Authors: Árni Már Einarsson, Rasmus Helles, and Stine Lomborg; Center for Tracking and Society, University of Copenhagen, Copenhagen, Denmark
Question: How does the use of algorithmic agenda-setters by media organizations, which are deployed “to optimize performance, relevance, engagement, and customer conversion,” impact the quality and diversity of political news presented to readers – especially ahead of key elections?
Significance: There are a variety of reasons media organizations rely on commonly-used, algorithmic news recommenders to increase “clicks.” The same cost-utility motivation could be further enhanced by new, AI-enabled systems. To mitigate potential biases of news recommender systems (NRSs) powered by AI, media organizations need to understand the salience, exposure, and tilt of news and actors that current systems promote. As explained in the study: “Personalized filtering systems are adopted by media organizations to add relevance, enhance user experiences, and increase revenues, but also denote a fundamental shift in news curation, replacing handheld curation of a single interface with automatic curation of an infinite number of varied interfaces. From the editorial perspective, this shift causes concerns for the alignment between editorial values and filtering systems, biases introduced to news exposure, and the societal role of news media as agenda-setters and gatekeepers.”
Theory: Researchers based their construction on agenda-setting theory – which posits that the media and communications channels, through their ability to identify, frame, publicize, or ignore certain issues, can pivotally sway public opinion towards certain perspectives. The theory also suggests that media bodies can subsequently shape which problems receive attention from governments and organizations. In this case, researchers examined what agendas are more regularly set by algorithmic recommendations of news articles.
Experiment: Using agenda-setting theory, researchers conducted a controlled digital experiment on Ekstra Bladet, a Danish tabloid news outlet that regularly reaches about a third of the Danish population online. The experiment was conducted by presenting real-time news on Ekstra Bladet’s digital platforms to two randomly selected groups, one exposed to personalized recommendations (n = 63,051) and one not exposed (n = 16,051). Taking place in the period leading up to the Danish election, from October 3rd to October 31st, 2022, researchers examined differences in exposure of topics (collated into “hard” and “soft” news), issues, and political agendas. Exposure was measured not just by concrete visibility on the website, but also through behaviors triggered by exposure: reduced scanning, further research, or exploring lists of similar articles.
Findings: As for the impact of news recommenders on political agenda-setting, there were four major findings from the Danish study.
-
Significant, but small, exposure effects: The study found that the effects of NRSs on news exposure were significant but small.
-
Shifts towards softer news content: Personalized recommendations led to a higher exposure to softer news content and a less diverse range of topics, political actors, and issues.
-
Certain political actors were exposed more: The study noted that right-wing parties received more exposure than left-wing parties. This disparity was not directly due to political bias in the algorithms but rather an indirect effect. Right-wing parties were more often featured in stories combining tabloid and political news, which garnered more clicks due to the combination of soft and hard topics.
-
News exposure becomes less diverse: Analysis of diversity in news topics, political issues, and featured actors showed lower diversity for the treatment groups exposed to NRSs.
Reflection for Democracy: Due to the pivotal importance of healthy media ecosystems to democratic political outcomes, the findings raise particular democratic concerns about the choice to deploy algorithmic news recommenders with content-based metrics. Firstly, exposure effects arose not just from the re-ranking of content, but specifically from altered consumption patterns – meaning that NRSs impacted reader interest in certain topics. Additionally, researchers raised concerns that the NRSs’ mirrored preference for “softer” news content could subtly influence public perception away from political issues by prioritizing more engaging, but less critical, news. Softer, less diverse news could concerningly lead to further audience fragmentation and the narrowing of public discourse, as it creates the circumstances for decreased public consensus and motivation to act on key issues.
Additionally, the third finding emphasized the role of journalists as agenda-setters. NRSs were more likely to promote articles about right-wing parties and figures, because of the choice of publications to report more tabloid-esque stories and softer news articles on the right-wing. This is a stark reminder that machine-learning models themselves don’t “think” or “recommend,” but rather, follow encoded protocol to regurgitate the text and stories that are already out there.
While the immediate effects of algorithmic news recommenders may seem minor, the long-term implications for democratic engagement and informed citizenship could be significant. In its conclusion, the study advocates for enhanced monitoring of outputs, alongside development of new methods, to ensure more balanced and diverse news exposure through algorithmic systems.
