Nudging recommendation algorithms increases news consumption and diversity on YouTube

Abstract

Recommendation algorithms profoundly shape users' attention and information consumption on social media platforms. This study introduces a computational intervention aimed at mitigating two key biases in algorithms by influencing the recommendation process. We tackle interest bias, or algorithms creating narrow nonnews and entertainment information diets, and ideological bias, or algorithms directing the more strongly partisan users to like-minded content. Employing a sock-puppet experiment ( $n=\mathrm{8,600}$ sock puppets) alongside a month-long randomized experiment involving 2,142 frequent YouTube users, we investigate if nudging the algorithm by playing videos from verified and ideologically balanced news channels in the background increases recommendations to and consumption of news. We additionally test if providing balanced news input to the algorithm promotes diverse and cross-cutting news recommendations and consumption. We find that nudging the algorithm significantly and sustainably increases both recommendations to and consumption of news and also minimizes ideological biases in recommendations and consumption, particularly among conservative users. In fact, recommendations have stronger effects on users' exposure than users' exposure has on subsequent recommendations. In contrast, nudging the users has no observable effects on news consumption. Increased news consumption has no effects on a range of survey outcomes (i.e. political participation, belief accuracy, perceived and affective polarization, and support for democratic norms), adding to the growing evidence of limited attitudinal effects of on-platform exposure. The intervention does not adversely affect user engagement on YouTube, showcasing its potential for real-world implementation. These findings underscore the influence wielded by platform recommender algorithms on users' attention and information exposure.

Keywords: computational social science; filter bubbles; news exposure; recommendation algorithm; social media.

Fig. 1.

Overview of the research design. Participants were initially screened for their YouTube browsing habits, and then qualified participants were asked to install the ResearchTube browser extension. The extension assigned participants to either treatment or control conditions and then monitored YouTube activities and administered interventions.

Fig. 2.

Average daily percentage of News videos watched by and recommended to participants in A) algorithmic nudge, B) user nudge, and C) control groups over total watches and recommendations. W1, W2, and W3 refer to the end of weeks 1, 2, and 3, respectively. The treatment ran between W1 and W3.

Fig. 3.

Average daily percentage of Political videos watched by and recommended to participants in A) algorithmic nudge, B) user nudge, and C) control groups over total watches and recommendations. W1, W2, and W3 refer to the end of weeks 1, 2, and 3, respectively. The treatment ran between W1 and W3.

Fig. 4.

Effects of treatments on the consumption of news videos and political videos. Dots represent coefficients and horizontal bars represent 95% CIs.

Fig. 5.

Political slant distribution of the videos recommended and watched, corresponding to the A) Very liberal, B) Liberal, C) Moderate, D) Conservative, and E) Very conservative participants. The three phases of the experiment are specified as pre-intervention (week 1), mid-intervention (weeks 2 and 3), and post-intervention (week 4). The top violin plot in each phase corresponds to the distribution of slant of recommendations and the bottom violin plot corresponds to the distribution of slants of videos watched. The extreme ends represent the interquartile range and the values in-between are the mean slants of the distributions. We see that for the and participants, the recommendations were not as partisan as the videos watched, suggesting that participants self-selected into their ideology rather than recommendations guiding them to do so.