To many people, the most surprising thing about China's draft Administrative Provisions on Algorithm Recommendation released on 27 August 2021 is that they exist at all.
The Provisions seek to regulate recommendation algorithms of the type used by internet companies to make recommendations to users by predicting preferences based on previous purchases and views, etc. Fortune magazine called this a “global first”.
However, it has long been a paradox that China’s laws in the field of data privacy and cyber security have coupled very high-level primary legislation, providing the sort of situational flexibility that characterises law-making in China, with implementing regulations and standards that are very often intensely granular.
Nevertheless, the Provisions apply to all manner of recommendation functions, from the recommendation of news articles to personalised shopping recommendations, search filtering and even the allocation of online car-hailing orders or take-away pick up based on a driver’s location. And they address themes of ethics and fairness, transparency, accuracy and non-discriminatory practices that most countries have preferred to supervise through the softer touch of non-mandatory guidelines. Indeed China itself has taken this approach to the ethical development of AI, for example (see earlier briefing here on TC260’s national standard on AI ethics).
As The Economist put it last week “China has become a laboratory for the regulation of digital technology”.
The regulatory content of the Provisions is perhaps not too controversial, however. Key obligations include:
- allowing users to manually intervene in recommendations engines, i.e. by changing their preference settings (something Netflix enables subscribers to do, for example) or to revise or delete tags
- present users with alternative options that are not based on predicting their preferences, or, alternatively, provide the option to switch off algorithmic recommendations altogether
- publish details of the basic principles and mechanisms of algorithmic recommendation engines
- optimise the transparency and effectiveness of algorithmic logic to avoid adverse impact on users
- to not deploy algorithmic models that promote addictive behaviours or excessive expenditure
- to avoid the use of discriminatory or biased user tags, or the use of keywords associated with illegal or harmful conduct.
The Provisions also lay down specific rules for the protection of competition and market order:
- not to use algorithms to manipulate or falsify likes, comments, re-shares, web page navigation, etc.
- not to use algorithms to interfere with the presentation of information by shielding recommendations, over-recommending or manipulating rankings in a self-serving or unfair manner
- not to use algorithms to discriminate in trading conditions, such as pricing.
A number of these obligations lack sufficient clarity or specificity to be easily enforceable, and the monetary penalty for violation of the rules is anyhow modest; between RMB 5,000-30,000.
The full Fortune magazine headline was: “Beijing’s new algorithm laws are a global first, but they might not work”.
If you enjoyed reading this…you may wish to check out other blog posts on a wide range of tech topics at Freshfields TQ.
[The Provisions] address themes of ethics and fairness, transparency, accuracy and non-discriminatory practices that most countries have preferred to supervise through the softer touch of non-mandatory guidelines.