On 25 October 2023, the public consultation closed on the National Information Security Standardisation Technical Committee of China (TC260) draft Basic Requirements for Security of Generative Artificial Intelligence (AI) Services. In particular, the draft specifies that Generative AI service providers are required to implement measures to ensure the legality and quality of the training data and cannot train their algorithms with more than 5% of data considered illegal or harmful. The output of the Generative AI model would have to undergo tests to confirm its compliance with 31 security risk categories. The providers would have to ensure that the generated content doesn’t violate socialist values or the legal rights of others or contain discriminatory elements. The content quality should have a success rate of 90%. Furthermore, the Generative AI model would have to be tested to ensure that inappropriate queries are rejected at a success rate of 95%.
Original source