China’s AI Companion Law went into effect on July 15, 2026, prompting the immediate shutdown of personalized AI features on ByteDance’s Doubao and Alibaba’s Qwen, two of the country’s most widely used chat applications.
What the new rules require
The regulation, formally titled the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, was issued on April 10, 2026 by the Cyberspace Administration of China and four partner agencies. After a three‑month grace period, it became legally binding on the date of the shutdown.
Unlike broader AI rules, these rules draw a clear line between services that aid work and those that act as companions. Customer‑service bots, knowledge‑base tools, workplace assistants and educational apps remain permitted as long as they avoid sustained emotional engagement. The law targets services that simulate “personality traits, thinking patterns, and communication styles of natural persons” to provide continuous emotional interaction.
Key obligations include anti‑addiction systems, mandatory notices after two consecutive hours of use, instant‑exit options, and real‑time detection of unhealthy emotional dependence. For users under 14, parental consent is required, and virtual intimate relationships are banned for anyone under 18. Services that launch new anthropomorphic features or reach one million registered users—or 100 000 monthly active users—must undergo a tiered security assessment and file their algorithms with regulators.
Seven categories of content are prohibited, covering material that encourages self‑harm, induces emotional dependence, or manipulates users into unreasonable decisions.
Related: New Drug Shows Promise Against Alzheimers
Why Doubao and Qwen went dark
Both platforms relied on persistent‑memory agents that keep conversation history and maintain a stable persona across sessions. That architecture makes it difficult to insert the interruptions the law mandates without fundamentally altering the product.
Industry observers note that retrofitting anti‑addiction layers onto such systems would require rebuilding the core design, a cost many companies deemed prohibitive. The two companies therefore chose to cease the companion features altogether rather than attempt a partial compliance fix.
ByteDance reported 345 million monthly active users on Doubao. Those users can view their agent configurations and chat logs in read‑only mode until October 15, 2026, after which the data will be handled under the company’s standard privacy policy and will no longer be recoverable inside the app. Alibaba gave Qwen users no comparable grace period; the company confirmed that agent data is already being permanently erased with no migration path announced.
Criticism surfaced on Chinese social media, where users lamented the loss of months or years of personal conversation history. One post described the agents as “long‑standing emotional support,” highlighting the human impact of the abrupt change.
For many users, the shutdown means losing a digital confidant that helped them study, role‑play, or simply pass the time. The abrupt removal of those tools may push some to seek alternative outlets, potentially increasing the risk of unsupervised interaction with less regulated platforms.
Related: Taobao launches regional bidding for Billion Dollar Subsidy
China’s move reflects a broader engineering dilemma.
Persistent‑memory companion AI, whether in China or the West, depends on three core features: cross‑session memory, stable persona, and emotionally calibrated interaction. Anti‑addiction rules demand breaking that continuity, inserting alerts, and resetting engagement after extended use. The result is a technical incompatibility that no major provider has yet solved.
Western platforms such as Character.AI and Replika face the same design challenge, though they are not currently subject to China’s rulebook. In the United States, California’s SB 243, effective January 1, 2026, requires disclosures that bots are not human and adds safeguards for minors, but it stops short of imposing structural changes. The EU’s AI Act also adopts a risk‑based approach without mandating the deep architectural revisions seen in China.
Regulators worldwide are watching the Chinese example.
The Chinese rules are the first national framework explicitly targeting emotionally interactive AI, a category not yet singled out in other jurisdictions. While the stated goal is user protection from psychological manipulation and addiction, the policy also signals a shift toward more granular tech governance.
Related: AMD Restores Ryzen Memory Encryption in July
Companies aiming to operate in China must now consider the compliance cost of any anthropomorphic feature.
App stores are required to verify compliance status, and non‑compliant products can be removed. The requirement to file algorithms and undergo security assessments adds another layer of oversight that may influence product roadmaps.
The upcoming deadline for Doubao data provides an early test of user response. If many users export their histories and migrate to ByteDance’s new companion offering in the Maoxiang app, the company may retain a portion of its audience while meeting regulatory expectations. Alibaba’s lack of a migration path could draw further scrutiny if discontent among Qwen users grows.
In practice, the shutdown shows how regulatory design can force a fundamental rethink of product architecture. Developers will need to decide whether to rebuild companion AI from the ground up with built‑in compliance mechanisms or abandon the market segment altogether.
For now, the immediate effect is clear: millions of Chinese users have lost access to the AI companions they built, and the industry faces a new benchmark for how emotional AI must be engineered to survive regulatory scrutiny.