In July 2026, China became the first country to write dedicated rules for AI that acts like a person. ByteDance's Doubao and Alibaba's Qwen responded by disabling their humanlike custom-agent features outright rather than rebuilding them to comply. The regulation targets a specific thing — AI "companion" services that simulate a personality and hold a sustained, emotional, one-to-one relationship with a user — and it explicitly leaves assistants, Q&A bots, and productivity tools alone. That line, between the agent that keeps you company and the agent that does a job, is the single most important idea for anyone producing content with AI avatars and synthetic personas, because it tells you where the regulatory pressure is actually pointed and why broadcast persona content sits on the safe side of it. This guide breaks down what the rules say, the taxonomy every creator should internalize, the wider disclosure landscape from the EU AI Act to the FTC and platform labels, and how to build a persona-content operation that stays durable as transparency requirements only tighten.
On July 15, 2026, China's Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services take effect — the first regulation anywhere written specifically for AI that acts like a person. In the run-up, ByteDance's Doubao and Alibaba's Qwen disabled their humanlike, user-customizable AI agent features rather than rebuild them to fit the new requirements, with Qwen's humanlike and user-created agents reported to go offline around July 10 and both firms' broader agent functions on July 15. The full news timeline is covered in [ByteDance and Alibaba are disabling their humanlike AI agents](/news/bytedance-alibaba-disable-humanlike-ai-agents); this guide is about what the shift means for anyone producing content with AI avatars and synthetic personas.
The one idea to carry out of the whole episode is the line the regulators drew. The rules go after the agent that keeps you company — a chatbot that simulates a personality and holds a sustained, emotional, one-to-one relationship with a user — and they explicitly leave alone the agent that does a job: customer-service bots, knowledge question-and-answer tools, workplace assistants, and education and research tools, as long as those do not slide into sustained emotional interaction. Producing content with an AI persona is a third thing again, and it is the safest of all: a one-way broadcast to an audience, not a private relationship with an individual. Understanding those three modes — companion, assistant, broadcast — is what lets a creator read every future rule correctly instead of panicking at the headline.
The measures were reported to have been co-issued on April 10, 2026 by the Cyberspace Administration of China alongside several partner agencies — including the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the market regulator — and to take effect on July 15, 2026. The scope is narrow and specific: services that "simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction." That last phrase is load-bearing. It is not "AI that talks like a human"; it is AI built to sustain an emotional relationship over time.
For services that fall inside that scope, the obligations are heavy. Providers must run anti-addiction systems and detect unhealthy dependence in real time, issue mandatory usage notifications, and offer an instant exit from the interaction. They are barred from offering virtual-companion or virtual-family-member services to minors and must obtain guardian consent before serving users under 14, with dedicated "minor modes" that cap usage time and nudge users back toward real-world interaction. And they must detect users in acute distress — signs of self-harm, suicidal behavior, or serious financial loss — and intervene, escalating to guardians or emergency contacts. These are the requirements of a duty-of-care regime for something regulators have decided can form a genuine emotional bond, and they are why ByteDance and Alibaba chose to pull the features rather than build all of that. The stated concerns behind the rule read the same way: dependence and addiction, privacy leakage, exposure to extremist ideas, and mental and physical harm.
Note what is not on the list. Nothing in the rule restricts making a video with an avatar, generating a branded image with a consistent AI face, or publishing a post written in a persona's voice. The regulated surface is the interactive, emotional, one-to-one relationship — not the synthetic identity itself and not the content it produces.
Because the coverage collapses everything into "humanlike AI," it is worth separating the three modes cleanly, since they carry completely different regulatory weight.
A companion agent is a chatbot a user talks to privately, over time, that role-plays a personality and is designed to feel like a relationship. This is what Doubao's and Qwen's custom agents let people build, and it is squarely what the rule targets. The risk model regulators care about — dependence, emotional manipulation, harm to minors, failure to catch a user in crisis — only exists because the interaction is sustained, two-way, emotional, and one-to-one. If your use of AI does not have all four of those properties, you are almost certainly not the target.
An assistant does a task: answers a question, drafts a document, files a ticket, tutors a lesson. It can be conversational and even personable, but its purpose is utility, not companionship. The regulation carves these out by name precisely because the emotional-dependence risk is low. Most business uses of AI live here, and they are untouched by the anthropomorphic rule as long as they do not drift into sustained emotional interaction dressed up as productivity.
Broadcast content is the mode most creators actually operate in and the one that barely intersects the rule at all. You use a synthetic persona — an avatar, a face-locked identity, a defined voice — to produce videos, posts, carousels, and graphics that you publish to an audience. There is no private relationship, no two-way emotional loop, no individual user forming a bond with the persona. It is a one-to-many broadcast, the same structure as a brand mascot, a spokesperson, or a channel host who happens to be generated. The mechanics of building that kind of consistent synthetic identity are covered in [identity-first AI video](/guides/identity-first-ai-video) and [AI avatars in video](/guides/ai-avatars-in-video); the point here is jurisdictional: broadcast persona content sits outside the companion-agent rule by construction.
Hold those three apart and the news stops being alarming. A rule that disables companion agents in China says nothing about your ability to publish avatar videos to nine platforms. It does, however, carry a signal worth heeding — which is the next section.
If broadcast persona content is safe, why care at all? Three reasons, none of which is "the rule bans your work."
First, dependency risk. The Doubao and Qwen shutdowns are a clean case study in how a platform feature can vanish overnight — not because the business failed, but because a regulator moved. Anyone who built a workflow, an audience touchpoint, or a product on those agent features got a hard deletion deadline with no migration path in Alibaba's case. The lesson generalizes: building your content operation on a single external feature — one avatar vendor, one model, one platform tool — means a rule change or a strategy shift can strand you. This is the same single-vendor exposure that hit creators when [OpenAI shut down Sora](/news/openai-sora-shutdown), arriving through regulation instead of a product wind-down.
Second, the disclosure climate. Even though this specific rule is about companion agents, it lands in a broader moment where "humanlike AI" is under scrutiny and audiences are primed to react to anything that feels like deception. The wider backlash against AI-forward content — documented in [the AI marketing backlash](/guides/ai-marketing-backlash) — and the specific unease about synthetic personas manipulating trust, covered in [the AI influencer manipulation trend](/guides/ai-influencer-manipulation-trend), mean the reputational cost of an undisclosed or deceptive AI persona is rising independent of any statute. Regulation is a lagging indicator of that unease, not the whole of it.
Third, the direction of travel. China is first to write a dedicated anthropomorphic-interaction rule, but it is not operating in a vacuum. The obligations pointing at synthetic identities and AI interaction are multiplying, and a persona-content operation built today should assume they keep growing rather than betting they stall.
Zoom out from the single Chinese rule and the pattern is transparency, not prohibition. The EU AI Act requires that people be told when they are interacting with an AI system unless it is obvious, and it requires providers to mark synthetic audio, image, video, and text as artificially generated in a machine-readable way — a labeling obligation that reaches deepfakes and AI-generated media directly. In the US, there is no single federal AI law, but the FTC has pursued deceptive AI endorsements and impersonation, and a patchwork of state rules touches synthetic likeness and disclosure. And the platforms themselves — the surfaces you actually publish on — increasingly require creators to flag AI-generated or AI-altered content, with their own labels applied on top.
The through-line across all of it is that you are rarely told you cannot use a synthetic persona; you are told to be honest that it is synthetic. That is a very different compliance posture than "stop," and it is one a well-run persona operation can meet without changing what it makes — only how transparently it labels and frames it. The practical implication: disclosure is a design default to bake in now, not a fire to fight later.
Translate all of that into how you actually run an avatar-and-persona operation, and a handful of concrete practices fall out. None of them requires abandoning synthetic personas — they require using them as a broadcast medium with honesty and governance built in.
The cleanest way to stay outside the companion-agent rules is to never build a companion agent. Use your persona to produce and publish content, not to run sustained one-to-one emotional conversations with individual users. If you would not put a human spokesperson in that role, do not put a synthetic one there either. Broadcast is not a limitation here; it is the safe, scalable shape of persona content, and it is where the audience and the revenue actually are.
Label AI-generated content according to the platform and jurisdiction you are publishing into, and do it as a standing habit rather than a case-by-case decision. Disclosure is cheap, it is increasingly required, and it defuses the exact "deceptive synthetic media" concern that draws both regulators and audience backlash. A brand that is upfront about using an AI persona is in a far stronger position than one caught pretending otherwise.
The sharpest legal and ethical edge in synthetic media is using a real person's face or voice without permission. Keep your persona a defined, owned identity — a consistent synthetic character that is yours — rather than a clone of someone who has not agreed to it. This keeps you clear of likeness and impersonation exposure and of the reputational damage that follows a non-consensual deepfake, and it is simply the honest way to build a persona brand.
Automation is fine; unattended emotional or brand-defining output going live with no human looking at it is where things go wrong. A review step before anything customer-facing publishes is what catches the off-brand generation, the tone-deaf take on a sensitive topic, and the disclosure you forgot to add. It is also, not coincidentally, the same governance the winning brands use to stay on the right side of the [AI marketing backlash](/guides/ai-marketing-backlash).
Read the four practices back and they describe a particular kind of tool — one whose entire architecture is broadcast persona content with governance and a human gate, rather than an interactive companion. That is what [Kompozy](/) is, and the fit with the regulatory line is structural, not marketing. Kompozy's AI Influencer personas exist to render finished, published output: Persona Shorts and HeyGen avatar video, face-locked Persona Photos and Persona Tweets, brand-exact carousels and quote graphics, plus text posts, blogs, and newsletters — eighteen formats fanned across nine social platforms plus email and blog. Every one of those is a one-to-many broadcast. There is no private conversation, no sustained emotional loop, no individual user forming a bond with the persona. By construction, it operates in the mode the anthropomorphic-AI rules leave alone.
The governance the tightening climate rewards is also built in, not bolted on. A [Persona Brief](/glossary/persona-brief) defines one owned, consistent identity and voice, so your persona is a deliberate synthetic character rather than a drifting default or an unauthorized clone of a real person. Gemini face-lock keeps that identity visually consistent across clips instead of morphing into someone recognizable by accident. And a per-post review pipeline keeps a human gate on what publishes — you can run trusted sources on autopilot, but a person approves the batch before it ships, which is exactly where a disclosure check or a sensitivity call belongs. That combination — broadcast structure, an owned persona, a review gate — is what lets you produce synthetic-persona content at real volume while staying firmly on the content-tool side of every line the regulators have drawn.
The honest boundary: no tool makes disclosure decisions or reads every jurisdiction's rules for you, and Kompozy does not claim to. Labeling AI content correctly, deciding what to disclose, and not impersonating real people are your calls — the engine gives you the broadcast shape and the review gate that make those calls easy to execute, but the judgment stays with you. What it removes is the temptation to cut corners for throughput: because generating and publishing on-brand persona content at scale is already handled, there is no pressure to reach for the riskier, murkier uses of synthetic AI to keep up. You get the volume from the safe mode.
China's anthropomorphic-AI rules are a milestone — the first dedicated attempt to govern AI that acts like a person — and ByteDance and Alibaba pulling Doubao's and Qwen's humanlike agents to comply is a real signal that the era of unregulated synthetic interaction is ending. But the rule is precise about what it targets: the companion agent that holds a sustained, emotional, one-to-one relationship, not the assistant that does a job and not the persona used to broadcast content to an audience. For creators working with avatars and synthetic personas, the takeaway is not fear; it is orientation. Stay in broadcast mode, disclose honestly, own your persona instead of cloning a real one, keep a human on the final yes, and avoid single-vendor dependency — and the tightening rules become a set of guardrails you were already inside, not a wall in your way.
The Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, effective July 15, 2026, target AI "companion" services — assistants built to simulate a human personality and provide sustained, emotional, one-to-one interaction. They require anti-addiction and dependence-detection systems, usage notifications, an instant exit, minor protections, and distress intervention. They explicitly exclude customer-service bots, knowledge Q&A tools, workplace assistants, and education and research tools that do not involve sustained emotional interaction.
No. The regulation is about interactive companion agents — a two-way, one-to-one emotional relationship between a user and a chatbot. Using an AI avatar or persona to generate videos, posts, and graphics you broadcast to an audience is a one-way content medium, not a companion service, and sits outside the rule. The regulated activity is emotional companionship, not producing published content with a synthetic identity.
Both companies chose to remove Doubao's and Qwen's humanlike custom-agent features around July 10–15, 2026 rather than rebuild them to meet the new anti-addiction, distress-intervention, and minor-protection requirements. Compliance for a genuine companion product is heavy — real-time dependence detection, guardian consent flows, mandatory exits — so pulling the feature was the faster path. Users were told to export agent data before deletion deadlines.
The direction of travel points that way, though the exact form varies. The EU AI Act already requires disclosure when a person is interacting with an AI system and labels for synthetic media, the US FTC has pursued deceptive AI-endorsement and impersonation cases, and platforms increasingly require AI-content labels. China is first to write a dedicated anthropomorphic-interaction rule, but transparency and companion-safety pressure is broad. Build on the assumption that disclosure requirements only grow.
Stay on the broadcast side of the line, disclose AI use where it matters, and govern the output. Use synthetic personas to produce content you publish to an audience rather than to run one-to-one emotional companion interactions; label AI-generated content per platform and jurisdiction rules; keep a consistent, honest brand identity rather than impersonating a real person without consent; and keep a human review gate on anything customer-facing. A governed engine like Kompozy is built around exactly that broadcast-and-review shape.
China's Interim Measures on anthropomorphic AI interaction, effective July 15, 2026, are the first dedicated rules for AI that simulates a human personality — and they target companion services that hold sustained, emotional, one-to-one relationships, not content tools. ByteDance's Doubao and Alibaba's Qwen disabled their humanlike custom agents to comply. For creators, the key line is assistant-and-broadcast (allowed) versus emotional companion (restricted): using an AI persona to publish content stays on the safe side.
Get started → · ← All guides · Compare Kompozy vs other tools