China's generative AI regulation takes effect

China’s Interim Measures for the Management of Generative AI Services came into force, becoming the country’s first binding regulation aimed specifically at generative AI. The Cyberspace Administration of China and six other agencies had jointly issued the measures the previous month, giving providers roughly five weeks to prepare.

The rules applied to any service generating text, images, audio or video for the public within China. They required that training data be lawfully sourced and respect intellectual property, that generated content be truthful, accurate and free of discriminatory bias, and that AI-generated content be labelled. Providers had to identify illegal content and act on it — suspending generation, removing outputs and reporting incidents to regulators — and to give users a mechanism for complaints. Services judged to carry “public-opinion attributes” or “social-mobilisation capabilities” needed to complete a security assessment with the Cyberspace Administration and file their underlying algorithms before launch.

The final text was notably softer than an April draft that had circulated for comment. Provisions requiring real-name verification of users, banning certain forms of user profiling, and imposing a fixed three-month window to correct non-compliant content were all dropped, as were specific monetary penalties for violations. Language emphasising industry development and innovation, rather than restriction, was strengthened in its place — read by many observers as a deliberate softening once regulators weighed the compliance burden against China’s ambition to compete in generative AI globally.

The measures became a reference point in the wider international conversation about AI governance: an example of a major economy moving from voluntary guidance to a binding, if narrowly scoped, legal instrument for generative AI well before the United States or European Union had equivalent rules in force, even as the substance of the requirements — labelling, filing, content review — stopped well short of a licensing or safety-testing regime for frontier models.