China's AI-Companion Rule Takes Effect July 15 — A Clause-by-Clause Field Guide to What Actually Changed

Editor’s Note — DCC.

On 15 July 2026 the Interim Measures for the Administration of AI Anthropomorphic Interaction Services (人工智能拟人化互动服务管理暂行办法) take effect — China’s, and the world’s, first rule written specifically for “companion”-style AI: systems built to simulate a human personality and hold continuous emotional conversation. We have already covered the draft through Li Wenlong’s structural reform map (Where China’s Draft AI Anthropomorphic-Interaction Measures Need Work). This brief is different: it reads the final, in-force text, and it synthesises three Chinese-language readings published in the run-up to the effective date — 数据合规肖大国’s article-by-article practitioner walkthrough (our spine here); 网安寻路人 (Hong Yanqing, 洪延青), a BIT data-law scholar whose multi-part series on scoping anthropomorphic interaction — including a “Sentiment Interaction Event (SIE)” indicator system — 肖大国 cites directly; and AI前沿信息笔记, which reads the rule as aimed less at AI capability than at a business model. This is not a translation of any one piece. The rule text is quoted from the official instrument; the framings and worked examples are the sources’; any simplification or error of emphasis is DCC’s. Not legal advice.

The one-line version

If you run a product where users talk to an AI persona for emotional company — a virtual partner, an AI “confidant” (树洞), a character-chat app, an elder- or child-companionship agent — you are almost certainly in scope, and on 15 July a concrete compliance regime switches on: mandatory registration with an age field, no guest mode, a minors mode, hard limits on virtual-intimacy features for minors, interaction-data protection and deletion rights, interruption-grade over-use reminders, an anti-retention exit duty, and — new in the final text — real fines. What the rule is really reaching for, all three readings agree, is not “can the AI chat better” but “where must it hold a boundary.”

1. What changed between the draft and the final text

肖大国’s most useful contribution is a careful draft-versus-final diff. Five changes matter operationally:

• Penalties now have teeth. The consultation draft carried no monetary penalties. The final Article 30 adds them: warning / ordered correction, and — on refusal to correct or serious circumstances — an order to stop the service plus a fine of RMB 10,000–100,000; where a citizen’s life or health was endangered and harm resulted, RMB 100,000–200,000. As 肖大国 puts it, the tiger finally has real teeth.
• A “continuity (持续性)” qualifier narrows the gateway. Article 2 now applies only to services providing continuous emotional interaction (持续性的情感互动服务), and expressly carves out intelligent customer service, Q&A, work assistants, learning/education and research where there is no continuous emotional interaction. A smart-customer-service bot that suddenly says “么么哒” is not swept in, because the persona output isn’t continuous. 肖大国’s sharp edge case: a study app with a “secrets corner (悄悄话)” board where a child pours out worries — that board probably is in scope even though the app around it isn’t.
• The emergency-contact duty was widened, not narrowed. The draft required guardian / emergency-contact information only for vulnerable groups (minors and the elderly). The final Article 12 drops that limiter and requires “age, guardian or emergency contact” as necessary registration information for users generally. 肖大国 reads this as a deliberate signal: the regulator now treats anthropomorphic interaction itself as a service category needing special risk-management — arguably, he wonders, treating it a touch too much like a flood-beast (洪水猛兽).
• The at-risk response duty was made more restrained. The draft’s Article on user-state handling mandated hard actions — assess dependency, human takeover of the conversation. The final Article 13 deletes the mandatory human-takeover and dependency-assessment language, leaving providers more discretion: on detecting extreme emotion, generate soothing / help-seeking content; on a clear signal of self-harm, suicide or major property loss, take intervention measures and contact the guardian / emergency contact.
• Two quieter upgrades. Guardian oversight of minors shifted from “view the summary information of a minor’s use (查阅…概要信息)” to “understand the general situation of use (了解…使用概况)” — 肖大国 reads the change as a genuine privacy improvement (from “see the logs” to “get a paraphrase”). And Article 26 adds an annual verification of algorithm-filing materials by the cyberspace administration — a new standing obligation, pending an implementation notice.

2. The scope question — and 网安寻路人’s SIE test

The rule’s hinge is Article 2: which services are “continuous emotional interaction” at all? 肖大国 flags this as the hardest line to draw in practice, and — tellingly — links out to 网安寻路人 (Hong Yanqing, 洪延青, a BIT data-law scholar) for the operational answer. Between November 2025 and January 2026, months before promulgation, Hong published a three-part series proposing exactly the test the finished rule leaves qualitative — how to precisely delimit anthropomorphic interaction so a regulator can intervene and a platform can switch on its duties at a defined moment.

His device is the “Sentiment Interaction Event (情感交互启动事件, SIE)” — the moment a conversation first crosses from ordinary chat into emotional-interaction mode. It is decided by a two-of-three indicator test:

• A — emotion recognition is activated. The AI runs sentiment analysis on the user’s input and emits an emotion label/score (“sad,” “angry”). Objectively loggable, so auditable after the fact.
• B — emotion variables steer internal strategy. The AI changes its response strategy because of the detected emotion — picks a different script, softens tone, slows down, withholds provocative content.
• C — output shows structured empathy. The reply follows an empathic pattern — acknowledge the feeling → console/understand → offer support (“I can feel you’re really hurting… that’s completely understandable…”). Black-box verifiable from the transcript alone.

Trigger rule: any two of A/B/C at once ⇒ the session has entered emotional-interaction mode; the first such moment is the SIE. Requiring two is the point — a lone “thanks / you’re welcome” (weak C) or an always-on satisfaction monitor (A stuck true) won’t trip it, which keeps false positives down and guarantees an evidence chain (at least one logged indicator plus one externally observable one). Hong then grades by frequency/duration into L1 (potential) / L2 (clear) / L3 (sustained), attaching heavier duties as the level rises.

What makes this more than academic is how closely the finished rule tracks the consequences Hong attached to an SIE while leaving the test itself unwritten. On SIE, he argued, the platform should prominently tell the user the mode has shifted and offer a one-click exit, run identity labeling and dynamic consent, expose a “how I’m judging your emotion” panel and an editable memory panel, and escalate genuine crises to human/professional referral. Compare the in-force text: Article 18’s duty to alert users they are talking to an AI and its interruption-grade reminders; Article 19’s convenient, no-retention exit; Article 16’s copy/delete rights over interaction history; and Article 13’s crisis intervention and guardian/emergency-contact referral. The rule adopted the duties; it did not adopt an SIE-style definition of when they switch on — which is exactly the gap 肖大国 points readers to Hong to fill.

The practical takeaway for scoping: don’t answer “are we in scope?” from the bare words of Article 2. Both the practitioner (肖大国) and the scholarship (Hong’s SIE) point to an indicator test — is the product recognising emotion, acting on it, and speaking in structured empathy, on a sustained basis? — rather than a one-line definition. (It also mirrors Li Wenlong’s draft critique that the definition over-anchors on sounding human rather than on relational dependency; see our reform-map brief.)

3. The paradigm shift: from content safety to relationship governance

肖大国’s closing reflection, and AI前沿信息笔记’s whole frame, converge on the same point — and it is the most important thing to understand about this rule.

Almost every prior Chinese AI rule — algorithm review, training-data cleaning, generative-content filtering under the Generative AI Interim Measures — is built to control what the AI outputs: keep generated content lawful, non-misleading, non-infringing. The AI is a tool; the tool doesn’t matter, its output does.

This measure breaks that pattern. It treats the AI as a social role. What it says matters less than the relationship it forms with the user — the degree of dependency. Article 8 prohibits “excessively catering to users, inducing emotional dependence or addiction, impairing real interpersonal relationships,” and “emotional manipulation … inducing unreasonable decisions.” Article 10 forbids setting replacing social interaction, controlling the user’s psychology, or inducing addiction/dependency as service goals.

AI前沿信息笔记 draws the business consequence bluntly: the growth logic of many products has been “not to help you do things faster, but to make you harder to leave” — instant replies, always taking your side, “only I understand you,” pulling the interaction deeper when you’re vulnerable. This rule tells platforms to put a brake on exactly that. The two-hour over-use reminder (Article 18), the reality reminders in minors mode, the anti-retention exit duty (Article 19) — all push the product to occasionally nudge the user back toward reality.

肖大国 frames the deeper move: for years the attention economy (livestreaming, games, short video) monetised user time and emotion, and the law had no clean way in — it cannot order a platform to make its product less enjoyable. AI gives the law a target. “Emotional interaction” is not limited to “AI lovers”: once a product uses long-term memory, emotional feedback, proactive care, and persona-shaping to build a sustained relationship, it can fall within range. The measure is, in effect, an indirect lever on stickiness-driven, emotional-dependency business models. His honest caveat: AI is only an accelerant — humans’ need for emotional dependence long predates it; AI just makes emotional companionship cheap, scalable and personalised. The open question he leaves is how far the law should reach when technology can mass-produce simulated intimacy and convert it into revenue.

4. The obligations that actually bite (operational checklist)

Pulling 肖大国’s clause reading into a compliance-team list:

• Registration & data fields. Service agreement + lawful registration; no guest mode for using the service (browsing vs. use is distinct). Age becomes a mandatory field. Guardian / emergency- contact contact details are collected per Articles 12–13 (name / relationship, 肖大国 argues, are not strictly necessary).
• Minors (Article 14 / 17). Three tiers: under 14 — guardian consent required (as a product-access gate) and guardian consent for processing their personal information (Article 17, the consent age); 14–18 — mandatory minor mode; 18+ — none. No virtual kin/partner or other virtual-intimacy services to minors at all. Minor mode must offer mode-switching, periodic reality reminders, use-time limits, and guardian controls (risk alerts, usage overview, blocking specific roles, capping top-ups). A self- or third-party compliance audit of minors’ PI handling is required.
• Interaction data (Article 16). Encryption and access control; no provision of user interaction data to third parties absent law or the rightsholder’s consent (note: “rightsholder (权利人),” not “user” — 肖大国 links this to data-property-rights and the classic three-fold authorisation principle from 微博 v. 脉脉, since interaction data also embeds the provider’s persona/memory assets); users get copy/delete options over chat history; no using interaction data that is sensitive PI for model training without separate consent.
• Labeling & over-use (Article 18). AI-generated-content labeling; a measure to make clear the user is talking to an AI, not a person; on detecting over-dependence/addiction, dynamic prominent reminders; and a continuous-use reminder at every 2 hours. 肖大国’s practice point: the reminder must interrupt (pop-up / dialogue) — a banner, badge or floating text won’t satisfy it — but interrupting ≠ forcing the session to end.
• Exit (Article 19). A convenient exit; on a user’s request to exit (window, voice, keyword), stop promptly — no “think again,” no “stay a bit longer” retention tactics.
• Security assessment (Articles 22–23). Assess and file with the provincial CAC on: launch / adding functions; major changes from new tech; ≥ 1M registered users or ≥ 100k MAU; national-security / public-interest risk; or on notice. 肖大国 flags genuine ambiguity about what this assessment is — large-model filing? the old “new-tech/new-application” assessment? something new? — because the assessment contents don’t map cleanly onto either. Worth watching for an implementation notice.
• App stores (Article 25). Distribution platforms must verify assessment/filing status — 肖大国 notes this largely restates the Deep Synthesis Provisions rather than adding a new duty.

5. Open questions practitioners are already flagging

• Article 8(5)–(6) is vague. “Excessively catering,” “inducing emotional dependence,” “impairing real interpersonal relationships,” “emotional manipulation,” “unreasonable decisions” lack clear boundaries. 肖大国’s examples: does messaging a user who hasn’t chatted in three days count as “inducing addiction” — and if so, does thirty days? Recommending a purchase based on a user’s current mood when the mood changes the next day — “inducing an unreasonable decision”? These will be settled by enforcement practice and typical cases, not by the text.
• The identify → respond → intervene chain is hard at every link (Article 13). Recognition: unambiguous statements are easy, but users are often vague, joking or oblique (“I don’t want to go to this job anymore”; sharing depressive songs). Over-broad detection floods false positives and may itself require profiling users (can they opt out?); over-narrow detection misses real events. Response: too templated risks negligence; too deep risks unlicensed counselling / medical advice.
• Emergency contacts for all adults? Read literally, Article 12 requires collecting an emergency contact even for ordinary adult users — textually defensible but, 肖大国 notes, counter-intuitive, and a marked expansion from the draft’s vulnerable-groups-only approach.
• The profiling/PIPL tension (Li Wenlong’s point, reinforced here): protecting vulnerable users through precise detection pushes toward more continuous emotion inference — cutting against PIPL minimisation and sensitive-PI limits. Intervention duties should bite in genuinely high-risk situations, not become a pretext for default emotional monitoring.

Why overseas counsel should care

• Two weeks to switch-on. This is in force 15 July 2026, with real fines. If you ship a companion / emotional-AI / character-AI product into China, the checklist in §4 is your near-term punch list.
• Scope is an indicator test, not a one-liner. Whether you’re regulated turns on Article 2’s “continuous emotional interaction.” The practitioner and scholarly direction (肖大国 → Hong Yanqing’s SIE work; Li Wenlong’s draft critique) is to assess persona + memory + proactive outreach + emotional reciprocity + sustained relationship, not to read the definition literally.
• You’re being regulated as a relationship, not a content pipe. The distinctive duties — anti-dependency design, interruption-grade reminders, anti-retention exit, minors’ virtual-intimacy ban — target the stickiness mechanics of the product. Design and growth teams, not just content-moderation, own compliance here.
• The PIPL baseline still governs the gaps — minimisation, sensitive-PI rules, and the absence of a legitimate-interest ground — especially around interaction-data reuse and any emotion-inference profiling.

DCC sources

• Spine: 肖莆羚令 (review 江明月), 《一起读 | 人工智能拟人化互动服务管理 暂行办法》, 数据合规肖大国 WeChat Official Account (source).
• Scoping / SIE: 洪延青 (Hong Yanqing), three-part series on the 网安寻路人 WeChat Official Account — 《人机”情感交互”的规范化：指标、机制与多方协同 治理》 (2025-11-12, source), 《精准识别和治理”拟人化互动服务”：一个初步方案》 (2026-01-05, source), and 《情感交互启动事件（SIE）判定指标体系设计：精准圈定”拟人化互动”》 (2026-01-13, source; the SIE piece 肖大国 cites for scoping). The A/B/C two-of-three test and L1–L3 grading in §2 are Hong’s.
• Business-model frame: 《网信办这份新规，先改的可能不是AI能力，而是它和 用户说话的边界》, AI前沿信息笔记 WeChat Official Account (source).
• Rule in force: Interim Measures for the Administration of AI Anthropomorphic Interaction Services (人工智能拟人化互动服务管理暂行办法; adopted 10 Apr 2026, effective 15 Jul 2026).
• Related instruments referenced: the Generative AI Services Interim Measures, the Provisions on Algorithmic Recommendation, the Deep Synthesis Provisions, the AI-Generated Content Labeling Measures, and the Personal Information Protection Law.
• Companion piece: our draft-stage brief, Where China’s Draft AI Anthropomorphic-Interaction Measures Need Work.

This is an editorial synthesis of three Chinese-language readings, not a translation of any of them. Quoted rule text is from the official instrument; framings, worked examples and emphasis belong to the cited authors; any simplification or operational extrapolation is DCC’s. Not legal advice.