TL;DR
China, the United States and the European Union are activating major AI oversight gates within a 19-day period. The regimes share an interest in examining powerful systems before broad release, but differ sharply on approval, enforcement and the risks being tested.
China, the United States and the European Union are activating separate AI pre-release regimes within 19 days, placing new government checks around advanced systems before or as they reach users. China’s rules for human-like AI took effect July 15, while a US national-security framework is scheduled to harden August 1 and the EU AI Act reaches its main application date August 2.
China’s Interim Measures for AI Anthropomorphic Interaction Services, issued in April by five agencies, extend the country’s existing approval structure to companion systems, agents and other services designed to interact in human-like ways. The agencies include the Cyberspace Administration of China, the national development and technology ministries, public-security authorities and the market regulator.
According to the Thorsten Meyer AI dispatch, covered Chinese services face security reviews before public deployment, algorithm registration and continuing government oversight. The dispatch says incidents must be reported within 24 hours, government information requests answered within 48 hours and ordered algorithm changes carried out. The exact coverage of particular products will depend on how Chinese authorities apply the measures.
In the United States, Executive Order 14409 is described as creating a voluntary 30-day evaluation window for participating frontier-model developers. The framework uses classified benchmarks and offers trusted-partner status as a procurement incentive, according to the dispatch. On August 1, the National Security Agency is expected to designate covered frontier models, but participation remains voluntary rather than a condition for market access.
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 safety compliance software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Gates Test Different Risks
The dates form a shared regulatory moment, but the three systems do not amount to a common global standard. China emphasizes content control and social stability, the EU focuses on fundamental rights and product safety, and the US framework targets national-security risks. A developer operating in all three markets may need separate evidence, testing and release schedules.
The differences make jurisdiction part of product architecture. Model design, documentation, incident reporting and government access may vary by market rather than being handled through one global review. The Thorsten Meyer AI dispatch argues that a model could pass all three systems while being examined for largely different hazards.
The regimes may also favor established companies able to fund legal reviews, evaluations and documentation. That concern remains an interpretation rather than a measured outcome. It is also unclear how effectively any of the systems will reach open-weight models released abroad, which can cross borders without following a conventional product-launch process.
AI development monitoring tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Approval, Conformity and Voluntary Review
China has required security assessments and algorithm filings for public generative-AI services since 2023. Its new rules apply that model to the companion-AI and agent layer, allowing regulators to seek changes before or after deployment. This is the closest of the three regimes to direct pre-release approval.
The EU uses a broader conformity system. The AI Act’s staged rollout began with prohibited practices in February 2025, followed by obligations for general-purpose AI models in August 2025. Its requirements include risk classification, technical records, conformity assessments and post-market monitoring, with added duties for models classified as presenting systemic risk.
The US structure is lighter and tied to government access rather than general market authorization. Developers that opt in provide advance access for classified evaluation, while trusted-partner treatment supplies the incentive. The United Kingdom remains outside this pattern, retaining a sector-led regulatory model without a single formal pre-release gate.
“The shared signal is that some classes of AI system should meet the state before they meet the public.”
— Thorsten Meyer AI dispatch
AI regulation compliance kits
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
EU Timing and Offshore Models
The EU deadline carries a pending legal caveat. A Digital Omnibus package approved by the European Parliament on June 16 by 423 votes to 57, with 174 abstentions, would move certain high-risk-system deadlines. According to the source material, the package still requires Council adoption and publication in the Official Journal, so the proposed changes are not yet legally effective.
Questions also remain about which models the NSA will designate, how the classified US benchmark will be applied and how many developers will join a voluntary process. None of the three approaches clearly binds every foreign open-weight release. Enforcement reach, compliance costs and the treatment of systems assembled from models released in other jurisdictions are still unsettled.
AI model evaluation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
August Deadlines Put Systems to Test
Attention now moves to the August 1 US designation process and the EU’s August 2 application date. Developers serving several markets will need to map which products are covered, align release calendars and prepare separate records for Chinese approval, EU conformity and US evaluation.
EU institutions may still alter some high-risk deadlines if the Digital Omnibus completes the legislative process. In Washington, Congress could later turn parts of the voluntary framework into mandatory requirements, but no such change is confirmed in the supplied material. Early enforcement actions and developer participation will show whether the three gates materially affect launches or leave major gaps around offshore and open-weight models.
Key Questions
What changed on July 15?
China’s anthropomorphic-interaction measures took effect, extending oversight to companion AI, agents and other systems offering human-like interaction. Covered services may face security review, registration and incident-reporting duties.
Does the US framework block models from release?
No. Based on the supplied material, the US program offers a voluntary 30-day government evaluation using classified criteria. Trusted-partner status and procurement access provide incentives, but the process is not general market approval.
Is the EU AI Act deadline already delayed?
No confirmed delay is in force under the supplied material. Parliament approved proposed changes, but Council adoption and Official Journal publication are still required. Until those steps occur, August 2 remains the operative date.
Do the three regimes use the same AI safety test?
No. They examine different risk categories and impose different processes. China centers state approval and content controls, the EU uses risk-based conformity duties, and the US framework emphasizes classified national-security evaluation.
Are open-weight models covered worldwide?
Not necessarily. The source identifies a continuing gap around models released outside a regulator’s jurisdiction. How authorities will treat foreign downloads, local adaptations and services built on those models remains unclear.
Source: Thorsten Meyer AI
