2022.Sep.14

REPORTS

Report on the 11th B’AI Book Club
Ricardo Baeza-Yates “Bias on the Web” (2018)

On March 31, 2022, the 11th meeting of the B’AI Book Club, a book review by project members of the B’AI Global Forum, was held online. Atsuko Sano, a project researcher of the B’AI Global Forum, introduced the article “Bias on the Web” (2018) by Ricardo Baeza-Yates, which was published in the journal of Communications of the ACM (Association for Computing Machinery). ACM is one of the most influential international conferences in the field of computer science. This paper shows how the display of the Web is biased due to various biases in the data, the unconscious bias of developers, and the cultural backgrounds and skill differences of users, considering the data usage mechanism of the Web system. The author argues that to cope with the various biases that arise in the development and usage process, we must first be aware of the existence of such biases and then design Web systems to meet the needs of users while recognizing that the developers and users also have their own biases that are internally manifested.

Baeza-Yates’ argument can be summarized in the figure “Vicious Cycle of Bias” presented in the paper. The figure explains how biases are generated and amplified in the Web system, starting with Web bias, which includes users’ historical data. Then it categorizes these biases and shows, with evidence, how each bias causes differences in Web displays. After introducing the paper, the reviewer, Project Researcher Dr. Sano, questioned why engineers try to deal with this bias only within the system, without considering countermeasures to deal with the data bias they are utilizing, i.e., user skills and cultural and social differences.

The members’ arguments on this question were summarized into two points. First, whether it is really possible to create a bias-free space, regardless of whether it is online or real, and second, it is regarding the term “bias.” In the former issue, the members almost reached a consensus that the scope of use of AI should be limited. For example, as deep learning’s automatic generation and hands-off nature should be an advantage, the scope of use should be limited to only those areas that are not affected by bias. Should there be any influence of bias, it would continue to require modifications by humans, which would not be efficient.

The latter issue of the term became discussed when the reviewer told that it was difficult to translate “bias” into Japanese to introduce this paper in Japanese. In the paper, a variety of biases were introduced from several angles, such as statistical deviations, cognitive biases of developers and users, and differences in contributions to the Web, and differences and prejudices caused by individual environments, such as habituation to the Web and skills behind such differences. Obviously, it is very confusing to refer to all these as “bias” in Japanese. In particular, gender researchers have a negative image of bias because it reminds them of unconscious bias, which is said to be an obstacle to gender equality. However, from the developer’s point of view, simply eliminating the functions to prevent bias would have larger disadvantages. In order to reduce the load on the Web and display web content efficiently and quickly, past search history is utilized as well as customized display for individual users to make it more convenient.

Such phenomena in which images brought about by words and translations affect our understanding of reality have some similarities to other terms such as “artificial intelligence” and “social networks.” We may be perceiving social networking sites as “social” spaces that are equally accessible to all, although their content is indeed limited and displayed in a different order due to the intervention of advertisements and algorithms, or we may be under the illusion that AI is artificial human because of the word “intelligence.”

In addition, as Baeza-Yates admits, there must be other gaps and disparities that could be termed biases that affect the presentation of the Web, in addition to those presented in this paper. For example, there could be a digital gender gap in skills, access, and leadership as presented in W20 (Women20）, which is one of an official engagement group of G20, or biases among countries due to the intervention of authorities, such as not being able to speak on the Web due to censorship.

The purpose of this book review project is to consider the development of AI in Japan with reference to foreign literature. The paper by Baeza-Yate showed us once again that when incorporating such foreign findings into Japan, we must keep an eye on them from the perspective of language and translation, such as the terms “bias” and “artificial intelligence.” Furthermore, Baeza-Yates himself is from a non-English-speaking country, and the fact that he presented such a paper in opposition to the English supremacy of the Web may indicate that the promotion of diversity in the development environment is effective in reducing various biases.