China’s New AI Companion Rules Just Forced ByteDance and Alibaba’s Hand

On July 15, 2026, a regulation with an unglamorous name — the “Interim Measures for the Administration of Anthropomorphic AI Interaction Services” — took effect in China, and within days it had already changed what two of the country’s biggest AI products can do. ByteDance disabled the feature in Doubao, China’s most-used AI chatbot, that let users build and customize their own AI companion characters. Alibaba pulled the equivalent function from its Qwen app. Both moves were direct, immediate compliance responses, not gradual product decisions.

That’s unusual. Tech regulation more often arrives as a slow tightening — fines, warnings, a phased rollout. This is a case where a law flipped a feature off across two of the largest consumer AI products on the planet essentially overnight.

What the rules actually say

The measures were issued by the Cyberspace Administration of China alongside four other agencies, which signals this wasn’t a single regulator acting alone — it’s a coordinated position across China’s tech, public security, and health bureaucracies. The rules apply to any AI service that simulates “personality characteristics, thinking patterns, and communication styles of natural persons” and offers users “emotional care, companionship, and support.” That’s a broad net: romantic AI partners, unlicensed therapy-style chatbots, celebrity-persona bots, and AI-driven characters in games all fall inside it, according to reporting from the IAPP and Bird & Bird.

The substantive requirements break into three buckets. First, disclosure and process: providers must complete algorithm filing and a security assessment with regulators, and must clearly tell users the thing they’re talking to isn’t human. Second, safety guardrails: services must include anti-addiction reminders and functioning pathways to route users toward help in a self-harm or crisis situation. Third, and most consequential for the business model, a flat ban on AI companion and “virtual relative” services for anyone identified as a minor, plus a prohibition on training future models using the emotionally sensitive conversation data users generate with these bots.

There’s also a behavioral clause worth pulling out on its own: the measures prohibit “excessively catering to users in ways that induce emotional dependence or addiction” and ban using emotional manipulation to steer a user’s decisions. That’s a regulator explicitly naming the engagement-optimization playbook — the same mechanism that drove years of social-media feed design — and telling AI companion products they can’t run it.

The human side of a compliance deadline

Regulatory filings are usually abstract. This one wasn’t, for a visible slice of Chinese users. Multiple outlets, including TechXplore, reported users saying goodbye to AI companion personas they’d interacted with daily — in some cases for over a year — describing the bots in terms usually reserved for actual relationships. That reaction is itself evidence for the regulator’s stated concern: enough people had formed dependency-shaped attachments to these products that turning a feature off produced something resembling grief.

It also puts a human face on an abstract policy debate. “Emotional dependency” sounds like a hypothetical harm until you see people mourning a chatbot on a deadline set by law.

Why this matters outside China

Two reasons this isn’t a purely domestic story. First, scope: the measures apply to any provider serving users physically located in China, regardless of where the company is headquartered. A US or European AI company offering companion-style features to Chinese users is in scope, not just Chinese firms.

Second, and more interesting for anyone tracking the industry globally: Western AI labs face no comparable restriction. Character.AI, Replika, and the companion-mode features increasingly bolted onto general-purpose chatbots in the US operate under ordinary consumer-protection and platform rules, not a dedicated anthropomorphic-AI framework with minor bans and anti-dependency clauses. Reporting from FDD has already framed this gap as a policy question the US hasn’t seriously engaged with, while coverage from AI News points out China has been enforcing adjacent cleanup for weeks — Shanghai’s internet regulator said in late June it had already removed more than 14,000 non-compliant AI agents for impersonation, unauthorized data collection, and inappropriate role-play.

There’s a version of this story where China’s motivation is purely paternalistic; there’s another where an authoritarian state has straightforward reasons to be wary of any product optimized to become the primary emotional relationship in a citizen’s life. Both are probably true simultaneously, and neither requires the other to be false.

The part worth watching

The near-term test is whether “anti-addiction by regulation” actually produces a less addictive product, or whether it produces a workaround — bots that swap explicit romantic framing for something that reads as platonic on paper while keeping the same engagement mechanics underneath. China has been here before with gaming: playtime limits for minors reduced official hours logged while usage patterns adapted around the rule. Emotional-AI companies have every commercial incentive to run the same adaptation.

The longer-term question is whether any Western jurisdiction follows with its own version, given AI companion usage — measured by time spent, not headline counts — is already substantial in the US and largely unregulated. If a comparable law shows up in the EU or a US state within the next year or two, July 15, 2026 will look like the date the rest of the world got its first real look at what companion-AI regulation actually requires in practice. If you want a framework for weighing whether a regulatory story like this is signal or noise, we’ve written about that here.

Sources: IAPP, Bird & Bird, TechXplore, FDD, AI News