Full steam ahead: China publishes interim generative AI measures

New week, new measures: such is the current pace of regulatory change regarding AI, each week can bring major new developments in the regulatory landscape. On 13 July 2023, the Cyberspace Administration of China (CAC) published a set of ‘Interim Measures for the Administration of Generative Artificial Intelligence Services’ (Interim Measures) that build upon the draft measures consulted upon by the CAC a mere few months ago (and which we covered in this recent post).

Whilst the revised generative AI measures are expressly titled “interim measures”, suggesting that they may be updated further, the Interim Measures are nevertheless slated to enter into legal force on 15 August 2023. Businesses active in generative AI (GAI) in China, and which may be affected by the Interim Measures, should start considering the steps required to comply with their provisions.

In this blog post, I highlight the key areas of movement in the revised version of the generative AI regulation and offer some reflections on what this may signal about the likely direction of travel for Chinese AI regulation.

What’s new in the Interim Measures?

The key takeaway is that the Interim Measures reflect a degree of rebalancing between security, regulation and innovation related objectives, and are—to a certain extent—more flexible in their approach to achieving these objectives than the draft measures published in April this year.

The Interim Measures state explicitly that the Chinese government “adheres to the principles of attaching equal importance to development and security, promoting innovation and lawful governance [and] employing effective measures to encourage the innovative development of generative AI”. A new section has been added to the Interim Measures to reaffirm the government’s support for the innovative development and application of GAI technologies.

Notable changes include:

• Provision of services to the public: the Interim Measures clarify that their provisions do not apply to organisations that research, develop and apply GAI technologies but which do not provide GAI services to the public within China. This provides a pro-innovation carve-out for GAI R&D, as well as internal applications of GAI technologies, narrowing the scope of the Interim Measures.
  
  
• Effective measures: rather than the prescriptive, outcome-based requirements imposed by the draft measures (to ensure, for example, the truth and accuracy of both input data and output content), the Interim Measures shift to a process-based standard that requires GAI service providers to take “effective measures” to (amongst other things) “boost the transparency of GAI services and the accuracy and reliability of the content generated” and “enhance the authenticity, accuracy, objectivity and diversity of training data”. Further, strict time limits on the period for rectifying or retraining non-compliant models have been removed.
  
  
• User service agreements: the Interim Measures signal that providers of GAI services may allocate a degree of responsibility for compliance with the Measures to users through service agreements setting out the rights and obligations of the parties.
  
  
• Enforcement: under the Interim Measures, new provisions specify that the CAC is empowered to notify relevant authorities to take “technical measures and other necessary measures” to deal with non-compliance by extraterritorial entities providing GAI services to the public within China. In addition, references to capped fines have been deleted—though this does not affect the ability of regulatory authorities to impose financial penalties under other related legislation.

What should we make of the new Interim Measures?

In my previous post on China’s draft generative AI regulation, I suggested that jurisdictional differences in approaches to AI regulation should not be overstated. Regulators in key jurisdictions such as the EU, UK and China appear to be ultimately oriented towards similar substantive objectives (regarding, for example, input data quality; the normative risks posed by AI-generated content; and the importance of advancing safety, transparency and accountability standards).

In their current form, the revised Interim Measures suggest that a degree of regulatory convergence is already occurring. Not only do the Interim Measures move away from prescriptive technical requirements in favour of more open-ended approaches to achieving regulatory outcomes, references to implementing “categorised and graded regulation for GAI services” suggest, perhaps, that Chinese regulators may be open to developing a risk-based approach AI regulation along the lines of that adopted under the EU’s forthcoming AI Act. 

Risk-based triaging is already visible elsewhere in the Interim Measures: any providers of GAI services the nature of which is related to “public opinion or capable of social mobilisation”, for example, are required to conduct security assessments, a tightening of the equivalent requirement set out in the draft measures.

The Interim Measures also signal China’s outwards-looking stance on AI regulation, explicitly encouraging “international exchange and cooperation” and Chinese participation in “formulating international rules relating to GAI”. As the EU and US race to formulate a global AI code of conduct, China may be eyeing the ways it can leverage first mover advantage to set a floor for GAI regulation and frame the global regulatory discussion going forwards.

More commentary and resources on deploying AI in your business and on regulating AI more generally, can be found on our Regulating AI hub.