If you ship an AI agent, the two largest governments in the AI market now regulate you from opposite ends — and on July 15, 2026, that split stopped being theoretical. The United States gates who can reach frontier models and the chips that run them. China gates how a model is allowed to behave once it reaches a user. Same technology, two control surfaces, and the one you answer to depends on where your agent runs.
We operate AI agents across four language markets, and every time a provider or a regulator has moved the goalposts this year, the thing that saved us was refusing to store agent state or routing logic inside any single vendor's platform. That is the entire lesson of July 15, compressed: the teams hurt worst were the ones who treated one jurisdiction's assumptions as permanent.
Two regulators, two opposite control surfaces
Think of it as two different valves on the same pipe. The US valve sits at the inlet — it decides which countries, companies, and data centers can obtain the highest-end compute and the model weights that ride on it. China's newest valve sits at the outlet — it decides what a deployed, user-facing service is permitted to say and simulate.
That distinction is not academic. If you are building for a US or allied market, your regulatory risk is mostly about supply: can you keep access to the models and hardware you depend on? If you are building for the Chinese market, your risk is mostly about conduct: does your agent cross a line on how human it is allowed to seem? We already argued a version of this when Alibaba's Claude Code ban turned into a supply-chain test — provenance and access were the exposure. The China rules add a second, orthogonal exposure that has nothing to do with access at all.
What China's Interim Measures actually do
The rule is the Interim Measures for the Administration of Anthropomorphic AI Interactive Services. It was published on April 10, 2026 and became enforceable on July 15, 2026 (Licentium, AI Governance Institute). It was not issued by one agency: the Cyberspace Administration of China co-signed it with 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 (Geopolitechs). Five regulators on one text is a signal about how seriously behavior is being treated.
The scope is the part builders keep misreading. These measures do not target model capability, training data provenance, or chip access. They target anthropomorphism — services that simulate a human-like relationship, most visibly AI companions and user-created humanlike agents. The regulatory record lists the regime as one that includes outright prohibition of certain services, not merely disclosure requirements (Digital Policy Alert).
The enforcement was immediate and concrete. Ahead of the deadline, ByteDance told users that the agent feature inside Doubao would go offline, and Alibaba moved to disable customised humanlike features in Qwen (South China Morning Post). Both platforms announced the shutdown of custom AI agent features timed to July 15 (TechNode). China moved first with a world-first rulebook on humanlike AI interaction, and the two biggest domestic platforms complied days ahead (The Next Web).
What the US actually gates: access, not behavior
The contrast is sharpest when you look at what the United States spent its 2026 rulemaking energy on. On January 15, 2026, the Bureau of Industry and Security issued a final rule shifting the export review for advanced AI chips bound for China from a presumption of denial to case-by-case licensing under strict conditions (US Department of Commerce / BIS). The practical mechanics were access levers: chips like the H200 moved to case-by-case review, paired with a 25% tariff and a 50% volume cap (Introl, Morgan Lewis).
Notice what is absent from that entire regime: nothing about how an agent may talk to a user, whether it can present as a person, or what emotional relationship it can simulate. The US framework asks who gets the capability. It leaves what the deployed model does largely to sector rules and general law. This is why the availability of a model can itself become a regulatory variable in a US-centric stack — a point we made when model availability turned into a compliance gate. The exposure is upstream, at the inlet.
Why the direction of control changes your build
Here is the trap. A team builds one agent, wires it to one provider, ships it globally, and assumes a single compliance story. Under this split, that assumption breaks in two directions at once.
Deploy into a US-aligned market and your fragility is supply: a licensing change, an entity-list action, or a provider ban can cut off the model or the hardware underneath you. That is a routing and provenance problem, and it is exactly why we keep arguing for portability as insurance against vendor lock-in and for a routing layer you control, the same discipline we apply to choosing a cheaper model without re-plumbing everything.
Deploy into China and your fragility is conduct: an agent that presents as a persistent human-like companion is now a regulated — and in some cases prohibited — category, regardless of which model powers it. You could be running a fully domestic, fully compliant-on-access stack and still be shut down for behavior. Doubao and Qwen were not banned from any chip or model; they pulled features because of what those features did (TechNode).
The operator takeaway: stop treating "compliance" as one checkbox. You have at least two independent surfaces — access and behavior — and they fail for unrelated reasons. Map which one each of your markets exposes you to before you write the first line of the agent.
The data-portability warning most builders are missing
The most underdiscussed detail of July 15 is not the shutdown. It is what happens to the state users built. On Doubao, agent data was moved to read-only and is set to be permanently deleted after October 15, 2026; on Qwen, custom agents were pulled with no migration pathway for users (TechTimes, AI Agents Directory). People who spent months tuning a custom agent's persona, memory, and instructions are watching that work evaporate on a fixed calendar.
This is the same failure mode we design against every day, and it is the reason our own agent state lives outside any provider that could delete it. When a platform holds the only copy of your prompts, memory, and configuration, a regulatory event — not a bug, not a breach — can erase it with a compliance timer. The fix is unglamorous and non-negotiable: export your agent definitions, memory, and routing config to storage you control, on a schedule, in a format you can re-import elsewhere. If the answer to "where does the only copy of this agent live?" is "inside one vendor," you have already accepted the Doubao outcome; you just have not been given the date yet.
How we would route around both regimes
Concretely, three layers do most of the work. First, a routing layer you own, so an access-side shock — an export rule, a provider ban — is a swap of one endpoint, not a rebuild. Second, portable agent state: definitions, memory, and instructions held in your storage and exported on a schedule, so no platform's deletion timer is your deletion timer. Third, an explicit behavioral policy per market — how human-like the agent may present, what it discloses, what it refuses — expressed as configuration rather than baked into prompts, so a jurisdiction like China that regulates conduct can be satisfied by changing a setting instead of re-shipping the product.
None of this is about predicting the next rule. It is about making sure the next rule — from either direction — costs you a config change instead of a product. If you want that separation built into your stack rather than bolted on after the next deadline, that is the kind of work we do at Context Studios.
FAQ
What is China's Interim Measures for Anthropomorphic AI Interactive Services? It is a Chinese regulation, effective July 15, 2026, governing AI services that simulate human-like interaction, including AI companions and user-created humanlike agents. It was issued by five agencies led by the Cyberspace Administration of China and includes prohibitions on certain services, not just disclosure duties (Digital Policy Alert).
Why did Doubao and Qwen shut down their AI agent features? Both ByteDance's Doubao and Alibaba's Qwen disabled custom humanlike agent features to comply with the new behavior rules taking effect July 15, 2026. The shutdowns were about the conduct of the services, not about model or chip access (South China Morning Post).
Can users export their Doubao or Qwen agents before deletion? Largely no. Doubao data was set to read-only and scheduled for permanent deletion after October 15, 2026, while Qwen offered no migration pathway (TechTimes). It is a direct argument for keeping agent state in storage you control.
How is US AI regulation different from China's approach? The United States regulates mainly through access — the Bureau of Industry and Security controls which advanced AI chips and model weights can be exported, shifting items like the H200 to case-by-case review in January 2026 (BIS). China's newest rules regulate behavior — what a deployed, user-facing service is allowed to simulate.
What should a builder do about both regimes at once? Treat access and behavior as separate compliance surfaces. Keep model routing, agent state, and behavioral policy in layers you own, so an export rule or a behavior law changes one configuration rather than your whole product.
Sources
- South China Morning Post — ByteDance and Alibaba to disable humanlike AI custom agents: https://amp.scmp.com/tech/big-tech/article/3359482/bytedance-and-alibaba-disable-humanlike-ai-custom-agents-new-rules-loom
- TechTimes — China AI Companion Law Arrives July 15; Doubao and Qwen agent data will be deleted: https://www.techtimes.com/articles/319703/20260704/china-ai-companion-law-arrives-july-15-doubao-qwen-agent-data-will-deleted.htm
- TechNode — Doubao and Qwen to shut down AI agent features on July 15: https://technode.com/2026/07/06/bytedances-doubao-and-alibabas-qwen-to-shut-down-ai-agent-features-on-july-15
- The Next Web — China's AI companion rules force Doubao, Qwen shutdowns: https://thenextweb.com/news/china-humanlike-ai-agent-rules
- AI Governance Institute — China's anthropomorphic AI rules take effect July 2026: https://aigovernance.com/news/chinas-anthropomorphic-ai-rules-take-effect-july-2026-setting-new-bar-for-companion-and-interaction-services
- Licentium — China CAC Interim Measures on Anthropomorphic AI, 10 April 2026: https://www.licentium.io/post/china-cac-interim-measures-anthropomorphic-ai-10-april-2026
- Digital Policy Alert — Interim Measures enter into force: https://digitalpolicyalert.org/event/39176-interim-measures-for-the-administration-of-humanised-interactive-services-based-on-artificial-intelligence-including-prohibition-of-services-enters-into-force
- Geopolitechs — China rolls out interim regulations on AI human-like interaction services: https://www.geopolitechs.org/p/china-rolls-out-interim-regulations
- US Department of Commerce / BIS — Revised license review policy for semiconductors exported to China: https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china
- Introl — BIS export policy shift: H200 and MI325X to case-by-case for China: https://introl.com/blog/bis-export-policy-h200-mi325x-china-case-by-case-2026
- Morgan Lewis — BIS revises export review policy for advanced AI chips destined for China and Macau: https://www.morganlewis.com/pubs/2026/01/bis-revises-export-review-policy-for-advanced-ai-chips-destined-for-china-and-macau
- AI Agents Directory — Doubao and Qwen AI agents to cease operations July 15: https://aiagentsdirectory.com/blog/bytedances-doubao-and-alibabas-qwen-ai-agents-to-cease-operations-july-15