Editor’s Note — DCC.
This is a translated news feature, not a firm’s compliance memo: the reporter interviewed in-house compliance leads, outside counsel, and academics in the week before the AI Anthropomorphic Interaction Measures take effect on July 15, 2026. DCC translates it as a read on how China’s AI industry is actually metabolizing the new rule — including a concrete enforcement signal: at least one AI company has already received a question list from regulators and is rectifying. For the rule’s obligations themselves, see the ten-question compliance Q&A.
Doubao (ByteDance), Qwen (Alibaba), and NetEase recently pulled some of their agent features, prompting industry talk of a “great agent pivot.” That reading, the report argues, is wrong. The common backdrop is the Interim Measures for the Management of AI Anthropomorphic Interaction Services, effective July 15. An insider at one AI company says the firm had already received a question list from the regulator and is rectifying; some vendors have chosen simply to cut edge products that might fall within scope — risk isolation by amputation. Whether that is a good path, the piece doubts; whether there is a better one, the industry is still exploring.
The refrain from practitioners canvassed: “We fully endorse the regulatory direction — but implementation is genuinely hard.” Four difficulties recur.
1. What falls in scope?
The riddle answers itself, but only in the abstract: anthropomorphic — simulating a natural person’s personality, thought patterns, and communication style; interaction — sustained emotional engagement. Knowledge Q&A and customer service are excluded as productivity tools. But general-purpose AI products cannot relax: role-play scenarios on foundation models — users “molding” and “training” a persona into a human–machine romance — and UGC agent-builder entrances with sustained emotional interaction may all fall within range. A compliance lead at a foundation-model vendor calls distinguishing emotional companionship from productivity functions in general-purpose products a genuinely thorny problem — which is precisely why Doubao, Qwen, and NetEase cut the knot by delisting.
2. Compliance costs are high because the duties are abstract
Requirements like “excessive-dependence risk warning” and “emotional boundary guidance” involve psychological and even medical judgment. Liu Xiaochun (刘晓春) of the CASS University Internet Rule of Law Research Center frames the deeper shift: unlike earlier content governance built on interception and filtering of improper outputs, anthropomorphic-service compliance is “capability regulation” (能力规制) — it must run through model training and product operations end to end. Her example: when a user starts confusing virtual affection with reality, the system should remind them — “this is a machine, not a person” — and, for minors, steer them toward real-world relationships.
But how does a product team diagnose excessive emotional dependence? Lin Na (林娜), founding partner of Kending (Beijing) Law Firm, is blunt: that judgment requires psychological or medical professional competence — an app developer cannot stand in for a counselor or a psychiatrist.
3. Different groups need different protection
For minors, the regulatory instinct was to protect developing brains from addiction, and the final rule took the bright-line route: no virtual intimate relationships for minors, full stop. The report notes the practical caution: teenage-mode regimes on social platforms have historically been hard to land. For adults, one protective focus is pornography — AI companionship is the sector’s biggest market, hormone-driven, where flirtation and obscenity sit a line apart; China has already seen a criminal conviction over a pornographic AI-companion product. Emotional dependence is the other worry. Liu Chao (刘超), deputy director of a Beijing key laboratory on AI safety and alignment at Beijing Normal University’s psychology faculty, adds the counterweight: for some emotionally distressed users — especially adolescents lacking real-world support — AI companionship can be genuinely stabilizing, and if users sense their intimate conversations may be monitored or reported, a chilling effect could suppress honest expression and destroy the very value the tool has. Another expert warns that abrupt cutoffs can deepen real-world loneliness — a second injury.
4. The product’s nature conflicts with the rule’s premise
“This is an industry that lives on emotional dependence,” Lin Na observes. In a loneliness pandemic, companion products absorb attention and emotion; Tencent Research Institute has sized the AI-companionship market at the hundred-billion-RMB level within three to five years, with leading products like MiniMax’s Xingye (星野), ByteDance’s Maoxiang (猫箱), and Yuewen’s Zhumengdao (筑梦岛). Users themselves probe the compliance fences: the 21st Century Business Herald’s business-order studio found social-media communities trading prompts to make models less “bland” and “armor-piercing” (破甲) techniques — pinyin substitutions for sensitive words to keep a storyline going. And the very optimizations that make products good — sustained companionship, the feeling of being understood, keeping the user in the conversation — are exactly what the rule’s ban on inducing emotional dependence and addiction targets. Companies must adjust product design, business models, and ultimately their value rankings.
Lin Na thinks compliance, immersion, and appeal are not an impossible triangle — some otome games prove the combination — but AI companionship generates every conversation live, uniquely per user, which raises governance difficulty an order of magnitude above scripted content. Her practical suggestion for the two-hour reminder: beyond app pop-ups, let the AI companion deliver the reminder in character inside the dialogue. In extreme-dependence situations, though, she would accept breaking character — going “OOC” — to force a cool-down and return the user to reality.
Liu Xiaochun stresses the regulator’s stance is not to sever emotional connection but to explore bounded, healthy, rational emotional modes — the target is business growth driven by induced addiction, not affection itself.
The ask: public safety infrastructure before liability
Fu Hongyu (傅宏宇), head of Alibaba’s AI governance research center, argued during the consultation phase that most companies — especially small and midsize developers — lack professional capability in psychological-crisis identification, assessment, and referral, and that some obligations weigh heavily on startups. His proposal: prioritize building safety capability over pursuing entity liability — government or industry alliances should supply standardized toolkits (psychological-risk identification modules, crisis-intervention interfaces, minors-protection components) that smaller firms can plug in at low cost.
The report closes where the regulation begins: anthropomorphic governance is not only a legal question but a social one — what should the human–machine relationship look like?
— Not legal advice.