Signal: Three Gates Close In Nineteen Days — The Pre-Release Regime Goes Global

China’s new rules for emotionally interactive AI took effect on July 15, opening a 19-day sequence in which the United States and European Union will reach separate AI oversight deadlines. The cluster matters for developers serving several markets, but the measures are not three equivalent approval gates: China imposes binding service rules, Washington is building a voluntary national-security review process, and the EU regulates market access through a broader risk-based law.

China’s Interim Measures for Artificial Intelligence Anthropomorphic Interaction Services, issued by five agencies led by the Cyberspace Administration of China, now cover services offered to the Chinese public that simulate human personality and communication through sustained emotional interaction. Providers must identify the service as AI, address dependency and psychological-safety risks, protect interaction data and restrict virtual intimate relationships for minors. The official text excludes customer service, knowledge tools, work assistants and research products that do not involve sustained emotional interaction.

In the United States, Executive Order 14409 gives federal agencies until August 1, 60 days after its June 2 signing, to develop a classified benchmarking process for advanced cyber capabilities and design a voluntary framework with AI developers. Participating companies could give the government access to designated covered frontier models for up to 30 days before sharing them with other trusted partners. The order directs the National Security Agency, working with other officials, to make model designations, but expressly rejects mandatory federal licensing or preclearance.

Most remaining provisions of the EU AI Act are scheduled to apply on August 2, including transparency duties affecting chatbots and certain synthetic content. General-purpose AI obligations have applied since August 2025. The EU has separately completed legislative approval of the Digital Omnibus on AI, which moves stand-alone high-risk system rules to December 2, 2027, and product-embedded high-risk rules to August 2, 2028. As of July 19, the signed amendment was still awaiting publication in the Official Journal, the final step before it enters into force.

Global Timelines Split by Design

The three dates show a shared policy preference for government scrutiny earlier in an AI product’s life, yet the authorities are examining different risks. China focuses on content, emotional dependency and social stability. The EU centers product safety, transparency and fundamental rights. The US process targets advanced cyber capabilities and national security while relying on voluntary participation.

For companies operating across borders, one model release may require several compliance tracks. Product scope, documentation, testing and release schedules may differ by market. Large developers may be better placed to absorb that work, while smaller laboratories could face higher relative costs. A model’s acceptance under one system also offers no automatic clearance elsewhere.

Oversight Moved Earlier

China has regulated public generative AI services since 2023 through rules that can require security reviews and algorithm filings. The July measures add service-specific duties for AI companions and other products built around continuing emotional relationships, rather than placing every AI agent under the same regime.

The EU AI Act entered into force in August 2024 and has applied in stages: prohibited practices began in February 2025, followed by governance and general-purpose model duties in August 2025. The US executive order takes a narrower course. Its 30-day access window concerns covered frontier models and trusted partners, not every commercial AI release.

“Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.”

— US Executive Order 14409

Scope and Enforcement Still Vary

Several operational details remain unsettled. The US benchmark and designation criteria will be classified, and it is not yet clear which developers will participate or how trusted-partner status will affect federal procurement. China’s enforcement practices under the new companion-AI rules are still emerging. In Europe, Official Journal publication will finalize the Omnibus timetable. Open-weight publication is not automatically exempt across all three markets; coverage can depend on where a provider operates, offers a service or places a model on the market.

August Deadlines Test Implementation

Attention now shifts to the August 1 US agency deadline and the August 2 EU application date. Developers will watch for details on American participation, NSA model designations and EU enforcement guidance. Regulators and companies will also track China’s first enforcement actions, which should show how broadly authorities interpret sustained emotional interaction.

Key Questions

Why is this described as a 19-day sequence?

The period from July 15 through August 2 contains 19 calendar dates when counted inclusively. The elapsed time between the first and final milestones is 18 days.

Does the United States require approval before releasing frontier models?

No. Executive Order 14409 calls for a voluntary framework offering government access for up to 30 days. It expressly disclaims mandatory licensing or preclearance.

Does China’s new rule cover every AI agent?

No. It targets services providing sustained emotional interaction that imitates human traits. Ordinary work assistants, customer-service tools and knowledge systems are excluded when they lack that feature.

Is the entire EU AI Act starting on August 2?

Most remaining provisions apply then, but the law already entered into force in stages. The approved Omnibus moves key high-risk system deadlines to December 2027 and August 2028, subject to Official Journal publication.

Source: Thorsten Meyer AI