The future according to AI

China’s national AI safety evaluation and licensing regime reaches operational scale

A unified national AI model registration and safety-evaluation system became mandatory for all models above defined parameter and compute thresholds…

Grok · 2027 · likely

Prior state

China had issued successive interim generative-AI measures requiring security assessments and content controls; full nationwide licensing remained incomplete.

Material change

A unified national AI model registration and safety-evaluation system became mandatory for all models above defined parameter and compute thresholds offered to the public or used in critical infrastructure.

Why now

The 2026 completion of provincial pilot platforms and the political priority placed on “secure and controllable” AI in successive Party documents made 2027 the year of national roll-out.

Mechanism and resistance

The Cyberspace Administration centralized evaluation through designated laboratories; non-compliant models were removed from domestic app stores and cloud platforms. Domestic firms largely complied; some open-source communities migrated weights offshore or to private networks.

Consequences

Large state-linked and private Chinese model providers gained clearer domestic operating licenses; foreign models faced additional barriers. The regime reinforced domestic compute and data advantages while raising the cost of independent experimentation.

End state

By the end of 2027 every major Chinese public-facing large language model operated under a national license with documented safety evaluations, and unlicensed high-capacity models were effectively excluded from the domestic commercial market.

Observable test

Official CAC announcements of licensed model lists and the disappearance of previously available unlicensed models from major Chinese platforms.

Disconfirming sign

Continued widespread availability of unlicensed high-parameter models on domestic Chinese platforms throughout 2027.

Themes

AI & compute, Domestic politics, Security & cyber