Dear editors of Data Science,

We are thrilled to submit a new and final version of our manuscript “The complex link between filter bubbles and opinion polarization” (Tracking #: 671-1651), coauthored by Michael Mäs and Marijn Keijzer. We apologize for the much-delayed resubmission and thank you for your patience. As you will see below, we again took the reviewers´ comments very seriously and improved the manuscript according to their suggestions.

We thank Reviewer #1 and Reviewer #3 for their nice words. We corrected the typos listed by Reviewer #1. Responding to her recommendation, we now mention explicitly data science in the introduction together with the argument that data science can profit from the complexity perspective. We also updated the literature overview, which has grown in the past months. However, we were also able to drop a few references to outdated research.

Reviewer #2’s evaluation remained negative. His comments, however, helped us further improve the manuscript. First, we agree with the reviewer that our paper is not a review, as we are not covering a literature but only point to those findings from a specific literature that are relevant for the debate about the personalization-polarization hypothesis. To avoid confusion, we do not write in the new manuscript that we review the literature, but argued that “we collected relevant insights from the literature on opinion dynamics in social networks, demonstrating that the direction and the strength of the effect of personalization on polarization may critically depend on aspects that have not been sufficiently studied by empirical research.”
Second, he argues that we need to make clearer in the discussion what data scientists can learn from our analyses from a “practical” and an “academic” point of view. To drive home this point, we now stress in the introduction that theoretical work has the potential to “point to gaps in the empirical literature that need to be filled before one can draw conclusions” about the effects of personalization. Likewise, in the conclusion section we now argue in more detail that “a purely empirical approach to testing the personalization-polarization hypothesis can lead to false conclusions. Assume, for instance, that an empirical study quantified the degree of personalization-induced homophily in various settings and found no correlation with opinion polarization in these settings. This finding certainly challenges the personalization-polarization hypothesis. In complex systems, however, effects can take long to unfold and can then be very abrupt and strong. In Panel A of Figure 1, for instance, polarization remained low for a long time, until it grew rapidly (57). In addition, personalization algorithms are still being developed and refined. The observation that they have not contributed to opinion polarization so far, does not imply that further advances in personalization will also remain without negative effects (96). This suggests that the empirical observation that personalization so far appears to be relatively mild and its effects on opinions modest (18, 44), should not lead one to conclude that personalization will remain an innocent technology in the future.“
Furthermore, we extended the concluding section to make explicit the practical contribution of computational modeling. We added somewhat visionary paragraphs describing an idea for the use of theoretical models to aid the public debate about online social networks and the design of digital communication platforms. In a nutshell, we propose that empirically calibrated computational models can be used to conduct virtual crash tests for online social networks, using a methodology that is standard in other fields applying a complexity perspective (e.g. earth system modeling, epidemiology, finance). This new contribution is also mentioned in the abstract and the introduction. While we refrain from continuing the debate about disruptiveness as a useful criterion in science, we hope that this vision illustrates the usefulness of theoretical analysis. We added the following paragraphs.

“In addition, the complexity perspective can inform the design of interventions combatting undesired effects of online media on public discourse and democratic decision making. First, the complexity perspective puts into questions those approaches that point to the individual level and advocate that educating individual users and enhancing their so-called digital literacy will prevent undesired effects of online media. Finkel et al., for instance, argued that one should be “encouraging them [social media users] to deliberate about the accuracy of claims on social media, which causes them to evaluate the substance of arguments and reduces their likelihood of sharing false or hyperpartisan content” (51). They also proposed that users should contribute to the identification of false or hyperpartisan content and, thus, augment professional factchecking. To be sure, we do not doubt that it is useful to educate users about dynamics unfolding on online media and about how their own behavior can contribute to problematic outcomes. However, it is hard to imagine how individuals can be put in the position to reliably evaluate the truth value of online content that may have reached them via a long and usually invisible paths through the network (148). What is more, fact-check labels on some but not all content in social media can backfire. According to the so-called implied-truth effect, unchecked content appears more truthful in the presence of labels on checked content (149). Furthermore, common indicators of trustworthiness such as the number of times a piece of content has been shared or liked can be manipulated and can affected by social dynamics that have the potential to make them unreliable (150, 151). From the perspective of complexity research, it is naïve to assume that individuals are able to gain the literary needed to evaluate the effects of a complex system on them and vice versa.
Pointing to the individual level has also been used to immunize online social networks against criticism. When asked why he refused to “at least admit that Facebook played a central role or a leading role in facilitating the recruitment, planning, and execution of the attack on the Capitol”, Zuckerberg pointed to “the people who spread that content, including the President but others as well, with repeated rhetoric over time saying that the election was rigged and encouraging people to organize. I think that those people bear the primary responsibility as well.”(152) From a complexity perspective, this reasoning is problematic. On the one hand, complexity research allows one to demonstrate how individual behavior aggregates to collective outcomes and sometimes even the behavior of a single individual can be responsible for dynamics of the overall system. On the other hand, we have shown that aspects on the local and the global level can have decisive impact too. In less abstract terms, local and global characteristics of online social systems can be designed in such a way that the behavior of the same individuals does not generate undesired effects. Thus, the designers of this communication technology can influence what dynamics emerge on their platforms. From this perspective, adjusting the design of online communication systems is an important contribution to combatting excessive opinion polarization.
So, how can complexity research contribute to a better design of online communication systems? The wide application of formal models in other fields where the complexity approach is used shows that empirically validated models of social influence dynamics are a potential game changer in the public and political debate about the effects of communication technology on opinion dynamics. To inform the climate-change debate, for instance, climate models are used to quantify the impact of specific economic sectors on climate change (153). To this end, modelers compare temperature rise predicted by models considering the emissions of a given sector (e.g. air travel) with predictions of the same model assuming no emissions from this specific sector. Similarly, one could compare the levels of opinion polarization in online social networks predicted by social-influence models assuming strong and weak personalization. Such an analyses would provide a rigorous measure of the contribution of a specific web technology on the degree of polarization in a society. Obviously, companies will question the assumptions of the models when findings attribute undesired dynamics to their services. However, this would elevate the public and political debate from a discussion about untested theories and anecdotal empirical evidence to a scientific debate about the processes causing polarization. What is more, companies would be given a strong incentive to conduct the empirical research needed to test model assumptions and further improve models.
Validated social-influence models will also make it possible to conduct computational crash tests for online communication systems. In medicine, for instance, simulation software is used to predict which route a given substance will take in the human body after administration, which organ will break it down, and where it will affect the organism. Such virtual tests are conducted in early stages of drug development - before exposing humans or animals to an unexplored product. Likewise, tech companies could use social-influence models to conduct virtual crash tests before they install new algorithms on their platforms, predicting whether and under what conditions they have undesired effects. In some industries, companies are legally required to conduct such tests. Like car manufacturers who are required to put their cars through a battery of tests before putting them on the road, tech companies could be forced to demonstrate that new algorithms have no undesired effects before implementing them in system as complex as a communication network. Since digital communication technology has the potential to interfere with democratic processes on a global scale, implementing it without rigorous pre-testing is¬¬ reckless.”