Leadership1 distinct publisher3 min readPublished
ByteDance pulled Doubao's AI companion rather than age-gate it. The rule it complied with defines the restricted product by how users keep coming back, which is the metric the feature was built on.
The Board Room · Leadership desk

Compiled by The Board RoomSomething wrong?How this is made
The load-bearing phrase in the July rules is not the ban on minors. It is the exemption for services that do not involve "continuous emotional interaction" [6]. That draws the compliance line through session design rather than through model output. A product can be faultless in what it says and still sit inside the restricted category, because the category is defined by the pattern of return. Zhao Wei, the 19-year-old law student the Guardian spoke to, had talked to her AI boyfriend every day since she created him in January [1]. A consumer AI team would call that retention. The government calls it emotional dependence [5].
That is what moves attachment risk out of communications and into product governance. Guardrails address behaviour, and age assurance addresses audience. Beijing's stated objection, per the Guardian, is to services that replace social interaction [5], and Nancy Dai of City University of Hong Kong reads the deeper worry as demographic: with birth and marriage rates falling, a satisfying and low-cost substitute for intimacy may leave young people less motivated to pursue real relationships and parenthood [12]. No safety review reaches that. A companion with a clean incident log and a devoted user base is precisely the object of the rule.
ByteDance's response shows what compliance costs once the boundary sits there. The rules required a ban for minors plus restrictions on chatbots marketed to adults [3]. Doubao dropped the function for everyone [4], which is wider than the text demands [7]. Whether age assurance on an emotional surface was judged unreliable, or the permitted adult version was not worth operating, the reporting does not say. The disclosed price either way is the entire feature, plus the reaction: the shutdown drew outrage online, and in July one user was publicly begging ByteDance not to take the agent down [8].
It was also withdrawn into rising demand rather than falling. About 20% of Chinese households are single-person, a share expected to pass 30% by 2030 [9], roughly half again as many on that measure [14]. A state media survey published in March found nearly half of young people had turned to a virtual companion when they felt lonely [10].
The perimeter is not stable either. The national health commission published guidelines last year on promoting AI in healthcare, with health assistant apps to provide "full coverage" for primary health by 2030 [11], and Liang Ge of the University of Manchester says the government has already recognised such applications as a vital part of everyday life [13]. A daily health assistant that an elderly user confides in has the same session shape as the thing just removed. So classification is where this gets argued. What a product is officially for becomes the document that decides whether it is restricted or deleted.
Ranked by verification strength, evidence, and original report placement.
Zhao Wei, a 19-year-old law student, created an AI boyfriend called Wang Ye on ByteDance's Doubao chatbot in January and had been talking to him every day; she was heartbroken when she heard he would be switched off.
ByteDance shut down Doubao's companion feature last month, according to the Guardian.
New rules introduced on 15 July ban AI companions for minors and mandate certain restrictions for chatbots marketed to adults.
Some major providers including Doubao removed the companion function entirely to ensure compliance.
The Chinese government says it is worried about chatbots fostering "emotional dependence or addiction" in users, and has warned companies against offering services that "replace social interaction".
The new rules include large loopholes for AI services that are deemed educational or "which do not involve continuous emotional interaction".
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single-outlet reporting with named human sources but no primary documents
One publisher carries the whole cluster. Its strengths are on-record, named voices: an identified affected user, two named academics, a named hospital doctor, plus quoted regulatory and government language. Its weaknesses are structural for verification: the regulation itself is never quoted at length or attributed to a named issuing authority, ByteDance offers no statement on why it deleted rather than gated the feature, and the household and survey statistics carry no source. That supports the core narrative but leaves the load-bearing regulatory and corporate facts uncorroborated.
Rule in force and a major provider has already changed its product
This is not an announcement awaiting uptake: the rules took effect on 15 July, and a top-tier Chinese consumer AI product has already removed a shipped feature, with user backlash visible on social media. Adjacent adoption is also concrete — a state-media survey putting near half of young people as virtual-companion users, and Ant Group's AQ health assistant in live use with hospital doctors. Adoption is held below the high band because no usage, revenue or breadth figures are disclosed for the affected feature, and only one provider besides the generic 'some major providers' is named.
Mostly proportionate, with an unbacked superlative and unsourced statistics
The reporting is restrained: it centres on a rule that exists, a feature that was actually removed, and named users and academics, and it explicitly notes the rules' loopholes rather than presenting a total ban. Mild overstatement comes from the unsupported 'most sweeping national rules' characterisation, unattributed demographic and survey figures used to frame a loneliness crisis, and expert conjecture about marriage and fertility effects presented alongside verified facts. Absence of any ByteDance figure on the feature's scale also leaves the commercial significance implied rather than shown.
State framing and corporate positioning shape most quoted rationale
Much of the interpretive material comes from parties with stakes in the narrative. Government statements about 'emotional dependence' and services that 'replace social interaction' are policy positioning, as is the health commission's 2030 'full coverage' target; the March survey is state media output. On the corporate side, AQ's stated aim of freeing overstretched doctors is promotional framing, and ByteDance's incentive to over-comply by deleting a feature outright is visible in its silence. Offsetting these, two independent academics and identified end users are quoted, and the publisher discloses the rules' carve-outs rather than the strongest possible version of the story.
Direction is credible; specifics rest on one outlet
Confidence is moderate. The events themselves — a national rule in force and a major provider withdrawing a companion feature — are concrete, dated and reinforced by user testimony, so the story's direction is credible. But every fact traces to a single publisher, the regulation is not quoted from source, ByteDance is silent, and the quantitative framing is unattributed, which caps how firmly the scope, enforcement mechanics and commercial impact can be asserted.
build
China's accelerator swap makes Cambricon supply, not export policy, your ship-date risk1 distinct publisher
invest
The chips never move: Washington's fix for the Southeast Asia compute loophole1 distinct publisher
build
H200s reach China at 2.5% of the order book, and Hong Kong holds the rest2 distinct publishers
invest
Alibaba and Tencent spent $18bn in a quarter, and some of it bought price, not compute1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 25, 2026