On April 8, 2022, the Cyberspace Administration of China issued a notice to take the lead in carrying out the special action of "Qinglang • 2022 Algorithm Comprehensive Governance", and by the end of this year, it will focus on inspecting large websites, platforms and products with strong public opinion attributes or social mobilization capabilities. In the notice, "actively carry out algorithmic filing algorithm filing" is regarded as one of the main objectives.
In the era of digital economy, the application of algorithms has greatly improved the efficiency of data processing, but at the same time, there have been negative events that use algorithm models to kill big data, discriminate against algorithms, induce addiction, "trap takeaway workers in the system", affect public opinion, and other unreasonable use of algorithms that are detrimental to market order and are not conducive to maintaining socialist core values.
In order to effectively deal with many problems caused by the illegal or unreasonable use of algorithms, on March 1, 2022, the Provisions on the Administration of Recommendation of Internet Information Service Algorithms (hereinafter referred to as the "Provisions on Algorithm Recommendation") came into effect, proposing a new scheme for "algorithm filing".
Starting from the introduction of the world's main algorithm transparency governance model, this article will show the focus and basic requirements of algorithm governance in different countries and regions, and focus on the specific requirements of the mainland algorithm filing system, and at the same time, from the perspective of comparative law, a brief analysis of the new direction of algorithm governance in the future mainland will be made.
1. Overview of the main governance models of algorithmic transparency
1. Eu: Emphasis on algorithmic transparency of large-scale network platforms and AI systems
At the beginning of 2022, the European Union passed the Digital Services Act, which builds on the General Data Protection Regulation and puts forward new requirements for algorithmic transparency of web platforms, especially very large web platforms.
On the one hand, the network platform should state the main parameters of the recommended algorithm system in the terms of service, and the parameter information should be made in a clear and clear language and publicly provided in an accessible form; on the other hand, when the European Commission requests access to the algorithm and database of the enterprise ex officio, the enterprise should provide the complete algorithm and database and make a clear interpretation of it.
In response to the issue of algorithm transparency of artificial intelligence systems, the European Commission's proposal on the formulation of uniform rules for artificial intelligence (the "Artificial Intelligence Act") issued by the European Commission in April 2021 divides artificial intelligence systems in different application scenarios into four levels: unacceptable risk, high risk, limited risk and low risk, and in order to ensure the transparency of algorithms for high-risk artificial intelligence applications, the institutional arrangement of the registration and filing of artificial intelligence algorithms (Register) has been carried out.
In accordance with Articles 51 and 60 of the Rules, the EU establishes an independent database of high-risk AI systems in which providers or authorized representatives shall register and file high-risk AI systems before they are placed on the market or put into use, and stipulate that the information contained in the EU databases should be accessible to the public.
2. United States: Emphasize different algorithmic transparency requirements for businesses and government agencies
For enterprises, the main focus is on algorithmic transparency in the field of commercial consumption. In its document" "Aiming for Trust, Fairness, and Equality in Companies' Use of AI," the U.S. Federal Trade Commission (FTC) emphasized that companies should use transparent frameworks and independent standards when applying algorithms, and that their algorithms should be regularly tested before and after using AI to ensure that it does not discriminate on the basis of race, gender, etc.
In terms of government application algorithms, in order to promote public acceptance and trust in the government's use of artificial intelligence and algorithm technology in decision-making, the United States has made detailed regulations on the application of algorithms by administrative agencies and defense and intelligence departments.
For example, the U.S. Executive Council (AC) stipulates in its Guidelines for the Use of AI for Government Agencies that the objectives of algorithmic processes and the rationale behind them should be publicly determined to facilitate internal or external review of agencies' AI-based decision-making.
The U.S. Department of Defense (DoD) requires in the Intelligence Agency's Framework for the Ethics of Artificial Intelligence to ensure algorithm transparency and interpretability, using understandable and interpretable methods to the extent feasible, in accordance with the intelligence transparency principles of the intelligence services, so that users, supervisors, and the public understand, as appropriate, how and why AI produces its output.
3. United Kingdom, Singapore: Implement the principle of algorithmic transparency through transparency reports and algorithmic audits
On 21 May 2021, the United Kingdom published the Draft Online Safety Act, which introduces uniform algorithmic transparency standards for the use of AI algorithm systems by government public agencies, and Article 49 of the draft requires providers of platform services to prepare annual transparency reports for each service they provide, and the information provided in the reports must be complete and accurate.
On 21 January 2020, Singapore's Personal Data Protection Commission (PDPC) released the second edition of the Model AI Governance Framework, establishing guiding principles for "transparency" and "people-centricity" of AI algorithms. The framework proposes audit requirements for AI algorithms in the appendix section, stipulating that algorithm audits should be performed if there is a need to actually operate the algorithms in the Discover model.
The framework requires that algorithm audits must be conducted at the request of a regulatory body with jurisdiction over the organization (as part of a forensic investigation) or performed by an AI technology provider to ensure the transparency of the algorithm.
4. China: Ensure algorithm transparency in automated decision-making and algorithm recommendation services
Returning to the mainland's threshold, the mainland emphasizes the transparency of algorithms applied by automated decision-making technology for personal information processors. Legislatively, the E-commerce Law and the Personal Information Protection Law both provide for automated decision-making.
For example, Article 24 of the Personal Information Protection Law requires personal information processors to ensure transparency in decision-making, ensure fairness and impartiality, and achieve equal transactions. This article also stipulates the "right to know" and the "right to refuse" that the information subject has when the personal information processor makes a decision that has a significant impact on the rights and interests of the individual.
In addition, in September 2021, the State Internet Information Office and nine other ministries and commissions jointly issued the "Guiding Opinions on Strengthening the Comprehensive Governance of Internet Information Service Algorithms" (hereinafter referred to as the "Guiding Opinions"), requiring the use of about three years to gradually establish a comprehensive governance pattern of algorithm security with sound governance mechanisms, perfect regulatory systems, and standardized algorithm ecology.
The comprehensive regulatory system mainly includes five aspects, namely algorithm security risk detection, algorithm security assessment, scientific and technological ethics review, algorithm filing management, and handling of algorithm-related violations of laws and regulations.
At the end of 2021, the Cyberspace Administration of China issued the Provisions on Algorithm Recommendation, the first departmental regulation in mainland China to regulate algorithm recommendation services, which aims to regulate the provision of information to users through algorithm technology, protect users' rights to know, choose, refuse, and remedy, and strengthen the protection of the rights and interests of minors, the elderly, workers, and consumers in the Internet age.
Among them, the algorithm filing system, as a new solution for the governance of algorithm transparency in mainland China, is also an important part of the comprehensive governance pattern of algorithms, which is particularly worthy of attention.
Second, the mainland algorithm filing needs to pay attention to the key issues
The Provisions on Recommendation of Algorithms first clearly stipulate one of the globally recognized regulatory principles in the field of algorithm governance: algorithm transparency, while the algorithm filing system implements the principle of algorithm transparency. The Provisions on Recommendation of Algorithms clearly stipulate the obligated subjects of the algorithm filing system, the algorithm technology that should be filed, the content of the algorithm filing, the publicity of the algorithm filing, and the dynamic filing of the algorithm, and innovatively stipulate the administrative law responsibility for the filing of violations.
1. Which entities need to fulfill the obligation of algorithm filing?
Article 24 of the Provisions on Algorithm Recommendation stipulates that the subject that needs to be filed is "an algorithm recommendation service provider with public opinion attributes or social mobilization capabilities". First of all, the subject of the filing obligation is limited to the algorithm service provider, excluding the algorithm developer, designer and algorithm end user (user). Second, the subject of the filing obligation must have the attribute of public opinion or the ability to mobilize society.
According to the Provisions on Security Assessment of Internet Information Services with Public Opinion Attributes or Social Mobilization Capabilities issued by the Cyberspace Administration of China in November 2018, Internet information services with public opinion attributes or social mobilization capabilities include the following circumstances:
Open information services such as forums, blogs, microblogs, chat rooms, communication groups, public accounts, short videos, webcasts, information sharing, mini programs, etc., or attach corresponding functions;
Establish other Internet information services that provide channels for the expression of public opinion or have the ability to mobilize the public to engage in specific activities.
Specifically, applications such as WeChat, Weibo, Douyin, and Douban belong to algorithm recommendation service providers with public opinion attributes or social mobilization capabilities.
2. Which algorithm recommendation techniques need to be filed?
Specifically, the main "application algorithm recommendation techniques" specified in the Provisions on Algorithm Recommendation include information services such as generating synthetic classes, personalized push classes, sorting selection classes, retrieval filtering classes, and scheduling decision-making classes.
The generation synthesis class mainly regulates content aggregation, content automation generation and content synthesis services, including the use of deep learning, virtual reality and other technologies to produce text, images, audio, video, virtual scenes and other information, such as speech synthesis, AI face change, video automatic editing, metaverse, etc.
Personalized push mainly regulates personalized push services for content, goods or services based on user portraits, such as short videos, news, advertisements, gift packages and other recommendation services.
The sorting selection category mainly regulates the layout page management of content service providers, such as home screen, hot search, featured, list, pop-up window, etc.
The search filtering class mainly regulates services or behaviors such as search engines, content retrieval, and content intervention, such as search engine bad information filtering and sensitive word recognition.
Scheduling decision-making mainly regulates work scheduling services, such as takeaway, online ride-hailing platforms, etc.
3. What is included in the algorithm ICP Filing?
The "Provisions on Recommendation of Algorithms" stipulate that the entity subjects subjects of filing obligations shall, within 10 working days from the date of providing services, fill in information such as the name of the service provider, the form of service, the field of application, the type of algorithm, the algorithm self-assessment report, and the content to be publicized through the Internet information service algorithm filing system, and perform the filing formalities.
According to the instructions of the Internet information service algorithm filing system, the information filled in by enterprises mainly includes three parts: algorithm subject information, filing algorithm information, and product and function information (see figure below). Among them, it should be noted that in the product and function information filling, the enterprise must complete three steps, one is to add product information, the second is to add product function access path, and the third is to add functional information under the path.
After the filing is passed, the enterprise can also change the filing information or cancel the filing according to the business changes.
4. Do I need to file a record to change or terminate the algorithm service?
Where there is a change in the filing information of the algorithm recommendation service provider, the change formalities shall be completed within 10 working days from the date of the change. Enterprises can modify the subject information, algorithm information, product and function information through the filing system, but after the filing information is modified, the Internet information department will review it again.
Where algorithm recommendation service providers terminate services, they shall go through the cancellation filing formalities within 20 working days of terminating the services, and make appropriate arrangements.
5. How to fulfill the obligation of algorithm service interpretation?
On the one hand, the Provisions on Algorithm Recommendation require that algorithm recommendation service providers should inform users of their provision of algorithm recommendation services in a conspicuous manner, and publicize the basic principles, purpose intentions, and main operating mechanisms of algorithm recommendation services in an appropriate manner. On the other hand, the algorithm recommendation service provider that completes the filing shall indicate its filing number and provide publicity information in a conspicuous position on the websites and applications it provides external services.
Regarding the direct interpretation obligations of algorithm recommendation service providers to users, the main relevant platforms have responded. The content of the algorithm recommendation service interpretation is basically divided into two aspects:
First, personalized content recommendation;
The second is personalized advertising recommendation.
And most of the platform's public content is located in the "settings" interface and the interface related to personal information protection.
The public content of Douyin is more detailed. For personalized content recommendations, Article 1.5 of the "Douyin" Privacy Policy explains the functions of personalized recommendations, the scope of personalized recommendation services, the information collected for predicting preferences and machine calculations, and the feedback and optimization of recommendation models.
For personalized ad recommendations, programmatic and personalized ads are described separately in the Settings - "Understand and manage ad push" section, and users can turn off these two different ads separately.
WeChat's personalized content recommendation service explains in Article 1.34 of the "Information We Collect and Use" section of the WeChat Privacy Protection Guidelines that WeChat will use the collected information to "conduct indirect portraits of people based on feature tags and provide more accurate and personalized services and content".
At the same time, there is a "personalized advertising management" in "Settings" - "Privacy" - "Personal Information and Permissions", and there is a section of "Understand the working principle of WeChat personalized advertising" to introduce the process of advertising, and users can turn off personalized advertising.
Douban has a special "personalized recommendation and advertising" content in the "settings" - "personal information protection" list, which introduces in more detail the basic mechanism and principle of algorithm recommendation of personalized content and algorithm recommendation of personalized advertising. At the same time, users can choose to turn off personalized content recommendations and personalized ad recommendations.
Weibo is more brief in terms of public content. Weibo briefly explains the personalized advertising recommendation technology in "Settings" - "Privacy Settings" - "Personal Information and Permissions", does not explicitly mention the algorithm, does not explain the personalized content recommendation algorithm, and only allows users to turn off personalized advertising recommendation and personalized content recommendation.
In addition, Weibo mentions in the Privacy Policy document that the user's "device information, IP address, location information, personal page information, and behavior information" will be used to personalize content recommendations.
In summary, it can be seen that since it is still in the early stage of the implementation of the "Algorithm Recommendation Provisions", there is no unified implementation standard, and the degree of detail of the content of the algorithm publicity is mainly grasped by the algorithm recommendation service provider. The author will also continue to pay attention to the standards and regulatory standards for algorithm recommendation service providers to perform the obligation of algorithm filing and publicity.
6. What are the consequences of filing a violation?
First, the legal liability that the algorithm should be filed for the record but not filed. When an enterprise fails to perform its filing obligations, the network information department and the relevant departments such as telecommunications, public security, and market supervision shall, in accordance with their duties, give warnings, circulate criticisms, and order corrections within a time limit; if they refuse to make corrections or if the circumstances are serious, they shall be ordered to suspend information updates and shall be fined between 10,000 and 100,000 yuan.
Second, the legal liability for algorithm filing defects. Where algorithm recommendation service providers with public opinion attributes or social mobilization capabilities obtain filings through improper means such as concealing relevant circumstances or providing false materials, the state and provincial, autonomous region, or directly governed municipality internet information departments are to revoke the filings, give warnings, and circulate criticisms; where the circumstances are serious, order the information to be suspended, and impose a fine of between 10,000 and 100,000 yuan.
Third, the legal liability after the algorithm is filed. Based on the legal nature of the algorithm filing, even if the algorithm activities have been legally filed, they should still bear legal liability for the damage when they cause damage. The filing of the materials submitted by the obligor by the regulatory authority only means that the obligor's filing behavior is valid and does not have any impact on the algorithm activity itself. If the recorder uses the recorded algorithm to harm the rights and interests of users or the public interest, it still has to bear the corresponding legal consequences, and the procedural factual act of filing will not exempt it from legal liability.
3. The prospect of improving the algorithm filing system
As mentioned above, there are still few countries in the world that have implemented algorithmic filing related systems, and only the European Union and the mainland are the major economies that adopt the method of prior filing. In fact, the EU Artificial Intelligence Act only specifically stipulates the filing obligation of high-risk artificial intelligence algorithms, but there are no specific provisions on the connotation, procedure and significance of filing, such as whether it is necessary to consider its own technical nature, application scenarios, sensitivity and volume of processing data when determining which AI algorithms need to perform filing procedures.
On the basis of the Artificial Intelligence Act, the Mainland's "Provisions on Algorithm Recommendation" improve and refine the specific rules and legal consequences of algorithm filing, and make them linked with algorithm security assessment and algorithm inspection, forming a multi-dimensional algorithm supervision framework. [1]
However, the promulgation and effectiveness of the Provisions on Recommendation of Algorithms is still only the starting point of algorithm governance in mainland China, and it still needs to be continuously developed and improved.
First of all, for the obligatory subjects of algorithm filing, it should be further expanded in the future. At present, the "Algorithm Recommendation Provisions" are completely aimed at algorithm service providers, that is, market entities, but in the future, in public government governance, it is also necessary to apply instrumental and artificial intelligence algorithms, and the use of algorithms by government and other management agencies also needs to be regulated.
Therefore, we should learn from the us regulation of the difference between the private industry field and the public government field algorithm, as well as the United Kingdom's requirements for the transparency of the government's use of artificial intelligence algorithms, and implement the algorithm filing system for the government to use the algorithm in the future to strengthen the transparency and public acceptance of the algorithm.
Secondly, as far as the scope of algorithm filing is concerned, a classification and grading method adapted to the development of algorithm technology should be adopted. The EU sets the obligation to file algorithms in the field of high-risk artificial intelligence, and the "algorithms" required by the mainland for filing are not limited to deep learning algorithms, but extend to five types of algorithms, such as generative synthesis, personalized push, sorting selection, retrieval and filtering, and scheduling and decision-making.
However, as algorithms are updated and iterated, classification ratings for specific algorithm types will diminish the effectiveness of algorithm transparency governance. In the future, the mainland may be able to learn from the practice of the Ai-Fiat Act to promote its shift to the classification of "instrumental algorithms" and "AI algorithms" and a broader risk classification. [2]
Third, looking at the whole process of algorithm supervision, relevant rules should be refined to promote enterprises to perform filing obligations. In the whole process of algorithm supervision, the rules for the initial filing of algorithms and dynamic filings should be refined to balance algorithm security and algorithm development. At present, the Provisions on Algorithm Recommendation only roughly stipulate the process of changing and terminating the filing of algorithm services, but in practice, when and under what circumstances it is necessary to change it remains to be clarified.
In addition, the Provisions on Algorithm Recommendation only impose algorithm filing obligations and responsibilities on enterprises, and can explore incentive measures for enterprises to adopt algorithm filings, promote enterprises to actively file and truthfully file, and ensure the benign development of the algorithm filing system.
attach:
1. Permission, Liu Chang: "On the Algorithm Filing System", in Artificial Intelligence, No. 1, 2022.
2. Ibid.