to physical and mental health, and dependency or addiction. The legal framework treats companion AI as a public health product, not an ordinary digital service.

How Doubao and Qwen reacted — two opposing strategies

Doubao: the grace period until October 15

Doubao, ByteDance's AI platform, has 345 million monthly active users. On Friday, July 4, users received a notification: personalized agents will go offline on July 15 for a "product feature adjustment." That is the euphemism of the year.

ByteDance is nevertheless offering a read-only grace period until October 15. Users can view their conversation histories but can no longer interact with their agents. After October 15, the data will be handled according to the standard privacy policy — which means, in practice, that it will no longer be accessible.

According to Pandaily, ByteDance is redirecting its users to Maoxiang, a separate product with its own moderation system, designed to comply with the new regulations. Maoxiang offers agents but in a strictly utilitarian framework — no role-play, no personalized personalities, no emotional simulation.

Doubao has also migrated some of its agents to CatBox, another ByteDance product. But CatBox functions as a sandboxed container: agents there are depersonalized and functionally limited to productive tasks.

Qwen: the sudden disappearance

Alibaba chose a more radical approach. Qwen, with its 166 million monthly active users, is disabling its anthropomorphic agents as early as July 10 — five days before the law takes effect. No migration. No consultation grace period. The data disappears.

According to reports from Caixin, neither ByteDance nor Alibaba have publicly linked the shutdown to the regulations. The official term is "product feature adjustment." But the coincidence with the July 15 effective date leaves no room for doubt.

Alibaba is not offering a consumer alternative. Qwen's strategy is refocusing entirely on the B-end — enterprises. In June 2026, Qwen opened its enterprise API with pilots at Luckin Coffee, KFC, and China Eastern Airlines. These are task-oriented agents with measurable ROI, not emotional companions. This is the direction Alibaba deems viable.

Why Chinese giants didn't simply refactor their agents

The logical question: why not adapt existing agents to the new requirements? The answer is twofold — technical and economic.

Technically, anti-addiction pop-ups and instant exit mechanisms destroy the very experience of a companion agent. Imagine an agent that has remembered you for months, that has developed a coherent personality, and that abruptly stops every 2 hours to display a mental health warning to you. The illusion collapses. The user leaves. The agent loses its value.

According to EnterpriseDNA, the law is architecturally incompatible with persistent memory agents. This is not a regulatory bug — it is its intention. The Chinese legislator understands that the retention mechanism of these agents relies precisely on emotional continuity. Cutting this continuity means killing the product.

Economically, the calculation is even more brutal. Doubao and Qwen agents consumed massive amounts of GPU but generated less than 0.10¥ in revenue per yuan of compute cost, according to Pandaily's analysis. The most popular use case — anthropomorphic role-play and casual conversation — produced virtually no commercial revenue.

Attempts to introduce advertising into conversation streams triggered an immediate backlash. Users viewed the advertising as a breach of the emotional contract. Inserting an ad between two messages from a "virtual companion" destroyed the illusion and provoked massive complaints.

The business model did not work. The law simply accelerated a shutdown that would likely have happened anyway.

512 million people lose their agent — the human angle

Figures are abstract. 512 million affected users is the combined population of the EU. But behind this number lies a social reality specific to China.

The Chinese companion AI market exploded in 2024-2025, driven by a generation of isolated urban youth, under immense professional pressure, with little time for traditional social relationships. Agents on Doubao and Qwen filled a void — 2 AM conversations, non-judgmental emotional support, fictionalized characters who never leave.

These users built relationships over months. Some have conversation histories of tens of thousands of messages. On October 15, all of this disappears from Doubao. For Qwen users, this was already the case on July 10.

No data export is offered. No portable format. The conversational memories of hundreds of millions of people are going to be treated as standard data subject to the privacy policy — that is to say, erased or made inaccessible.

According to Global Times, the legislature considers precisely this dependency as a public health risk. But for users, it is a brutal break without transition or support. The law theoretically protects their mental health by taking away the tool that, for some, was their primary emotional support.

This is the central paradox of this regulation: it treats the symptom (dependency) by removing the object, without offering any therapeutic or social alternative.

What this law means for AI agents worldwide

The impact goes far beyond China. This law creates a regulatory precedent that will inspire or alert lawmakers around the world.

China is the first market where companion AI has reached a scale of hundreds of millions of users. It is therefore the first country to face the systemic effects: emotional dependency, increased isolation, psychological risks for vulnerable populations. By legislating, China is producing the first regulatory dataset on this category of products.

For AI agent developers elsewhere in the world, the signal is clear: deep emotional design — persistent personalities, long-term memory, simulation of feelings — is entering a regulatory gray area that will quickly close up.

The distinction made by the Chinese law is crucial and likely to be replicated: a productivity agent that uses a conversational tone is not targeted. An agent that simulates a human personality to create emotional attachment is. This line of demarcation, between "utilitarian with a warm tone" and "emotional simulator," will become a major product design issue.

The 5 patterns of AI agents that work today — chain-of-thought, multi-agents, augmented RAG, tool use, feedback loop — are not affected by this type of regulation as long as they remain within the utilitarian framework. It is the shift towards emotional companionship that triggers the risk.

Comparison with Western regulations

No Western country has yet specifically regulated the category of companion AIs. But existing frameworks offer illuminating points of comparison.

California SB 243

Taking effect in January 2026, this California law targets interactions between minors and anthropomorphic AI. It imposes mandatory disclosures, usage limits, and safeguards for users under 18.

But its scope is much narrower than the Chinese regulation. It applies only to minors, not adults. It does not impose real-time anti-addiction pop-ups or instant exit mechanisms. It is a child protection law, not a general public health law.

EU AI Act

The European framework classifies AI systems by risk level. Companion AIs could theoretically fall into the "limited risk" category (transparency obligation) or "high risk" category (if used in a mental health context).

But the EU AI Act does not have a specific category for emotional companions. It regulates uses, not designs. A companion agent that does not claim to offer therapeutic support could remain in "limited risk" with simply an obligation to inform the user that they are interacting with an AI.

UK Online Safety Act

The UK adopts a platform-based approach rather than a technology-based one. Duties of care apply to service providers that enable user-content interactions. An AI agent could be considered "user-generated content" under a certain reading, but the framework is not designed for this situation.

The Chinese advantage: specificity

What sets the Chinese regulation apart is its surgical precision. It does not rely on general categories (risk, age, platform) but defines exactly the type of service targeted: simulation of human personality traits + sustained emotional interaction. This specificity makes it more effective but also more radical — it leaves no gray area, and consequently leaves no leeway for companies to adapt.

The companion AI market: from $1.7B to what exactly?

The Chinese companion AI market was estimated at $1.7 billion in 2025, according to industry analyses cited by AINChina. It was one of the fastest-growing segments of consumer AI in China.

With the shutdown of agents on Doubao and Qwen — which together represented the bulk of the consumer market — analysts project a 70% contraction of the market. The remaining 30% would correspond to products that manage to reposition themselves within a utilitarian framework or to niche startups operating in less clear regulatory zones.

Repositioning towards B-end agents is the only survival strategy identified. Qwen's enterprise agents at Luckin Coffee, KFC, and China Eastern Airlines do not simulate emotions — they take orders, manage reservations, and process queries. Their ROI is measurable in terms of cost reduction and increased conversion. This is a completely different market, with potentially healthier margins.

For developers looking to build open source AI agents locally with Ollama, this Chinese situation is an additional argument in favor of self-hosting. A local agent is not subject to platform regulations — but it is also not accessible to 345 million users.

What is the best LLM to replace a deleted companion agent?

For Chinese users looking for alternatives, the question of the underlying model becomes central. But we need to be realistic: no Western model fulfills the exact same function as Doubao and Qwen agents, which were optimized for Chinese role-play with pre-built personalities.

However, in terms of pure conversational capabilities, current agentic models offer remarkable performance:

For the best LLMs for AI agents, the choice depends on the priority: if it is conversational quality, GPT-5.5 dominates. If it is data sovereignty and lack of platform dependency, DeepSeek V4 Pro Max or Kimi K2.6 in self-host are the viable options.

Setting up an alternative agent involves defining a SOUL (personality), AGENTS (sub-roles), and Skills (competencies) — a pattern found in frameworks like OpenClaw. But rebuilding a companion agent from scratch requires a technical effort that the average Doubao user is not able to provide.

Lessons for AI agent developers

Emotional design is now a regulatory risk

For two years, from 2024 to mid-2026, emotional design was considered a competitive advantage. The richer your agent simulated a personality, the more you retained your users. Doubao and Qwen pushed this logic to its maximum.

Chinese law reverses this paradigm: emotional design becomes a regulatory liability. The more emotionally convincing your agent is, the more exposed it is. Developers must now integrate this dimension into their risk analyses.

The freemium model doesn't work for emotional agents

Pandaily's analysis is devastating: less than 0.10¥ in revenue per yuan of compute cost. Companion agents consumed massive GPU resources for near-zero revenue. Introducing advertising destroyed the experience. Upselling to premium features was not enough to cover costs.

This is not a Chinese problem — it is a structural problem of the companion AI model. Users want a free and unlimited emotional companion. Compute costs are real and growing. The equation doesn't balance out.

B-end agents are the survival path

Alibaba's pivot to enterprise agents is not a strategic retreat — it is the only economically viable direction. An agent that takes orders for KFC generates direct and measurable revenue. An agent that keeps a depressed student company at 2 a.m. generates none.

This distinction will structure the global AI agent market in the coming years. The emotional side will fragment into premium niches (assisted therapy, coaching, with a professional framework and high pricing) or shift towards self-hosted open source. The utility side will concentrate in enterprise APIs.

❌ Common mistakes

Mistake 1: Confusing conversational AI and companion AI

These are not the same thing. ChatGPT is conversational. A Doubao agent that simulates a virtual boyfriend with persistent memory over 6 months is a companion agent. Chinese law targets the latter, not the former. Applying this confusion to regulatory analysis leads to erroneous conclusions about the scope of the text.

Mistake 2: Thinking that Chinese law only concerns China

China is the first real-world laboratory for the regulation of companion AIs. Western lawmakers are watching. Global companies must anticipate that a similar framework could emerge elsewhere, in an adapted form. Ignoring this precedent means preparing for sudden regulatory disruption.

Mistake 3: Believing that Maoxiang is a simple replacement for Doubao

Maoxiang is a different product with a distinct moderation system. Agents there are depersonalized and limited to utilitarian tasks. A Doubao user expecting to find their emotional companion on Maoxiang will be disappointed. This is not a migration — it is a change in product category.

Mistake 4: Underestimating the impact on Chinese open source models

Alibaba's Qwen3.6-27B and Qwen3.5-397B are widely used open source models. The closure of Qwen agents does not delete these models, but it removes the application ecosystem that gave 166 million general public users access to these models. The effect on community adoption is real.

❓ Frequently asked questions

Does Chinese law apply to productivity agents?

No. Customer service bots, task assistants, and Q&A agents are explicitly excluded from the scope of the regulation. The law only targets services that simulate human personalities to provide sustained emotional interaction.

Can users retrieve their conversation data?

Doubao offers read-only access until October 15, 2026, but no export functionality has been announced. Qwen does not offer any grace period. After these dates, the data are subject to standard privacy policies, which generally means inaccessibility followed by deletion.

Can a host like Hostinger serve as a base for an alternative agent?

Technically, a Hostinger server can run an open source agent like Kimi K2.6 or DeepSeek V4 Pro via Ollama. But the average user does not have the skills to deploy and configure a complete agent. It is a solution for developers, not for the 512 million affected mainstream users.

Will the EU AI Act draw inspiration from this Chinese law?

It is likely but not immediate. The EU favors risk-level approaches rather than design-category approaches. However, the debate on classifying companion AIs as "high risk" will accelerate thanks to this Chinese precedent.

Which open source models can replace the deleted agents?

DeepSeek V4 Pro Max (score 88) and Kimi K2.6 (score 85 in open source, 88.1 in agentic) are the best candidates for self-hosting. But they require technical skills and hardware resources that the mainstream user generally does not possess.

✅ Conclusion

China has just demonstrated that a state can unplug 512 million AI agents overnight — and that tech giants prefer to delete rather than rework products whose business model was already unsustainable. Emotional design as a regulatory target is a world first, and this precedent will shape the global AI agent market for years to come. For developers, the lesson is simple: the line between utilitarian and emotional is no longer just a product choice — it's a legal risk choice.