Black-box Audit of YouTube's Video Recommendation: Investigation of Misinformation Filter Bubble Dynamics (Extended Abstract)

1 Introduction

In this extended abstract we describe the most important findings of an auditing study 1 originally published in [Tomlein et al., 2021], discuss their impact and outline next steps towards independent oversight of personalization behavior of large platforms, which is a general motivation of our work. Large platforms are routinely being accused of contributing to the spread of misinformation caused by their personalized routines and, at the same time, of being reluctant to revise them [Zuboff, 2019; Vaidhyanathan, 2018]. Even when they promise some changes, there is a lack of effective public oversight that could quantitatively evaluate their fulfillment. Auditing studies, such as ours, are tools that may improve such oversight.

∗This is an extended abstract of a paper that won Best paper award at Fifteenth ACM Conference on Recommender Systems (RecSys '21).

1 The implementation of the experimental infrastructure and data collected are available at https://github.com/kinit-sk/ yaudit-recsys-2021

In the study, we investigate the misinformation filter bubble creation and bursting on YouTube by simulating user behavior on the platform, recording its responses (e.g., search results, recommendations) and manually annotating them for the presence of misinformative content. Then, we quantify the dynamics of misinformation filter bubble creation and also of bubble bursting, which is the main novelty of the study. While previous works examined how a user can enter a filter bubble [Abul-Fottouh et al., 2020; Spinelli and Crovella, 2020; Hussein et al., 2020; Papadamou et al., 2020], no audits have covered if, how or with what effort can the user burst the bubble. Thus, our study extends the previous works with investigation of this important aspect.

The first contribution of this work is the analysis of YouTube's personalization behavior in a situation when a user with a history of watching misinformation promoting content and with a developed misinformation filter bubble starts to watch misinformation debunking content in an attempt to burst that misinformation filter bubble. The key finding is that watching credible content generally improves the situation, albeit with varying effects and forms, depending on a particular misinformation topic.

We align our methodology with the work [Hussein et al., 2020] which also investigated the creation of misinformation filter bubbles using sockpuppeting auditing technique. We replicate parts of Hussein's study by re-using the maximum of Hussein's seed data (topics, queries, videos), using similar scenarios and the same data annotation scheme. Therefore, we are able to directly compare the outcomes of both studies. Although we expected to see less filter bubble creation behavior than Hussein et al. due to recent changes in YouTube policies [YouTube, 2020], this was generally not the case.

As the second contribution, we report changes in misinformation video occurrences on YouTube, which took place since mid-2019 when the study [Hussein et al., 2020] was carried out. We observe worse situation regarding the topics of vaccination and (partially) 9/11 conspiracies and some improvements (less misinformation) for moon landing or chemtrails conspiracies.

2 Background and RelatedWork

Misinformation filter bubbles can be defined as states of intellectual isolation in false beliefs or a manipulated perceptions of reality. They can be characterized by a high homogeneity of recommendations/search results that share the same positive stance towards misinformation. The existing studies confirmed the effects of filter bubbles in YouTube recommendations and search results. Spinelli et al. [2020] found that chains of recommendations lead away from reliable sources towards extreme and unscientific viewpoints. Similarly, Ribeiro et al. [2020] concluded that YouTube's recommendation contributes to further radicalization of users. Abul-Fottouh et al. [2020] confirmed a homophily effect; anti-vaccine videos were more likely to be recommended for anti-vaccine videos than the pro-vaccine ones and vice versa. An algorithmic audit is a systematic quantitative probing of an online platform, used for quantification of this proportion [Sandvig et al., 2014; Hussein et al., 2020].

Crowdsourcing audit studies are conducted using real user data. Silva et al. [2020] developed a browser extension to collect personalized ads with real users on Facebook. Hannak et al. [2013] recruited Mechanical Turk users to run search queries and collected their personalized results. While crowdsourcing audits cover more realistic user conditions, this also means they are noisy (e.g., user behavior is influenced by confirmation bias). Additionally, uncontrolled environment makes comparison difficult, it is challenging to retain users, and the audits also have to tackle possible privacy issues.

Sockpuppeting audits solve these problems by employing non-human bots simulating user behavior in a predefined controlled way [Sandvig et al., 2014]. They, however, have their own methodological challenges [Hussein et al., 2020]. First, appropriate seed data such as the initial activity of bots or search queries must be selected. Second, the experimental setup must measure the real influence of the investigated phenomena while minimizing confounding factors and noise (e.g., of name, gender or geolocation). Another challenge presents labelling of the collected data for the presence of the audited phenomena, which can be expert-based/crowdsourced [Hussein et al., 2020; Silva et al., 2020] or automatic [Papadamou et al., 2020].

Existing auditing studies can be further distinguished by the domain they are applied on (e.g., social media [Silva et al., 2020; Papadamou et al., 2020; Hussein et al., 2020], search engines [Metaxa et al., 2019; Le et al., 2019; Robertson et al., 2018], e-commerce sites [Juneja and Mitra, 2021]), by adaptive systems they investigate (e.g., recommendations [Hussein et al., 2020; Spinelli and Crovella, 2020; Papadamou et al., 2020], search results [Papadamou et al., 2020; Hussein et al., 2020; Le et al., 2019; Metaxa et al., 2019; Robertson et al., 2018], autocomplete [Robertson et al., 2018]) and by phenomena they study (e.g., misinformation [Hussein et al., 2020; Papadamou et al., 2020], political bias [Le et al., 2019; Metaxa et al., 2019], political ads [Silva et al., 2020]). Recently, audits also focused on creation of misinformation filter bubbles [Hussein et al., 2020; Papadamou et al., 2020].

3 Study Design andMethodology

In the study, we let a series of agents (bots) pose as YouTube users by performing predefined sequences of video watches and query searches. They also log the platform's responses: recommended videos and search results. The predefined actions are designed to first create a misinformation filter bubble by purposefully watching videos containing (or leaning towards) misinformation. Then, agents try to burst the filter bubble by watching misinformation debunking videos. To prevent possible carry-over effects, the agents are idle for some time between their actions. The study is a partial replication of a previous study done by Hussein et al. [2020] (denoted onwards as the reference study), which allows us to directly compare quantities of misinformative content encountered by the agents in ours and in the reference study. Our research questions were twofold:

RQ1 (comparison to the reference study). Has YouTube's personalization behavior changed with regards to misinformative videos since the reference study?

RQ2 (bubble bursting dynamics). How does the effect of misinformation filter bubbles change, when debunking videos are watched?

From the reference study, we reuse two metrics: SERP-MS and normalized score. Both metrics quantify misinformation prevalence in a given list of items (videos) on the ⟨-1, 1⟩ interval. Videos are annotated as either promoting (value 1), debunking (value -1), or neutral (value 0). Normalized score is computed as an average of individual video annotations. Since it does not take order into account, it is suited for shorter lists, such as recommendations in our case. SERP-MS captures amount of misinformation and its rank, thus making it suitable for longer, ordered lists (search results in our case). It is computed as SERP-MS = ∑nr=1 (xi∗(n-r+1)) n∗(n+1) 2 , where

xi is annotation value, r search result rank and n number of search results in the list [Hussein et al., 2020].

The agents interact with YouTube following a scenario depicted in Figure 1. At the start of a run, the agent fetches its predefined configuration, including the YouTube user account and various controlled variables. It also fetches a topic with which it will work (e.g., "9/11"), including a list of 40 promoting videos, 40 debunking videos, and 5 search queries (e.g., "9/11 conspiracy"). The agent opens a browser in incognito mode, visits YouTube, logs in using the given user account, and accepts cookies. Then it visits YouTube, saves its homepage content and performs the search phase. In the search phase, the agent executes each predefined search query on YouTube in random order and saves the search results. To prevent any carry-over effect between the search queries, the agent waits 20 minutes after each query.

The agent then proceeds to create a filter bubble. It watches (for up to 30 minutes) promoting videos in random order. Immediately after watching a video, the agent saves video recommendations on that video's page and visits the YouTube homepage, saving video recommendations listed there as well. After every two videos, the agent performs another search phase. After watching all listed promoting videos, the agent follows the same procedure to burst the filter bubble using the debunking video list. After that, the agent clears YouTube history, making the used user account ready for the next run. For each selected topic, we run this scenario 10 times in parallel to reduce noise.

The agents are run from fixed geolocation (N. Virginia) to allow for better comparison with the reference study. The date of birth for all accounts is arbitrarily set to 6.6.1990. The gender is set as "rather not say" to prevent any personalization based on this attribute. The account names are composed randomly of the most common US surnames and unisex given names.

We use 5 topics in our study (same as the reference study): 9/11 conspiracies, moon landing conspiracies, chemtrails conspiracy, flat earth conspiracy, and vaccines conspiracy. The narratives associated with the topics are popular (persistently discussed), while at the same time, demonstrably false, as determined by the reference study. For each topic, the experiment requires two sets of seed videos. As a basis, we use data published in the reference study. Remaining seed videos needed for our experiment (we use more seed videos than the reference study) were collected using similar methodology as in the reference study.

Agents collect videos from three main components on YouTube: 1) recommendations appearing next to the watched videos, 2) home page videos, and 3) search results. We collect 20 videos that YouTube normally displays next to a currently watched video. For each encountered video, the agent collects: 1) YouTube video ID, 2) position of the video in the list, and 3) presence of a warning/clarification message that appears with problematic topics such as COVID-19. Other metadata, such as video title, channel, or description, are collected using the YouTube API.

To annotate the collected videos for the presence of misinformation, we use an extended version of the methodology from the reference study. Each video was viewed and annotated by the authors of this study. The videos are annotated as debunking, when their narrative provides arguments against a misinformation, neutral when the narrative discusses the related misinformation but does not present a stance towards it, and promoting, when the narrative promotes the related misinformation. We also annotate, whether the topic of a video is related to the run's topic (sometimes, misinformation from different topics appeared) or whether it relates to any misinformation topic at all. We also indicate (rare) cases of nonEnglish videos or cases where annotators' confidence about its content is uncertain. Some videos were already removed from YouTube before we managed to annotate them. We also identify videos that mock misinformation (rather than seriously debunking them). For the purpose of this study and given the constraints of manual expert annotation, we annotated only a subset of all collected videos. We annotated all search results as well as videos recommended for first 2 seed videos at the start of the run and last 2 seed videos of both phases (resulting in 6 sets of annotated videos per topic).

4 Results and Findings

We executed the study between March 2nd and March 31st, 2021. Together, we executed 50 bot runs (10 for each topic). The bots watched 3,951 videos (collected 78,763 recommendations associated with them, 8,526 of them unique), executed 10,075 queries (collected 201,404 search results, 942 of them unique), and visited homepage 3,990 times (collected 116,479 videos there, 9,977 of them unique). Overall, we recorded 17,405 unique videos originating from 6,342 channels. Five annotators annotated 2,914 unique videos (covering 255,844 appearances). In total, 244 videos were identified as promoting misinformation, 628 as debunking (including mocking videos), 184 as neutral, 1,829 as not about misinformation. Other videos numbered 29.

We report the results using SERP-MS score metrics for search results and mean normalized scores for recommendations. Since the metrics are not normally distributed with some samples of unequal sizes, we use two-sided MannWhitney U test to test our hypotheses. In cases where multiple comparisons by topics are performed, Bonferroni correction is applied on the significance level (in that case α = 0.05 is divided by the number of topics (5), resulting in α = 0.01).

Within RQ1, we investigated whether YouTube's personalization behavior changed since the reference study took place. Overall, we observe a small change towards the misinformation promoting content both for the same search queries in our and the reference data as well as for the top-6 recommendations (when comparing the up-next and top-5 recommendations together using last 10 watched promoting videos in reference watch experiments and last two watched videos in our promoting phase). In case of search results, the mean SERP-MS worsened from -0.46 (std 0.42) in reference data to -0.42 (std 0.3) in our data. In case of top-6 recommendations, the mean normalized score worsened from -0.07 (std 0.27) in reference data to -0.04 (std 0.31) in our data. However, neither of these differences is statistically significant.

More considerable shifts in the data can be observed when looking at individual topics. Regarding search, improvement can be seen within certain queries for the "chemtrails" conspiracy that show a large decrease in the number of promoting videos. On the other hand, search results for "flat earth" conspiracy worsened. Within the "anti-vaccination" topic, there is an increase in neutral videos (from 12% to 35%) and a drop in debunking videos (from 85% to 61%). This may relate to new content regarding COVID-19. Regarding recommendations, only the results for the "moon landing" and "antivaccination" topics are statistically significantly different. In case of the "moon landing" topic, we see an improvement (more debunking videos are recommended). On the other hand, we observe a drop in debunking videos (from 29% to 9%) in the "anti-vaccination" topic and a subsequent increase in neutral (from 70% to 78%) and promoting videos (from 1% to 8%). There were also more promoting videos on the "9/11" topic (27% instead of 18%), but the distribution is not statistically significantly different.

Within RQ2, we investigated what is the effect of watching debunking videos after the promoting phase, i.e., whether we will observe the "bubble bursting" behavior. We expected the metrics would improve due to watching debunking videos, i.e., that we would observe misinformation bubble bursting. Regarding top-10 recommendations, their overall distribution is significantly different when comparing the ends of promoting and debunking phases (U=7179.5, p-value=1.8e-9). Mean normalized score improves from 0.01 (std 0.31) to - 0.27 (std 0.27). Except for the moon landing conspiracy, we observe significantly different distributions for individual topics as well. The 9/11 topic shows a decrease in promoting videos, while other topics show an increase in the number of debunking videos.

We also examined the differences between the very start of the experiment and its end. We expected the metrics would improve due to watching debunking videos despite watching promoting videos before that. The distributions of SERP-MS scores in search results are statistically significantly different (U=36515, p-value=0.0002). Overall, we see an improvement in mean SERP-MS score from -0.39 (std 0.28) to -0.46 (std 0.29). Moreover, all topics except " 9/11" have significantly different distributions and improved. The improvement is due to increases in debunking videos, decreases in promoting videos, or reordered search results in some search queries. Similarly, top-10 recommendations at the end of the experiment come from significantly different distributions (U=6940.5, p-value=2.9e-7). Mean normalized score improves from -0.07 (std 0.24) to -0.27 (std 0.27).

5 Conclusions

We performed an audit of misinformation present in search results and recommendations on YouTube. We sought to verify, whether there is less misinformation present in both search results and recommendations after recent changes in YouTube policies [YouTube, 2020]. Unfortunately, we did not find a significantly different amount of misinformation in search results in comparison to the reference study of Hussein et al. [2020]. Only the "anti-vaccination" topic showed a statistically significant difference and that was in a worsening direction. Recommendations showed significant differences across multiple topics but were not significantly different overall. The"moon landing" topic improved normalized scores of recommendation, while the "anti-vaccination" topic worsened its scores. We suspect the changes in search results and recommendations were influenced mostly by changes in content (e.g., COVID-19 pandemic).

We can also conclude that users, even with a watch history of promoting conspiracy theories, do not get enclosed in a misinformation filter bubble when they search on YouTube, but they do (with varying degrees depending on the topic) in video recommendations. However, watching debunking videos helps in practically all cases to decrease the amount of misinformation that the users see.

Our study had several limitations. Only five topics were selected for investigation (to allow comparison with the reference study) and did not include more recent topics, e.g., "QAnon". Next, we included only a limited set of agent interactions with the platform (search and video watching).

We see several avenues for future work. First, auditing scenarios of a higher fidelity are needed, which would include a more human-like bot behavior such as liking or disliking videos, subscribing to channels, or clicking on the search results or recommendations. This should correspond to behavior of real user stereotypes present on the audited platform which could be learned automatically. Second, considering the cost of manual data annotation, automatic approaches labeling the presence of the audited phenomena should be researched. There have been first attempts, e.g., [Hou et al., 2019; Papadamou et al., 2020], but challenges such as robustness to concept drifts or being able to work with only limited labeled data remain to be addressed. Next, the current auditing studies are usually limited in time and scope. If they are to enable independent oversight of large platforms, they need to become continuous (longitudinal) [Simko et al., 2021] and more generic to allow comparison of the same phenomena across multiple platforms and languages. Lastly, the continuous automatized audits need to assure credibility and allow human oversight. This can be achieved by continuous monitoring of the performance of the automatic ML methods (e.g., the ones used for labeling), by enabling to examine the decisions of these methods via generated explanations and through interpretability of the used models, and by allowing the humans to override or correct these decisions (in line with human-in-command as defined in [AI HLEG, 2019]).

Acknowledgments

This work was partially supported by The Ministry of Education, Science, Research and Sport of the Slovak Republic under the Contract No. 0827/2021, and by TAILOR, a project funded by EU Horizon 2020 research and innovation programme under GA No. 952215.