TL;DR
China’s rules for human-like AI took effect on July 15, while a US voluntary evaluation framework and the EU AI Act reach key dates on August 1 and August 2. The three systems examine different risks, carry different legal force and leave gaps around open-weight models released beyond their jurisdictions.
China’s rules for human-like AI services took effect July 15, opening a 19-day period in which the United States and European Union are also scheduled to activate major pre-release controls. The closely grouped dates will make government review, documentation and market-access planning a more immediate part of deploying advanced AI across three of the world’s largest regulatory jurisdictions.
China’s Interim Measures for AI Anthropomorphic Interaction Services, issued in April by five government bodies, extend the country’s existing approval system to services designed to interact in human-like ways, including AI companions and agents. The agencies named in the Thorsten Meyer AI dispatch include the Cyberspace Administration of China, the national development and industry ministries, public security authorities and the market regulator.
On August 1, the US framework described in the dispatch is scheduled to reach its operational milestone under Executive Order 14409. Participating frontier-model developers would provide the government with 30 days of pre-release access for evaluation against classified benchmarks. Participation is voluntary, with trusted-partner status and potential procurement advantages serving as incentives.
On August 2, the EU AI Act is scheduled to become fully applicable after a staged rollout that included prohibited practices in February 2025 and obligations for general-purpose AI models in August 2025. The law relies on risk classification, conformity checks, technical records and post-market monitoring, with added evaluation and incident-reporting duties for general-purpose models classified as presenting systemic risk.
Three Models of State Control
The dates reflect a shared policy judgment that some advanced systems should face government scrutiny before public release. They do not amount to a common global standard. Each system applies a different test and gives regulators a different role.
China’s gate centers on approval, content controls and social stability, including security reviews and the ability to demand algorithm changes. The EU system focuses on fundamental rights and product safety through legally required conformity processes. The US model emphasizes national security but relies on voluntary participation rather than formal market authorization.
For developers operating across borders, that fragmentation means one model could pass all three systems after undergoing reviews that overlap only partly. Release schedules may need separate time for Chinese regulatory clearance, US government access and EU compliance work. Smaller developers may face heavier relative costs because they have fewer legal, security and documentation resources.
AI compliance and regulation guidebook
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Three Systems Differ
China has required security reviews for public generative-AI services since 2023. According to the dispatch, its registration process allows the Cyberspace Administration to seek design changes before issuing a filing number. Providers also face continuing duties, including a 24-hour incident-reporting period, responses to government information requests within 48 hours and implementation of ordered algorithm adjustments.
The EU’s approach is broader across product categories but does not require an individual government approval for every AI release. Obligations depend on how a system is classified and used. The US framework, by comparison, operates as an evaluation window rather than a legal approval gate. Its central benchmarks are classified, leaving outside researchers and much of the public unable to examine the criteria.
The United Kingdom remains outside this pattern. Its principles-based, sector-led model gives existing regulators responsibility for AI within their fields and does not currently impose an equivalent formal pre-release gate.
“Every serious jurisdiction has concluded that some class of AI system should meet the state before it meets the public.”
— Thorsten Meyer AI dispatch
As an affiliate, we earn on qualifying purchases.
Deadlines and Coverage Still Unsettled
The EU timetable could still change for parts of the law. The Digital Omnibus package, approved by the European Parliament on June 16 by 423 votes to 57, with 174 abstentions, would move some deadlines for high-risk systems. According to the supplied dispatch, it has not completed Council adoption or publication in the Official Journal, so the August 2 date remains legally operative for now.
It is also unclear how many US developers will enter the voluntary 30-day evaluation process, which models the National Security Agency will designate as covered, and how trusted-partner status will affect procurement decisions. The classified benchmarks limit public examination of what the government will test.
None of the three systems clearly closes every route for an open-weight model released outside the regulator’s jurisdiction. Enforcement against foreign developers, treatment of downstream modifications and the practical cost for smaller laboratories remain unresolved.
As an affiliate, we earn on qualifying purchases.
Next Tests Arrive in August
Chinese authorities will begin applying the anthropomorphic-interaction rules to covered services, making early filing decisions and enforcement actions the first indicators of their reach. Attention then shifts to the US framework on August 1 and the EU AI Act milestone on August 2.
Developers will be watching for US participation announcements, NSA model designations, final movement on the EU Digital Omnibus and national enforcement guidance across Europe. Those developments will show whether the three gates materially change releases or mainly add parallel compliance tracks to global deployment plans.
As an affiliate, we earn on qualifying purchases.
Key Questions
What changed in China on July 15?
China’s interim rules for anthropomorphic AI interaction services took effect, extending the country’s approval and oversight model to covered AI companions, agents and human-like services.
Does the United States require approval before releasing an AI model?
No. The framework described in the source is voluntary. Participating developers provide 30 days of pre-release access for government evaluation, with trusted-partner status acting as an incentive.
Does the EU AI Act require every model to pass a government review?
No. The EU system assigns duties according to risk category, model capability and intended use. Some providers must complete conformity work, maintain technical records and monitor systems after release.
Has the EU already delayed its August deadlines?
Not under the status reported in the supplied material. The Digital Omnibus would change some high-risk deadlines, but it still requires the remaining legislative steps before those changes carry legal force.
Do these regimes cover open-weight models released abroad?
Coverage remains incomplete and depends on where the developer operates, where the model is offered and how downstream services use it. Models released beyond a regulator’s reach may expose enforcement gaps.
Source: Thorsten Meyer AI
