On July 15, 2026, China's first dedicated framework for "anthropomorphic AI interaction services" takes effect — and Doubao and Qwen, the two platforms with the most to lose, chose to shut down user-created humanlike agents entirely rather than retrofit them at scale. The compliance logic is rational; the downstream consequences for the Western developers building on these models are not yet fully mapped. What remains genuinely open is how broadly the exemptions are drawn, and whether this regulatory frame migrates outward.
Let's start with the concrete sequence of events, because the headlines have scrambled the causality.
In late December 2025, the CAC — China's top internet regulator — circulated draft rules for public comment. The comment period closed January 25, 2026. On April 10, 2026, five agencies jointly published the final text: the CAC, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation. The July 15 effective date gave platforms roughly three months to comply.
Doubao's "Character" feature lets users build custom AI personas — name them, give them backstories, set their conversational styles — and in certain configurations, deploy those agents to interact with third parties who didn't create them. Qwen's developer ecosystem has enabled similar functionality for thousands of Chinese enterprise apps. Neither product configuration, as designed, could satisfy the new regulation's core requirement without fundamental rearchitecting: mandatory AI identity disclosure at login and at two-hour intervals throughout every interaction.
The industry's reading of why Doubao and Alibaba chose shutdown over retrofit: both platforms have accumulated hundreds of thousands or millions of user-generated agents. Each one would need to be individually reviewed, modified, and certified under the new compliance framework. That's operationally impossible to execute in three months. Rebuilding the feature from scratch — with disclosure baked in from the model layer up — is a cleaner engineering path, even if it means the interim period has no product. It's a rational calculation, not idealism about regulatory intent.
This is where Western observers tend to either over-read or under-read the story.
The Interim Measures apply to AI systems that "simulate human personality traits, modes of thinking, and communication styles" and/or "engage in emotional interaction with humans" through text, images, audio, video, or other means. Read that carefully: both conditions apply independently. A text-based chatbot that simulates a personality but involves no emotional content might still fall in scope. An emotional support AI that doesn't simulate human personality might also be covered.
What is explicitly out of scope: smart customer service, informational Q&A, work assistance, study and education tools, and scientific research applications — provided they do not involve "ongoing emotional interaction." That carve-out is doing a lot of work, and it's already generating compliance anxiety in China's enterprise AI sector. A sales assistant that comforts a frustrated customer: in scope or out? A tutoring agent that provides encouragement: same question. Chinese legal analysis published by Mayer Brown flags this ambiguity explicitly — the exemptions are defined by purpose and design intent, not by technical features, which means enforcement discretion is high.
What the regulation does not do: restrict the underlying models. Qwen's open-weight releases remain available. Alibaba Cloud's model APIs are not being pulled. The crackdown targets the product configuration layer — specifically the combination of user-defined persona, third-party interaction, and persistent non-disclosure — not the LLM capability beneath it.
To understand why this matters beyond China's borders, the numbers help.
Doubao crossed 345 million monthly active users in Q1 2026. For context: that's roughly comparable to the population of the United States using a single AI app, in a market where OpenAI cannot operate directly. ByteDance built that user base in roughly 18 months through aggressive subsidisation of compute costs, tight integration with TikTok's recommendation infrastructure, and — relevantly — features like AI companions and custom character creation that drove retention in a way that a standard chatbot doesn't.
Qwen is a different kind of story. Alibaba has published model weights across multiple generations — Qwen2.5, then Qwen3, now into Qwen3-Coder variants — on Hugging Face and Alibaba's own ModelScope. Qwen3-235B-A22B, the current flagship, reportedly benchmarks ahead of both DeepSeek V3.2 and Llama 4 on reasoning and general intelligence tasks. The Qwen API powers a large fraction of Chinese enterprise AI deployments via Alibaba Cloud. Those third-party apps inherit whatever compliance constraints the platform layer introduces — which means the July 15 shutdown touches a surface area well beyond Alibaba's own consumer products.
This is also the context for understanding the chip layer underneath all of this. Beijing's push to domesticate its AI stack — from Huawei's Ascend accelerators taking share as Nvidia sales in China stall to Qwen's open-weight strategy reducing dependence on US frontier models — is a separate story, but it runs in parallel. Regulation and hardware localisation are both expressions of the same structural intent: a Chinese AI stack that Beijing can govern end-to-end.
The regulation doesn't exist in isolation. Here's how the main platforms look from a Western builder's vantage point:
| Platform | Owner | Underlying Model | Agent Features | Regulatory Status (July 2026) | Open to Western Devs? |
|---|---|---|---|---|---|
| Doubao | ByteDance (Beijing) | Doubao-pro series | User character creation: offline July 15 | Complying via shutdown and rebuild | Limited — no formal Western market launch |
| Tongyi / DashScope | Alibaba (Hangzhou) | Qwen series | Humanlike agents: offline July 10–15 | Complying via shutdown; API unaffected | Yes — Alibaba Cloud international |
| Qwen (open-weight) | Alibaba | Qwen3 family | N/A (model weights, not product) | Outside scope of Measures | Yes — HuggingFace, ModelScope |
| Kimi | Moonshot AI (Beijing) | Proprietary | Agent tools; compliance status unannounced | No public announcement as of July 6 | Limited English API access |
| Ernie Bot | Baidu (Beijing) | ERNIE series | Enterprise agent suite; status unclear | No public announcement as of July 6 | Yes — Baidu AI Cloud, limited |
| Doubao-pro API | ByteDance | Doubao-pro | Developer API; product layer affected | API likely unaffected; persona layer restricted | Limited — early international expansion |
Note: "Compliance status" for Kimi and Ernie Bot reflects the absence of public announcements as of this writing, not confirmed non-compliance. The pattern at the larger platforms — preemptive action before enforcement — is an industry norm that smaller players may or may not follow. Watch for announcements in the July 10–15 window.
The disclosure requirement will become a de facto global standard, on a slower timeline. China's two-hour interval rule is stricter than anything binding in the EU or US today. But the EU AI Act's agent provisions are still being operationalized, and the FTC has been active on AI identity deception cases without passing binding rules. The likely trajectory: the Chinese floor becomes the reference point that other regulators cite when they move, not the ceiling they surpass.
Qwen's international footprint complicates the picture. Because Alibaba publishes open-weight Qwen models, the model is already embedded in applications across the US, Europe, and Southeast Asia — well outside CAC jurisdiction. But as Qwen model versions become more tightly coupled with Alibaba Cloud's infrastructure and agent tooling, the line between "open model" and "platform product" blurs. Watch how Alibaba designs Qwen4 (or whatever comes next) — whether the agentic capabilities are baked into the weights or offered as a cloud service layer will have regulatory implications that mirror what just happened.
ByteDance will use this as a product reset, not a retreat. A forced shutdown of a feature is also an opportunity to rebuild it correctly. The "Character 2.0" version of Doubao's agent capabilities — built around disclosed, compliant personas — is likely in development now. If ByteDance launches Doubao in Western markets in 2027, it will arrive with disclosure-by-design as an architectural feature, which is a cleaner compliance pitch than most Western competitors can make.
The enterprise agentic market in China won't shrink — it will bifurcate. Consumer AI companions (high churn, high regulatory surface area) and enterprise workflow agents (lower compliance burden under the exemptions, clearer ROI) are effectively being separated by this regulation. The consumer persona market pauses and rebuilds; enterprise deployments continue under the carve-outs. That bifurcation accelerates a market transition that was already underway as the initial consumer novelty wave plateaued.
Watch whether Kimi and Baidu publish compliance statements before July 15. Moonshot AI (Kimi) and Baidu's Ernie have not announced public responses as of July 6. If they remain silent through enforcement day, they either believe their products are out of scope — or they're betting on non-enforcement in the first cycle. The CAC's enforcement history since 2021 suggests the bet carries real risk.
Q: Does this regulation affect Qwen open-weight models on HuggingFace? No. The Interim Measures apply to "providers" offering services "to the public within mainland China." Open-weight model weights published internationally — and the applications built on them outside China — fall outside CAC jurisdiction. If you download Qwen3 weights and run them on your own infrastructure in Europe or the US, nothing in this rule touches you. The caveat: if you deploy on Chinese cloud infrastructure or serve Chinese users, the deployment context brings you back in scope.
Q: How does this compare to EU AI Act rules on AI identity? Stricter, at the interaction-session level. The EU AI Act (Article 50) requires disclosure for AI systems interacting with humans in natural language in "reasonably foreseeable" circumstances — but the implementation guidance is still being finalized, and there is no equivalent of China's two-hour interval reminder. The Chinese rule is also enforced by five agencies simultaneously, which concentrates compliance risk more than the EU's distributed enforcement model.
Q: Why did ByteDance choose to shut down rather than retrofit? The likely reason, consistent with industry reporting: the volume of user-generated agents on Doubao is enormous, and each one would need to be individually reviewed and modified to meet the new standards. The compliance math doesn't work. Rebuilding the feature from the ground up — with disclosure as a first-class design requirement — is faster and cleaner. This is a pattern from other Chinese platform compliance cycles: when retrofitting is impractical at scale, shutdown-and-rebuild is the rational path.
Q: Does the "work assistance" exemption cover enterprise AI agents? Probably, for most enterprise use cases — but the exemption is defined by the absence of "ongoing emotional interaction," not by the presence of a business purpose. An enterprise sales agent that maintains relationship continuity, an HR tool that offers emotional support, a manager-assistant AI with a persistent "personality" — these may not qualify for the exemption. Chinese legal counsel with regulatory affairs experience is the appropriate guide here; this is not a line where pattern-matching against the definition text is reliable.
Q: What happens to existing Doubao agents and chat data? Users retain read-only access to agent configurations and chat histories until October 15, 2026. After that date, the data is deleted. Alibaba has not announced a comparable migration solution for Qwen users. If your team has anything built on Alibaba's product-layer agent tooling, the October 15 deadline is the hard data-loss cutoff.
Q: Is this the end of AI companion products in China? No — and the Carnegie Endowment's February 2026 analysis frames China's regulatory concern accurately: the worry is dependency and psychological harm, not the category itself. The regulation narrows the feature set, doesn't abolish it. Compliant AI companions — with persistent disclosure, minor protections, and no virtual intimate relationships — remain permissible. The market pauses and recalibrates. It doesn't disappear.
Q: Should Western founders be building on Doubao or Qwen APIs right now? Qwen API access via Alibaba Cloud remains available and is not directly affected by the shutdown. Doubao's developer API is less mature and has limited English documentation — and the product-layer features most affected by the regulation are exactly the differentiated ones that made Doubao interesting to persona-product builders. For model capability, Qwen open-weight models remain one of the most accessible and performant open options globally. For product-layer agent features, the Chinese platforms are, at minimum, in a rebuild cycle.
Alibaba and Tencent have spent the last eighteen months building the foundation for physical AI — shifting engineering resources from chatbot interfaces toward embodied systems, with Alibaba's Qwen model family...
Three of China's dominant AI platforms — ByteDance's Doubao, High-Flyer Quant's DeepSeek, and Alibaba's Qwen — approached the 2026 FIFA World Cup with measurably different prediction philosophies, and the divergence...