TL;DR
China’s companion-AI rules took effect on July 15, while US agencies face an August 1 deadline to design a voluntary frontier-model review framework. Most remaining EU AI Act provisions apply August 2, although lawmakers have delayed major high-risk requirements.
China’s new rules for human-like AI services took effect on July 15, beginning a 19-calendar-date span that also includes an August 1 US deadline for a voluntary frontier-model review framework and the August 2 application of most remaining EU AI Act provisions. The cluster matters for developers operating across major markets, but the three systems regulate different products, risks and stages of deployment.
China’s five-agency measure applies to services offered to the public in China that simulate human personality, thought or communication patterns through sustained emotional interaction. It covers products such as AI companions and emotional-support services, while excluding customer service, work assistants and similar tools that lack sustained emotional engagement. Providers must conduct security reviews in specified circumstances, disclose that users are interacting with AI, protect personal data and issue reminders after each two hours of continuous use.
In the United States, Executive Order 14409 gave federal agencies until August 1, 60 days after the order was signed, to develop a classified cyber-capability benchmark and a voluntary framework for covered frontier models. Participating developers could give the government access for up to 30 days before release to trusted partners. The National Security Agency would determine whether a model meets the covered-model threshold after consulting other officials.
The US order does not establish mandatory approval. Its text expressly rejects a federal licensing, preclearance or permitting requirement for releasing AI models. In the European Union, most remaining AI Act provisions apply on August 2, including transparency duties for specified AI interactions and generated content, along with European Commission enforcement powers covering general-purpose AI obligations. The EU model relies on market-access rules, documentation and regulatory supervision, not routine government review of every model before release.
Three Gates Close in Nineteen Days
The Pre-Release Regime Goes Global
Same-day-verified · one instinct, three architectures — and none of them binds the open frontier
Anthropomorphic-interaction measures take effect: five agencies extend the CAC approval regime to companion AI and agents.
EO 14409’s classified benchmark and voluntary 30-day pre-release framework harden. NSA designates covered frontier models.
The AI Act becomes fully applicable — the staged rollout that began February 2025 reaches its final station.
Same instinct, three theories of a gate
STEELMAN: THE GATE-SKEPTIC CASE
Pre-release regimes structurally favor incumbents who can afford the process — and none of the three binds an open-weight release from a lab outside its jurisdiction. The gates go up exactly as the fastest-moving part of the frontier walks around them.
The signal: a model can clear all three gates having been evaluated for three almost non-overlapping things — content control, fundamental rights, national security. Jurisdiction is now an architectural property. If your deployment calendar doesn’t carry July 15, August 1, and August 2, it’s a calendar for a market you’re not in.
AI companion and emotional-support device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Markets, Three Compliance Tests
The compressed calendar gives multinational developers little room to treat compliance as a single global exercise. A companion application offered in China may need security reviews, dependency safeguards and user-intervention features. A provider participating in the US framework may need to arrange confidential government access. An EU-facing service may face disclosure, documentation and supervision duties tied to how the system is marketed and used.
The milestones also show the limits of calling all three systems pre-release gates. China is applying direct service controls, the EU regulates access to its market through a risk-based rulebook, and the US framework remains voluntary by executive order. A model could comply in one jurisdiction without satisfying another because the authorities focus on social and content risks, fundamental rights, product safety or national security.
portable SSD for machine learning datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Rules Converge, Designs Do Not
China has regulated recommendation algorithms, synthetic media and public generative-AI services through earlier measures. The July rules add a dedicated layer for emotionally interactive systems, including protections for minors, controls on sensitive interaction data and action when a service presents a major safety risk.
The EU AI Act entered into force in August 2024. Prohibited practices and AI-literacy duties began applying in February 2025, followed by governance and general-purpose AI obligations in August 2025. EU lawmakers have since approved the Digital Omnibus on AI, moving stand-alone high-risk obligations to December 2, 2027 and product-embedded high-risk obligations to August 2, 2028. That change means August 2, 2026 is a major application date, but not the start of every high-risk requirement.
“Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.”
— Executive Order 14409
smart badge printer for AI workplaces
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
US Benchmark Details Stay Classified
The US benchmark and covered-model threshold are not public, leaving developers without an open account of which capabilities will trigger review. It is also unclear how many companies will participate, what protections will govern model access or whether trusted-partner benefits will provide a strong enough incentive.
Enforcement practice under China’s new measure is still developing, including how authorities will apply its security-review triggers to updated or foreign-operated services. In Europe, companies still await parts of the implementation guidance and technical standards needed for consistent application. The broad claim that no regime can reach open-weight models released abroad also remains disputed: legal exposure depends on where a model or service is offered, its capabilities and the exemptions available.
AI news and regulation guidebook
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
August Deadlines Test Readiness
Attention now shifts to whether US agencies publish any unclassified account of the August 1 framework and which developers agree to participate. One day later, EU authorities and providers begin operating under the August 2 provisions that remain scheduled, including specified AI-transparency duties. Regulators’ first enforcement decisions will show whether this 19-date cluster produces practical alignment or three separate compliance tracks.
Key Questions
Do all three jurisdictions require government approval before an AI model is released?
No. China requires security reviews in specified cases, the EU uses market and conformity obligations, and the US framework is voluntary.
What services are covered by China’s new measure?
It covers public services in China offering sustained emotional interaction through simulated human personality or communication. Ordinary customer-service and work tools are excluded when they lack that feature.
What happens under the US framework?
Participating developers may provide a covered frontier model to the government for up to 30 days before sharing it with trusted partners. The review uses a classified cyber benchmark.
Does the full EU high-risk regime begin August 2, 2026?
No. Most remaining provisions apply then, but major high-risk duties were moved to December 2027 for stand-alone systems and August 2028 for systems embedded in regulated products.
Source: Thorsten Meyer AI
