Author | Fish three falcons
Edit | Wei Shijie
On March 1, China's first regulation specifically for algorithm recommendation, the Provisions on the Administration of Recommendation of Internet Information Service Algorithms, was officially implemented.
The Provisions require that algorithm recommendation service providers need to provide users with the right to know about algorithms, the right to choose algorithms (the convenient option to close algorithm recommendation services), and make specific specifications for algorithm recommendation service providers that provide services to minors, the elderly, workers, consumers and other entities.
Up to now, WeChat, Douyin, Today's Toutiao, Taobao and other apps have made initial improvements, and the algorithm recommended close key has been launched.
As one of the main productive forces of digital life in the 21st century, algorithms have gradually spread to all fields of society by virtue of the explosion of data volume and the improvement of computing power, affecting the distribution of information, the allocation of social resources such as commodities, and the implementation of automated decision-making.
From joyfully accepting and feeling its convenience, to being troubled by it and causing hidden worries, how do algorithms "dominate" our lives? Can users really abandon the algorithm? How do the new rules restrict the improper rights of "algorithms"? What kind of development has the relevant laws and regulations undergone? Will the content of the Provisions face practical difficulties?
Recently, we have had conversations with digital humanities scholars, Internet governance researchers, and former algorithm engineers at big factories.
Scrutiny: How do algorithms exercise "power"?
Li Dabai: Translator of "The Power of Algorithms" and scholar of digital humanities
The coordination, cooperation and control of the social order are inseparable from information exchange, and through the processing and application of a large amount of data, algorithms have become an effective carrier of information exchange in today's society, so it is not an exaggeration to say that "algorithms represent a power".
The author of The Power of Algorithms, British scholar and barrister Jamie Saskander has been concerned for years about the political, economic and social impact of digital technologies. In the book, he depicts the total "invasion" of private and public life by digital technology, of which "force", "censorship" and "perceived control" are the three ways in which algorithms exercise "power".
The Power of Algorithms by Jamie Saskander
"Force" means that a program that has been written cannot be selected or aborted during execution, and people can only obey the results automatically given by the algorithm.
In 2009, Oba President Ma presented 25 classic American films as gifts to visiting British Prime Minister Gordon Brown. However, brown returned to London and found that the films could not be played on his British DVD players; the reason was that DVD manufacturers and distributors had coded the ban on the DVD in order to protect commercial interests and enforce copyright laws, and even though Brown was already the most powerful person in Britain, he could not do anything in the face of the algorithm's prohibition instructions.
"Censorship" is mainly through the collection of various data to monitor, predict and regulate people's actions.
For example, the recent network broke a number of "algorithmic systems to monitor employee behavior", "algorithms predict the probability of employee departure" news, although the relevant Internet companies have come forward to deny this statement, but this incident itself shows that algorithms have become more and more commonly used to measure and evaluate people's work, and even predict when employees may leave, the result is to reproduce Foucaeus's "panoramic prison" scene: employees have to pay attention to the Internet behavior during working hours.
There is more discussion about the "perception control" of people by algorithms, and the core concepts are often referred to as "information cocoons" and "filter bubbles" over the years, which refers to the ability of algorithms to filter information and change the ordering of information, which in turn affects people's feelings and cognition of the outside world.
The information people receive is always subject to some sort of "filtering." Before the era of algorithms, the responsibility of filtering information was mainly done by mass media such as print, radio, and television, and today, recommendation algorithms have assumed the responsibility of filtering to a considerable extent, which is represented by today's headlines that were launched in August 2012 and focused on personalized content recommendation. At first, people didn't realize the impact it could have, but interest-based algorithmic recommendations almost fundamentally changed the logic of content production. In other fields, the scenarios of algorithm application have gradually enriched, and e-commerce, takeaway platforms, and ride-hailing software have also adopted more complex algorithm systems and done more dimensions of service matching, completely building an algorithm-led platform society.
At the beginning, because the algorithm can more accurately and widely control the information people transmit and receive, it was enthusiastically sought after by users when it was first applied, and it was invincible; however, with the emergence of negative problems such as "big data killing" and "takeaway riders trapped in the system", algorithms can "serve" people, but can also "use" people's concepts have also begun to penetrate the public.
There are also many related cases abroad. In the 2016 U.S. election, some scholars found that robot trolls "produced" about 19% of the relevant tweets to solicit votes for the candidates; in the same year, when the Brexit referendum was held, about a third of the traffic on Twitter came from bots, and they were almost all on the side of Brexit. It's hard to argue that the information these algorithms write doesn't interfere with people's public decisions.
Algorithms increasingly determine who is seen and who is hidden; who enters and who is out, which content spreads like a virus, and which content is destined to go unnoticed.
The Provisions on the Recommendation and Administration of Internet Information Service Algorithms, which came into effect in March this year, pointed out that users can choose to turn off algorithm push, which is undoubtedly justified, because it provides users with the necessary "exit mechanism" for the use of algorithms, so that users have the right to choose whether to accept the algorithm's shaping of self-perception channels.
The question is what will users see after shutting down?
I tried to tick the "Turn off personalized recommendations" button on a content platform, and then the content presented on the platform was almost disorderly, so I had no choice but to reopen "personalized recommendations".
If before algorithm-led content, we saw high-quality content selected by editors, then in the case of "algorithm" checks, the information flow without personalized recommendations is likely to be neither nutritious nor meet the reading needs of users. If in order to achieve the independent choice of algorithm recommendations, it is necessary to lose the quality of reading or use experience, which is obviously what most people do not want to see. If the platform grasps this psychology of the user and "treats" the user who closes the recommendation algorithm, then the user can only give up the right to choose in his hand.
In addition to the above problems, due to the increasingly wide application of algorithms in social scenarios, it has also raised questions about whether algorithms are fair and just.
In 2014, Amazon developed a resume screening system to identify and sort the resumes of the past 10 years, and the result was that the system had a clear preference for male candidates; in 2015, Google updated its photos app to add an automatic tagging function, but was discovered by a black programmer in New York that his photos were labeled "gorillas".
The algorithm itself may be "innocent", but when it crawls, learns from incomplete, or biased data, discrimination and injustice are naturally embedded in the algorithm. It is difficult to require every algorithm engineer to be a "philosopher king" who is familiar with moral philosophy and social policy, and to achieve a relatively fair algorithm requires the joint participation of various disciplines such as politics, law, humanities, and sociology.
Governance: The history of the "normative" web
Fang Shishi: Associate Researcher of the Institute of Journalism, Shanghai Academy of Social Sciences, Director of the Internet Governance Center
Digitization, digitization and computability are the barriers to entry into the algorithmic society. Through a variety of quantification and tracking technologies, we are experiencing a "data boom". How to use this data effectively, rationally and innovatively is not only a technical problem, but also a social problem.
In this regard, the EU got off to a relatively early start, with their 2016 General Data Protection Regulation (GDPR) calling it "the strictest privacy and data protection law in history."
In fact, as early as 1995, the European Union issued a data protection directive, but at that time there were very few people using the Internet, and the collection and processing of personal data was limited to user names, addresses and financial information. With the spread of the mobile Internet, the content of this directive has been unable to help the EU cope with the constant emergence of security risks.
Since 2012, the European Commission has been re-examining the legal norms for personal data protection and has progressively developed and refined the GDPR.
Also in 2016, the Cybersecurity Law was promulgated in China, which also mentioned strengthening the protection of personal information. The Data Security Law and the Personal Information Protection Law, which were successively issued last year, also make special provisions on online data and personal information.
But in a world where algorithms invade, it's not enough to just normalize the data.
The Provisions on the Administration of Recommendation of Internet Information Service Algorithms, which began to be implemented in China in March this year, clearly stipulate hot issues of social concern such as "big data killing", "unfair competition", and "protection of special groups". This is the first rule and regulation in mainland China specifically for algorithm recommendation, and it is also a pioneering move in the world, which to a certain extent represents the forefront of the current governance of algorithm technology.
The "Provisions" list upholding positive energy, prohibiting the recommendation of illegal information, and implementing the responsibility of the main body as the most important norms for algorithmic recommendation service providers. A review of the ethics of science and technology is on the agenda.
In view of the helpless position of users facing the algorithm, the regulations also require that users be provided with convenient service options for closing the algorithm recommendation. The latest news I pay attention to is that as of March 15, Apps such as WeChat, Douyin, Today's Toutiao, Taobao, Baidu, Dianping, Weibo, and Xiaohongshu have all been online with the algorithm off key.
WeChat personalized ads close the page
In the new regulations that have begun to be implemented, algorithm filing is a point worthy of attention. It requires algorithm service providers to open the algorithm black box from the inside, increase transparency, and through output, archiving, and retention, the filing system will become traceable.
On March 1, the "Internet Information Service Algorithm Filing System" has been launched, and now even ordinary users can go to the official website of the algorithm filing to inquire, which to a certain extent gives ordinary users the right to pay attention to, discuss and supervise algorithm issues.
With the introduction of various regulations, the "normative network" of algorithms in China is gradually being woven. At the same time, the hierarchical positioning of governance subjects is also more clear.
The State Internet Information Department is responsible for the overall planning and coordination of governance and supervision and management work, telecommunications, public security and market supervision jointly become the main body of governance, and the corresponding local departments are responsible for the relevant work of the administrative region. The provider of the algorithm recommendation service shall cooperate with the management department to carry out security assessment and supervision and inspection work, and provide necessary technical, data and other support and assistance.
Or we can think of the algorithm as the invisible "relationship line" in the network society, in the network of governance formed by the algorithm, the forces of different levels, fields, and sources are summoned out and enter the network of governance and play different roles. For example, platforms and large technology companies need to implement their responsibilities and cooperate more with relevant departments; the public can also use the rights they have to supervise the algorithms used by technology companies.
For the future of algorithmic governance, it can be seen as a network of multiple actors, and governance will tend to be standardized and stable, but it will not stop.
Doubts: Can algorithmic black boxes really be governed?
Liu Peng: Author of the public account of "Calculation Advertising"
I do not deny that the laws and policies introduced in recent years have a certain degree of positive significance, but I have doubts about the specific effects of such regulations.
Algorithms, it's really not so easy to regulate well.
After graduating from Tsinghua University, I first worked at Microsoft Research Asia for a while, studying artificial intelligence, and then went to a number of Internet companies to do computational advertising. An obvious experience is that the algorithms applied by large Internet companies are now becoming more and more inexplicable.
"Machine learning" is the mainstream implementation form recommended by the current algorithm, and the original material for learning comes from the basic information of each user, such as gender, age, education, etc., as well as the use data generated by the user in the process of use, such as likes, favorites, concerns, etc. Each type of information of the user is called "feature" feature.
The simplest type of algorithm is the linear regression algorithm, in which the weight of each feature of the user is clearly visible, so it is "explainable" and difficult to regulate.
But the problem is that simple algorithms can solve very limited problems, in order to improve efficiency, more and more Internet companies are using more complex networks of deep learning algorithms.
Many people use "alchemy" to describe deep learning.
In ancient times, alchemists would pour the collected materials into the furnace, and after a period of time to see if gold was refined, they did not know what kind of chemical reaction occurred in the process of refining, and could only try the material combination and refining process continuously.
Similarly, algorithm engineers train all kinds of available data information in complex deep learning models, and even if they find a more efficient algorithmic model, they can't explain how the data is computed and processed.
In this case, it is difficult to say which feature has a critical impact on the outcome of the algorithm's output, and the operation process of the algorithm becomes a "black box" that is difficult to explain, and the hidden problems are difficult to correct.
When the Draft provisions on Recommended Algorithms were first released in August last year, I interpreted it from a technical point of view on my video number.
I think one of the first things you should understand is that the world of algorithms is very different from the human world in terms of how they are organized.
For example, when we humans learn Go, we first learn the rules and chess theory, and after the integration is integrated, the chess strength will naturally win the game. But the process of algorithm learning is the opposite, it will first set a goal - "I want to win the game", in the algorithm we call this goal "objective function", and then call all the data that can be used to achieve this goal.
In the process of reaching this goal, even if you set many obstacles, a complex network of algorithms will bypass these restrictions and achieve the goal in other ways.
Therefore, the most important thing in the world of algorithms is actually that "objective function".
Assuming that a company's "objective function" is to optimize the user's retention time, the algorithm model will take other ways to improve the user's use time on the platform even if anti-addiction settings such as time limiting are carried out.
The "objective function" is the "lifeblood" of algorithmic governance, but this is often overlooked by legislators in various countries.
The current regulations only require companies to publish algorithmic referral services in an "appropriate manner", but do not impose restrictions on what is "appropriate".
If some companies only selectively publish the characteristics of the algorithm in key links when publishing the algorithm, it seems to be fair and transparent, but because you don't know what the objective function it really formulates is, the role of this disclosure is very limited, and may even be useless.
At the same time, there is another problem behind the disclosure of algorithms: some algorithms are related to the confidentiality of companies or commercial institutions.
Trade secrets are protected by law, so how do we define which algorithms should be published and which algorithms should be protected? There are still many questions that need further consideration on how to achieve meaningful algorithmic transparency without damaging the company's trade secrets.
*Header Image Source: Installation Art Parallels (MINI & UVA)
This article is the original article of Geek Park, please contact Geek Jun WeChat geekparkGO for reprint