What is Toutiao?
In Australia, you may not have heard of an app called Toutiao. But you must have heard of the company it belongs to: Bytedance’s other product, TikTok. In fact, in China, it has more than 120 million users every day. According to Hariharan (2023), in the Facebook – Q1 2016 earnings report, it is written that Toutiao users spend an average of more than 74 minutes a day on the app, which is far more than the average time users spend on Facebook.
Unlike traditional news applications, Toutiao is a news aggregator application; in other words, its defining characteristic is that it integrates the news content from other news applications, meaning that users don’t have to spend ages searching through massive amounts of information. News is no longer pushed by communicators like traditional media. Instead, Toutiao provides users with personalised news recommendations based on their activity and interests, which enables it to stand out among many similar products. After its launch in 2012, Toutiao quickly occupied the Chinese algorithmic news market through continuous technological innovation (Kuai et al., 2022). It has now developed into a news information creation, aggregation and distribution platform based on big data and artificial intelligence. It is second to none among Chinese news platforms in terms of the number of active users and user stickiness and is a leader in algorithmic news platforms(Kuai et al., 2022).
How Does Toutiao’s Algorithm Work?
According to the research of Haim et al. (2018), the working principle of the news recommendation algorithm is AI technology: the AI algorithm uses big data to analyse user characteristics such as age, gender, interests and social relationships, combined with the user’s historical footprint, such as previous searches, reading habits and other data. It then uses these analyses to establish corresponding user portraits, adapt the user portraits to labelled information products, and then distribute relevant news to achieve accurate personalised delivery. For example, as shown in the figure, when someone uses the Toutiao news application for the first time, they are first shown a list of channels from which they select the channel they are most interested in. Toutiao then recommends corresponding news based on the content of the channel and calculates the amount of time the user takes to read each article, their reading time per day, and the comments they post. Combined with the user’s big data, this allows for a highly personalised and accurate push of content. When the user reads the recommended content, more user data are captured using AI and entered into the big database, enabling a virtual circle for the development of the enterprise and the recommendation mechanism. If we think of the news content produced by traditional news media as a table of dishes, then the news recommended by Toutiao’s algorithm is a table of dishes that users like to eat. Under the algorithm recommendation mechanism, users do not need to spend a lot of time choosing the news content they like as long as they read the recommended news to meet their needs. Therefore, in this fast-paced digital era, it is clear that algorithmic news can beat traditional news and win more users.
However, although Toutiao’s algorithm mechanism has brought many advantages, there are still a lot of issues that need to be addressed.
Algorithmic Opacity and Algorithmic Bias
Flew (2021) defines algorithm as a computational and operational procedure for processing data. As an emerging technology, algorithm is dressed in high-tech coat. However, due to the complexity of technology and the public’s lack of understanding of and contact with it, there is a great deal of distance between it and the public. Xu (2022) writes that Toutiao’s algorithm is still a black box, and that its owner has never released the details of its algorithm. However, Pasquale (2015) argues that it is difficult to disclose the black box of algorithms and big data due to their complexity as they are often designed by a large team of engineers and hidden behind numerous confidentiality agreements. Even if the information is published, it would be highly difficult for nonprofessionals to interpret it. This shows that the algorithm has some potential dangers because it is neither fully transparent nor fully supervised, and various other issues, such as errors and biases, may arise.
At present, society contains some examples proving the objective existence of algorithmic bias. Kodiyan (2019) found that Amazon’s AI resume screening system generally gave lower scores to resumes submitted by women, while Facebook’s artificial intelligence was found to mislabel a video about black people as content about a “primate.” These examples reflect that algorithmic bias exists not only in terms of gender but also in terms of race and indicate the existence of various other biases.
According to a study conducted by artificial intelligence researchers at the University of California, Berkeley, the algorithm run by ByteDance, the company that owns Toutiao, has obvious algorithm bias (Gade, 2020). TikTok, another application owned by ByteDance, is a strong example of this bias as its recommendation algorithm recommends similar streamers and content based on users’ interests (Gade, 2020). These recommendations are not based merely on similarities between video content but are also based on the streamers’ race, gender, disability, etc (Gade, 2020). This shows that TikTok’s algorithm labels and classifies users behind the scenes, and the labels it assigns may contain prejudice and discrimination. When the algorithm forces the user to be “tagged”, it actually only gives a one-sided and incomplete interpretation of the user’s subjective personality. This incomplete treatment often leads to the creation of “algorithmic stereotypes”, which damage the user’s online personality and content style. This example shows how Toutiao’s algorithm, being a product of the same company, is bound to contain algorithmic bias as well, since they run under similar algorithms.
The causes underlying algorithmic bias may be surprising. Algorithms are not a completely value-neutral technology, and algorithmic bias is a product of social bias in the age of artificial intelligence. Jaume-Palasí and Spiekamp (2017) write that during the design stage of the algorithm development process, the algorithm absorbs and incorporates the designer’s own views about social systems, culture, practices, etc., and will eventually integrate the decisions of the designer and the institution to which the algorithm belongs to its entire lifecycle. If a bias exists in the designer’s subjective opinion, this will be absorbed into the algorithm and will eventually become part of the algorithm’s bias. In addition, the limitations of the knowledge principles on which the algorithm design is based can also result in algorithmic bias. Many ranking algorithms include popularity indicators, and such criteria represent the interests and values of the majority of people (Jaume-Palasí & Spielkamp, 2017). In short, since an algorithm is a man-made product, during its design, it inevitably incorporates the prejudices of the designers and even wider contemporary society. Therefore, algorithmic bias is actually just artificial intelligence and algorithms reflecting the prevalent prejudices in today’s society.
Filter Bubbles and the Information Cocoon
The concept of “filter bubbles” was proposed by American scholar Pariser (2011), who found that the personalised recommendation function creates a unique environment for Internet users, making their surroundings full of familiar and easily recognisable ideas. For example, Toutiao’s news recommendation mechanism will calculate the number of clicks and reading time of users regarding various types of news, so as to recommend a completely different range of information to different users. Such a function leads users to fall into a filter bubble environment, in which they are exposed to a large amount of homogeneous information for a long time. In addition, Lewis et al. (2019) found that users who like to browse news aggregation platforms generally show a tendency of low autonomy. They are often very passive, typically only browsing the news recommended by the system and not actively searching for other news. The algorithm recalculates the usage and reading habits of these users into the main database. In the long term, users are recommended increasingly similar information.
Meanwhile, in their study of recommendation algorithms, Haim et al. (2018) focused on the problem of long-term exposure to the same information, which may leave users under the influence of “information cocoon”. In other words, guided by the recommendation algorithm, users may only be exposed to one-sided and narrow information for a long time, just like being trapped in the “cocoon” of the algorithm’s recommendations. This will seriously weaken the information autonomy of the user.
Basic Paths for Effective Governance of Algorithm
Regulation of Algorithm Designers and Users from the Level of Legal Governance
In the governance of algorithms, it is essential to establish relevant laws for the accountability of algorithms. As Xu (2022) writes, algorithms are becoming increasingly involved in important decisions in people’s lives, while these algorithms often rely on, rather than eliminate, biased assumptions or data. The exposure of algorithm bias has made it clear that the attribution of responsibility and the construction of a legal system suitable for algorithms has become a practical problem that urgently needs to be solved.
At present, many countries and institutions have issued laws, regulations and principles related to governance algorithms. For example, in 2017, the American Association for Computing Machinery issued seven principles on algorithmic transparency and accountability (Xu, 2022). Similarly, on 1 March 2022, the Chinese governmeny implemented Provisions on the Recommendation and Administration of Internet Information Service Algorithm. This regulation clarifies that service providers who use algorithms need to protect the rights of users, including the right to choose and be informed about algorithms (Xu, 2022). This is also the relevant law that directly governs ByteDance, the Chinese company that owns Toutiao. Therefore, on the basis of respecting the logic of algorithm operation, algorithm bias can be reduced and potentially eliminated by clarifying the behaviour of algorithm designers and owners. The institution of formal laws and corresponding penalties can stimulate algorithm companies to make corresponding improvements to the algorithm.
Optimise algorithm design and improve algorithm transparency
The transparency of algorithms is another crucial factor for sufficient algorithmic governance as improving algorithm transparency can not only narrow the information gap with the masses but also create corresponding conditions for the establishment of more complete and careful algorithm accountability laws (Xu, 2022). According to Carlson (2018), transparency in the use of algorithms involves ensuring the objectivity and fairness of the production and recommendation process involved in the development of algorithmic news, which is related to the public’s trust in news media and their products. Improving algorithm transparency can not only be used to ensure and regulate the ability of algorithms but also improve the cultural and epistemological impact of algorithms.
Moreover, improving algorithmic technology is a fundamental way to govern algorithmic bias. At present, most algorithm training adopts the method of machine learning-supervised learning. The source of the data they learn is the label data input by the algorithm designer, and the algorithm system does not identify the existence or absence of bias and discrimination in these data; rather, it merely learns according to the established principles (Shalev-Shwartz & Ben-David, 2014). Because of the human participation in this process, it is difficult to guarantee that the algorithm data trained are completely objective and fair. Joseph et al. (2016) propose the introduction of the sociological concepts “discrimination index” and “opportunity theory” into machine learning, so as to create the idea of algorithmic scheme.
In conclusion, Toutiao is the leading aggregated news application in China. Its recommendation algorithm mechanism not only provides users with a convenient and highly personalised source of news and information but also has a profound impact on the success of Toutiao. However, it also contains several disadvantages and issues, such as low opacity, algorithmic bias, filter bubbles and information cocoons. The further introduction of laws and improvements to algorithm technology can reduce the prevalence of these shortcomings and make the algorithm a more convenient tool for users.
Carlson, M. (2018). Automating judgment? Algorithmic judgment, news knowledge, and journalistic professionalism. New Media & Society, 20(5), 1755-1772.
Flew, T. (2021). Regulating platforms. Polity Press.
Haim, M., Graefe, A., & Brosius, H. B. (2018). Burst of the filter bubble? Effects of personalization on the diversity of Google News. Digital journalism, 6(3), 330-343.
Hariharan, A. (2023). Hidden forces behind Toutiao: China’s content king. YC Startup Library. Retrieved from
Jaume-Palasí, L., & Spielkamp, M. (2017). Ethics and algorithmic processes for decision making and decision support. AlgorithmWatch Working Paper, 2, 1-19.
Joseph, M., Kearns, M., Morgenstern, J., Neel, S., & Roth, A. (2016). Rawlsian fairness for machine learning. arXiv preprint arXiv:1610.09559, 1(2), 19.
Kodiyan, A. A. (2019). An overview of ethical issues in using AI systems in hiring with a case study of Amazon’s AI based hiring tool. Researchgate Preprint, 1-19.
Kuai, J., Lin, B., Karlsson, M., & Lewis, S. C. (2022). From Wild East to Forbidden City: Mapping Algorithmic News Distribution in China through a Case Study of Jinri Toutiao. Digital Journalism, 1-21. https://doi.org/10.1080/21670811.2022.2121932
Lewis, S. C., Guzman, A. L., & Schmidt, T. R. (2019). Automation, journalism, and human–machine communication: Rethinking roles and relationships of humans and machines in news. Digital Journalism, 7(4), 409-427.
Pariser, E. (2011). The filter bubble: What the Internet is hiding from you. penguin UK.
Pasquale, F. (2015). The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard University Press. http://www.jstor.org.ezproxy.library.sydney.edu.au/stable/j.ctt13x0hch
Shalev-Shwartz, S., & Ben-David, S. (2014). Understanding machine learning: From theory to algorithms. Cambridge university press.
Xu, Z. (2022). From algorithmic governance to govern algorithm. AI & SOCIETY. https://doi.org/10.1007/s00146-022-01554-4