China AI Companion Rules Take Effect: Doubao, Qwen Cut

China’s first national rules governing humanlike artificial intelligence took effect on Wednesday, July 15, 2026, and the immediate consequence was visible inside the country’s most popular chatbots: ByteDance’s Doubao and Alibaba’s Qwen disabled the personalized, persistent AI companion features that hundreds of millions of users had built into their daily routines. The shutdown marks the first time a major government has forced a wholesale redesign of consumer AI products around a mandatory safety regime, and it draws a sharp line between how China and the West intend to regulate the most emotionally engaging category of AI.

What the regulation requires

The measures at the center of the shutdown are the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, issued in April 2026 by the Cyberspace Administration of China (CAC) together with four partner agencies: 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. The framework is the first in China written specifically for AI services that simulate human personality, emotion, and sustained relationships, rather than the general-purpose generative AI rules that came before it.

According to the text, the measures cover systems that “simulate human personality traits, thinking patterns and communication styles to provide sustained emotional” interaction. The regulators frame the category as carrying distinct risks — the spread of extremist ideas, privacy leakage, harm to users’ physical and mental health, and, most prominently, dependence and addiction. In place of the open-ended, memory-rich companions that platforms had been racing to ship, the rules impose a set of hard requirements:

• Anti-addiction systems that monitor and limit prolonged or compulsive use.
• Mandatory usage notifications that periodically remind users they are talking to software, not a person.
• Instant-exit mechanisms that let a user break off an interaction immediately and cleanly.

The friction is architectural. Persistent AI companions are designed around long-lived memory and an unbroken sense of continuity — the very qualities that make them feel like a relationship. Rules that mandate interruption, disclosure, and hard usage limits are, as several platforms acknowledged in their own notices, incompatible with how those agents are built. Rather than retrofit the products, ByteDance and Alibaba chose to switch the features off.

How Doubao and Qwen responded

Both companies began winding down their custom-agent features ahead of the deadline, but they handled the underlying user data differently — a distinction that matters to the people who spent months configuring companions and accumulating conversation histories.

On ByteDance’s Doubao, users retain read-only access to their existing data until October 15, 2026. After that date, the data will fall under ByteDance’s standard privacy policy and will no longer be recoverable inside the app. The three-month window functions as a grace period for users to review or export what they can before the companion layer is fully retired.

Alibaba’s Qwen took a harder line. The company confirmed that agent configurations and conversation histories will be permanently deleted, with no announced migration path. For users who had built elaborate personas or leaned on a companion for daily support, that means the relationship simply ends — the stored context that made each agent distinct is being erased rather than archived.

The scale is significant. Doubao and Qwen sit among the most-used consumer AI apps in China, and the companion and custom-agent features being disabled reached an audience measured in the hundreds of millions. Neither company framed the change as a product decision; both pointed to the incoming regulatory deadline as the trigger.

Why China is regulating companions specifically

The measures reflect a policy judgment that emotionally persuasive AI is a category apart from productivity tools or search assistants. A general-purpose model that drafts an email is treated differently from one engineered to feel like a confidant, and Chinese regulators have concluded that the second kind warrants its own guardrails.

That concern is not unique to China. Around the world, the rapid growth of AI companion apps has drawn scrutiny over their effect on minors, on vulnerable users, and on anyone prone to substituting a chatbot for human connection. What distinguishes China’s approach is speed and force: rather than issue guidance or open a consultation, Beijing wrote binding rules with a fixed effective date and let the deadline do the enforcing. The result is a live demonstration of what happens when compliance is non-negotiable and the product cannot be made to comply — the product goes away.

For the platforms, the episode also underscores how tightly consumer AI in China is now coupled to the state’s regulatory calendar. The same AI agents that Western labs are shipping with ever-longer memory and autonomy are, in China, being pared back to fit a safety-first template. Persistent memory — the feature that lets an agent remember you across sessions and is central to how a large context window gets used in practice — is precisely the capability the rules target.

A widening regulatory divergence

The shutdown sharpens a contrast that has been building for more than a year. In the United States and Europe, the dominant regulatory questions have centered on frontier-model safety, transparency, and market power. The EU’s AI Act, whose general-purpose AI obligations moved into active enforcement this year, imposes documentation and systemic-risk duties on the largest model providers but does not force consumer companion features offline. Western labs continue to ship humanlike agents and voice companions with few product-level constraints.

China has now taken the opposite tack on this specific category, and it has done so unilaterally and on a hard schedule. That divergence has practical consequences for the global market. Chinese platforms must now build two mental models of their products — one that satisfies domestic anthropomorphic-interaction rules and one calibrated for markets without them — even as their models increasingly compete for enterprise adoption abroad. It also feeds directly into the broader international debate over how AI should be governed, a debate playing out in venues such as the UN’s global AI governance dialogue, where the question of whether emotionally persuasive systems need bespoke rules is now unavoidable.

What it means

China’s companion rules are the first real-world stress test of a regulatory idea that many governments have flirted with but none had enforced: that AI designed to form sustained emotional bonds is a distinct product class requiring distinct controls. The July 15 shutdowns show what enforcement looks like when the rules and the product architecture cannot be reconciled.

Who loses. The most immediate losers are users — hundreds of millions of them — who built companions on Doubao and Qwen and now face either a three-month read-only window or outright deletion. The platforms lose a high-engagement feature and the retention it drove, and they absorb the cost of re-architecting or abandoning products they had invested heavily in. ByteDance and Alibaba both chose to switch features off rather than attempt a compliant redesign, which signals how deep the incompatibility runs.

Who benefits. Regulators gain a demonstration that binding deadlines produce compliance, and safety advocates worldwide gain a precedent — evidence that a large market can require anti-addiction systems, usage disclosures, and instant-exit controls and see them implemented within months. Rival platforms that never built heavy companion features are comparatively insulated.

What to watch next. The first signal is whether Doubao, Qwen, and their competitors return with redesigned, compliant companions — agents rebuilt around session limits, periodic reminders, and constrained memory — or whether the category stays dormant in China. The second is diffusion: whether other jurisdictions weighing companion regulation, particularly around minors, treat China’s measures as a template. And the third is the product-design fallout globally — if persistent memory and unbroken continuity are the features regulators find most objectionable, labs everywhere may face pressure to add the same friction that made these Chinese products impossible to keep running. For now, the clearest takeaway is that in at least one major market, the most humanlike AI is no longer something a company can ship on its own terms.