The collection used in this study was created using a double-phase approach outlined in Figure 1. We adopt a “streaming” approach, wherein every day the collector takes 2 actions: it first uses a list of relevant keywords to search for videos published in the previous day and selects the videos that pass several filters for inclusion in the dataset, and then it periodically recollects the metadata about the selected videos.

Specifically, we use the YouTube Search application programming interface (API) [30] to query for videos matching the following keyword list: vaccine, vaccination, novax, provax, antivaxx, antivaxxer (guided by previous work [4,19,31,32]). We restrict the time of collection to the previous day, the location to the “US,” and the language to English. The results are sorted by relevance, and we save the top 500. Next, we use the YouTube Video API [33] to collect the metadata about the posted videos (note that it would be at most 1 day old), and the YouTube Transcript API library [34] to download their transcripts. These videos are then filtered through several selection steps to make sure they contain enough content for analysis and that they are relevant to the topic of vaccination. We require that the videos have a transcript, be at least 1 minute and at most 60 minutes in duration, and that at least 1 keyword from the query can be found in their title or transcript, or at least 3 keywords can be found in the transcript, if none exist in the title. This filter results in about 8.3% (7213/86,387) of the videos passing (mostly because so many videos are very short, the analysis of which we leave for future studies).

During the periodic phase of the collection, we requery the metadata information about the videos daily in order to compose a temporal picture of the audience’s engagement with them, as well as to capture any potential moderation actions taken by the platform. We perform this requerying every day for the first week after the video is initially retrieved, and following that, every week. After 3 months of collection, we consider whether in the past 2 time segments the counts of views, likes, or comments increase, and if they do not, we conclude the collection for those videos. During this collection, we also gather the 1000 most recent comments posted on that video using the comment Threads API, as well as the information about the poster of the video using the Channels API [35]. The final dataset spans 6 months (November 20, 2024, to May 20, 2025) and is composed of 7213 videos posted by 3872 unique channels, and their accompanying 911,219 comments. Of the 7213 videos initially collected, 302 were excluded due to missing channel-level metadata (eg, hidden subscriber counts) or engagement timestamps required for the regression models, resulting in a final analytical dataset of 6074 videos.

This observational study analyzed publicly available YouTube content collected via the official YouTube Data API v3; it involved no direct interaction with human participants. Consequently, institutional review board (IRB) approval was deemed unnecessary, and the research complies entirely with YouTube’s Terms of Service. To protect user privacy, all individual usernames, user IDs, and identifying metadata associated with the comments were removed during data processing, and the comment text was deidentified. For transparency and reproducibility, examples of the public, high-profile YouTube channels analyzed in this study are listed in Multimedia Appendix 1. These channels represent public entities and major creators with no reasonable expectation of privacy regarding their broadcasted content.

Next, we take advantage of the regular requerying of the engagement statistics to examine the temporal trend of the views, likes, and comments the video receives after its posting. Figure 4 shows the trajectories of views, likes, and comments of the videos in each of the label categories. For each statistic, we also compute an average trajectory line and show it in color. We also show (by the vertical line) the number of days at which the average trajectory reaches 90% (P90) of the statistic. For instance, videos labeled as “strongly in favor” on average reach 90% of their views at day 32 (IQR 18-64), whereas those labeled as “strongly against” reach it at day 18 (IQR 8-38).

Although it is not possible for views to decrease over time, it is possible for the number of likes or comments to decrease when a like is unselected or a comment is deleted. This introduces some noise, especially for those videos that have only one comment or like. For this reason, the analysis of the likes and comments includes only videos that had more than 1 like or comment (respectively) in our dataset. See Section H in Multimedia Appendix 1 for plots that include all videos, as well as the same statistics at thresholds 5 and 10.

Considering all of the engagement metrics, we find that the videos labeled as “against” or “strongly against” vaccination reach their peak engagement sooner than those in favor. Visually, the trend lines become sharper at the “knee” (where the distribution starts reaching a plateau). The pattern remains similar at thresholds 5 and 10 (see Section H in Multimedia Appendix 1), though the magnitude of the difference between “strongly in favor” and “strongly against” decreases in the case of comments, which are sparser than likes.

To quantitatively characterize the temporal saturation of engagements, we use a nonnegative binomial Bayesian model. The center panel of Figure 3B models the speed at which videos accumulate their audience, predicting the number of days required to reach 90% of total engagement (P90). The results from our Bayesian negative binomial models show that vaccine-hesitant content reaches its engagement peak significantly faster than neutral or provaccine content. For the trajectory of likes, videos with an A stance are expected to reach their P90 approximately 24% faster than neutral videos (IRR 0.76, 94% HDI 0.658-0.864). Similarly, SA videos are approximately 20% faster (IRR 0.80, 94% HDI 0.705-0.905).

In contrast, provaccine content exhibits a slower rate of engagement accumulation. SIF videos are predicted to take approximately 12% longer to reach their P90 peak (IRR 1.12, 94% HDI 1.016-1.242), with a credible interval entirely above 1.0. The effect for IF videos is indistinguishable from neutral content (IRR 0.99, 94% HDI 0.882-1.112). These patterns were consistent across the models for views, likes, and comments. Furthermore, we found a strong effect for channel size, with each unit increase in LogSub associated with a trajectory that is approximately 17% faster (IRR 0.83, 94% HDI 0.803-0.856), suggesting that videos from larger channels accumulate their engagement more quickly.

To address endogeneity of temporal saturation, we conducted a sensitivity analysis by refitting the negative binomial P90 models with additional control for log-transformed total views. This specification tests whether stance effects on days-to-saturation reflect genuine timing asymmetries or merely mechanical confounding. The mechanism of potential confounding is straightforward: videos receiving fewer total likes or views will, by definition, reach a 90% engagement threshold sooner than videos with higher engagement ceilings. If stance predicts both engagement volume and timing, the observed temporal asymmetry could be entirely mechanical rather than reflecting genuine differences in how quickly vaccine-hesitant versus provaccine content accumulates engagement.

The controlled models substantially attenuated the stance effects on P90 days. For “strongly against” videos, IRR for temporal saturation decreased from 0.56 (44% faster saturation; 94% HDI 0.50-0.62) in the uncontrolled baseline model to 0.80 (20% faster saturation; 94% HDI 0.70-0.90) after controlling for log(TotalViews). This 24 percentage point reduction represents substantial confounding. The control variable itself had a positive coefficient (IRR 1.042, 94% HDI 1.021-1.061), meaning that videos accumulating more total engagement reached P90 more slowly, consistent with a pattern of sustained engagement growth rather than front-loaded spikes. Approximately 54% of the uncontrolled 44% effect appears attributable to mechanical confounding, while 46% persists as a genuine timing advantage. This controlled estimate of 20% faster saturation for vaccine-hesitant content represents a more conservative but still substantial temporal asymmetry, indicating that these videos not only receive less total engagement but also concentrate their engagement into a shorter time window relative to their final engagement level. All parameters achieved acceptable convergence (R <1.01). Detailed results are shown in Section G in Multimedia Appendix 1.

In summary, we find that vaccine-hesitant content has a “speed advantage.” By raw timing, while provaccine content accumulates its audience gradually over a month (32 days to P90), vaccine-hesitant content experiences a concentrated burst of activity, saturating within 18 days (44% faster). However, a portion of this temporal advantage reflects mechanical confounding: videos with lower total engagement reach absolute thresholds sooner by definition. Sensitivity analysis controlling for log-transformed total engagement attenuates the effect from 44% to 20% faster saturation, indicating that approximately 54% of the raw temporal advantage reflects mechanical confounding while 46% persists as genuine temporal compression. This means vaccine-hesitant videos not only receive lower total engagement but also concentrate what engagement they receive into a narrower time window relative to their eventual engagement ceiling. Both the 44% practical effect and the 20% controlled effect have policy relevance: practitioners face the 44% saturation speed; researchers identify a 20% genuine temporal compression. This rapid early-stage lifecycle implies that the digital footprint of hesitant content is established quickly, suggesting that the window for effective counter-messaging or moderation is much narrower for this type of content.

In summary, the 18-day versus 32-day difference corresponds to a descriptive 44% faster saturation for vaccine-hesitant content, whereas the controlled IRR from the negative binomial model corresponds to a more conservative model-based estimate of 20% faster saturation after adjusting for total engagement volume.

Beyond the quantity of audience interaction, we aimed to understand its nature. To this end, we classified comment stances relative to the video’s own stance, analyzing the results in aggregate, by channel topic, and through the CS.

The aggregate results, shown in the “all categories” plot in Figure 5A, show that videos with SA and A stances are more likely to foster echo chambers; their comment sections are dominated by agreement, with 39.9% (6184/15,508) and 35.6% (6055/17,030) of comments agreeing, respectively (with only 18.6% (2889/15,508) and 25.7% (4379/17,030) disagreeing). In contrast, videos with SIF and IF stances attract significant contention, with higher disagreement rates of 35.9% (2751/7650) and 31.7% (2362/7440) and lower agreement rates of 24.4% (1870/7650) and 27.8% (2072/7440). This suggests that while vaccine-hesitant content fosters confirmatory engagement, provaccine content often serves as a “battleground” for public debate. Disaggregating by authority, we see that the “battleground” effect is most intense in health and entertainment (eg, 47.4% [138/291] disagreement on SIF health videos), while the knowledge topics (like doctor) appear more balanced, with “strongly in favor” videos attracting the highest rate of agreement (422/820, 51.5%) for any provaccine content.

These patterns are further explored in the ternary plots of Figure 6A, which visualize the distribution of individual videos in the tree level of agreement. Vaccine-hesitant videos (colored red and orange) mostly cluster in the top “agree” corner, providing strong visual evidence of echo chambers, particularly in the politics and society topics. Provaccine videos (colored blue) are distributed more broadly, with a significant concentration extending toward the bottom-left “disagree” corner and into the central battleground region, confirming their role as contentious battlegrounds.

Building on these observed patterns, we then colored the same distribution by the CS to quantify the level of contention (Figure 6B). The results confirm that the CS effectively captures the distinction between echo chambers and battlegrounds. The lowest controversy scores (bright teal) are heavily concentrated in the corners of the plots, particularly the top “agree” corner, aligning with the confirmatory echo chambers identified in the vaccine-hesitant videos. In contrast, the highest controversy scores (deep purple) are found in the center of the triangle, corresponding to the “battleground” videos that attract a chaotic mix of agreeing, disagreeing, and neutral comments. By comparing the topic-specific panels, we can see that the health topic contains a prominent cluster of high-controversy (purple) videos in its center, quantitatively confirming it is the most contentious domain. This shows that the CS metric captures the observed discourse patterns, mapping high consensus to low controversy and high conversational diversity to high controversy.

Figure 3C analyzes the nature of the comment section itself by modeling the proportion of disagreement and the CS. The results statistically confirm our “battleground vs echo chamber” hypothesis with quantitative evidence. In the disagreement model, provaccine stances are associated with a higher likelihood of attracting dissent. Compared to neutral videos, a SIF stance increases the odds of a comment disagreeing by approximately 5% (OR 1.05, 94% HDI 1.003-1.094). In contrast, a SA stance is associated with a decrease in the odds of disagreement by about 7% (OR 0.93, 94% HDI 0.898-0.972), with its credible interval falling entirely below 1.0. This provides strong statistical evidence that SA videos foster more agreeable, echo-chamber-like environments.

The controversy score model further clarifies this dynamic. The SIF stance is associated with a statistically credible decrease in the CS of approximately 0.075 units (94% HDI –0.119 to –0.033) compared to neutral content. At first glance, this seems counterintuitive. However, it suggests that while SIF videos attract more direct disagreement (a 2-way debate), neutral videos foster a more chaotic and unpredictable mix of “agree,” “disagree,” and “neutral” comments. The CS metric captures entropy, not just disagreement: perfectly balanced agree/disagree comments have lower CS than chaotic agree/disagree/neutral mixtures. Thus, more focused contention (SIF videos) can have lower CS than less predictable patterns. The effects of other stances (IF, A, SA) on the CS are uncertain, with credible intervals that all broadly overlap zero. Across both models, channel size (LogSub) has a significant positive effect, indicating that videos from larger channels tend to host both more disagreeable and more broadly controversial debates.

Our analysis delineates 2 distinct social environments coexisting on the platform. Vaccine-hesitant videos function largely as “echo chambers” where users reinforce each other’s views with high agreement. In contrast, provaccine videos—especially those from health institutions—function as “battlegrounds” characterized by debate and dissent. This distinction is crucial for interpretation, as it clarifies that high engagement numbers on provaccine content often reflect controversy and argument rather than consensus or support.

Next, we show that the provaccine and vaccine-hesitant content largely originates from two distinct creator ecosystems. This section details the evidence for this sourcing asymmetry by examining the topical framing, authority signals, and institutional identity of the channels in our dataset.

To contextualize our findings, we first characterized the topical composition of the 7213 videos using the categories of the posting channel. Using a hierarchical resolution of YouTube’s metadata, we categorized videos into 3 primary topics: health (4027/6074, 66.3%), politics (918/6074, 15.1%), and society (628/6074, 10.3%). This hierarchical approach was necessary due to the prevalence of multitopic assignments in the original metadata (71.8% of videos), reflecting the inherently multidimensional nature of vaccine discourse. As shown in our subsequent analyses, this topical framing emerges as a significant moderator of engagement and comment discourse.

The distribution of video stances across these channel characteristics is presented in Figure 7. By topic (Figure 7A), health-related videos are slightly more provaccine (49.3% [1987/4027] “strongly in favor” and 21% [847/4027] “in favor”), whereas politics and society topics show more balanced distributions. This suggests health-framed discussions may attract greater institutional participation, while politics and society frames host a more diverse set of creator viewpoints.

This pattern is further intricate when considering the authority of a channel’s signals extracted via manual annotation (Figure 7B). Channels identified as news or about health predominantly post content in favor of vaccination, with 66% (870/1318) and 77.7% (237/305) falling into the combined “strongly in favor” and “in favor” categories, and with minimal representation in the strongly against category. In contrast, channels run by (self-declared) doctors post content both in support and against vaccination. Finally, entertainment and opinion channels display the most heterogeneous distributions, including significant portions of antivaccine content, suggesting these creators serve diverse audiences, often without the institutional constraints that guide health or news organizations.

Figure 7C shows a clear distinction between organizational and individual creators. Organizations are more represented in provaccine content production (1103/1598, 69%) of their videos fall in the “strongly in favor” and “in favor” categories combined, versus only 10.8% (172/1598) in “strongly against,” whereas individual creators exhibit greater stance diversity. Only 18.6% (132/708) of videos from individual creators are “strongly in favor,” while 25.1% (178/708) are “strongly against,” pointing to the need for engagement with individual posters to tackle such content.

These distributional patterns are reflected in the composition of the most prolific channels in our dataset (Section I in Multimedia Appendix 1). The 10 channels that posted the most videos are responsible for just 5% of the total volume (Gini=0.39), indicating that the content is not dominated by a few voices. This list includes a mix of news organizations, such as CTV and 9NEWS, which primarily post provaccine content, alongside individual creators like a self-described doctor or a showman, who are significant contributors to the vaccine-hesitant discourse. This mix illustrates the broader patterns in action: news media are an important contributor of vaccine-related content, but influencers and alternative voices have a strong and countervailing presence.

Finally, we investigated how the source of a video impacts audience discourse by analyzing comment agreement patterns across these channel types (Figure 5B,C). The results show that the nature of the creator significantly shapes the comment section’s dynamics. Within the Authority categorization, for instance, health channels exhibit the most polarized reactions, fostering strong echo chambers for “strongly against” videos (82% agreement) while their “strongly in favor” videos become “battlegrounds” (50% disagreement). When categorized by institution, a clear distinction appears: videos from organizations tend to generate stronger echo chambers, especially for vaccine-hesitant content, while individual creators host more balanced and contentious conversations across the board.

In summary, the analysis of channel characteristics shows that provaccine and vaccine-hesitant content largely originates from two distinct creator ecosystems. The provaccine discourse is predominantly driven by organizations, particularly news and health institutions, which reach cohesive audiences. In parallel, the vaccine-hesitant discourse is more strongly represented by a diverse array of individual creators, who foster more contentious and heterogeneous comment sections. This fundamental divergence in sourcing is crucial for understanding why these two types of content exhibit such different engagement dynamics and foster such different online communities.

In the course of the collection, 243 videos at some point failed to return metadata results during the periodic recollection efforts. We manually examined the errors the YouTube interface provided for these, and found that for 116, the platform simply tells the users that the “This video isn’t available anymore,” 66 have been made private by the uploader, 50 were unavailable due to the posting account having been terminated by the platform, 3 were removed by the uploader (a message distinct from the video simply being unavailable), 6 were removed due to violating YouTube’s Community Guidelines or Terms of Services, one because the uploader has closed their YouTube account, and one because of a copyright claim. We consider the video disappearances attributed to the author separately from all others (which we attribute to the platform), and find the rates shown in Figure 8. We find that the removals by the author of the video happen more for “strongly in favor” and “in favor” videos, whereas videos labeled as “against” or “strongly against” were more likely to be removed by the platform.

Considering the error codes more closely, we find that all videos removed due to violating YouTube’s Community Guidelines were either labeled as “against” or “strongly against.” On the other hand, videos that belonged to channels terminated by the platform span all 5 categories. Whenever the video was made private by its author, the label is more likely to be “in favor” or “strongly in favor” (77.3% [51/66] of such videos) than “against” or “strongly against” (9/66, 13.6%). Thus, we find a variety of reasons for the disappearance of videos, with some evidence of the platform’s policies affecting more the vaccine-skeptic side when it comes to sanctioning content, but not necessarily the accounts themselves.

The median time of the video disappearing is 14.4 (IQR 4.9-35.8) days from its posting, meaning that half of the videos were taken down within that time. However, many remained until almost the end of our 3-month collection period, pointing to a low but constant level of moderation across time. Since most of the interactions with the video happen within the first few weeks of a video’s posting, this policy likely fails to stem the majority of the video’s impact.

Upon the examination of the metadata collected for the videos of accounts terminated by the platform (as the videos themselves were not available at the time of the analysis), we find that they largely consist of vaccine-skeptic, anti-establishment content, as well as likely semiautomated channels that repost news snippets. Those labeled as “in favor” of vaccines sometimes use sensationalized language, such as “The truth might SHOCK you!” (such phrases are often considered “clickbait” to attract an audience). Some also engage in “keyword stuffing” by including many keywords in the description of the video to increase the chance of attaining a higher rank during a search. Yet other accounts mention vaccination in the context of providing other information, such as a channel promoting US visas. Unfortunately, beyond the error message, we cannot obtain any additional information on why these accounts might have been removed; however, the above observations exemplify the low-quality content that public health institutions compete with on this platform. The last engagement information our data collection captured for these removed videos reflects the overall trend of videos against vaccination, showing more engagement than those in favor (see Section J in Multimedia Appendix 1).

Prior YouTube vaccine studies have shown that vaccine-hesitant content receives more engagement than provaccine ones [10,11,44]. However, these snapshot-based analyses reveal only an engagement gap, leaving unanswered when and how fast this gap widens. By tracking engagement longitudinally, we document a previously unquantified temporal asymmetry: vaccine-hesitant videos saturate engagement significantly faster than provaccine content (44% faster in raw timing; 20% faster in model-based estimates that control for engagement volume). Furthermore, we challenge the assumption that polarization is symmetric [18,45]. Our comment analysis demonstrates that vaccine-hesitant videos function as agreeable echo chambers [23,46], whereas provaccine videos act as contentious battlegrounds. This suggests the platform hosts two distinct discourse architectures simultaneously, mediated by content stance.

Importantly, our analysis is purely observational: we do not observe recommendation paths or run interventions, and therefore we cannot identify causal effects of YouTube’s ranking or recommendation systems; we interpret our results as patterns of realized engagement rather than causal impacts of specific algorithms.

By quantifying engagement dynamics, we observe that vaccine-hesitant videos attract substantially higher normalized engagement rates (2.8 to 2.9× higher) and reach saturation significantly faster than provaccine content. This temporal dimension suggests that vaccine-hesitant content gains traction not through sustained growth, but through rapid early-stage saturation. While the observed speed advantage (44% faster descriptively) is partly driven by lower engagement ceilings, a genuine temporal compression persists even after controlling for this confounding (20% faster after controlling for total engagement). This distinction between threshold-crossing speed and engagement velocity deserves deeper attention.

The decomposition of this temporal asymmetry shows distinct audience dynamics. The portion of the speedup attributable to confounding indicates that vaccine-hesitant audiences are smaller but more efficient: they mobilize quickly from a self-selected pool, producing a “peak-and-plateau” dynamic characteristic of viral misinformation [47]. Unlike provaccine content, which accumulates engagement incrementally as it diffuses through heterogeneous networks, vaccine-hesitant content exhibits the rapid consolidation typical of echo chambers [18,48]. This aligns with our comment-level findings: the structural homogeneity of the vaccine-skeptic audience facilitates rapid, confirmatory engagement, while the heterogeneity of the provaccine audience results in slower, contentious accumulation.

These findings identify a mechanism-specific temporal signature. Prior work established that false news spreads faster [6,15], but this study suggests that the temporal advantage operates through distinct audience structures: concentrated early-stage mobilization of in-groups versus distributed, delayed expansion into out-groups. From a policy perspective, this mechanism suggests that the critical window for intervention is not merely short (as commonly noted [6,49]) but compressed into a narrow audience demographic, concentrated in the initial 2 weeks post-upload.

The policy implications of this rapid saturation are 2-fold. First, the 18-day saturation window implies that practitioners must adopt real-time monitoring, as the majority of a video’s impact crystallizes faster than traditional fact-checking cycles can respond. Second, since moderation actions in our dataset occurred with a median delay of 14.4 (IQR 4.9-35.8) days [50,51], reactive content removal often arrives after the audience has already been saturated.

Therefore, effective intervention must target early audience composition rather than relying on post hoc removal. These empirical patterns motivate, but do not themselves test, potential intervention strategies. Conceptually, prebunking approaches that inoculate audiences against misinformation narratives [52-54] are particularly relevant here as they work before the rapid saturation occurs. Furthermore, saturated content is potentially amplified via cross-platform diffusion [55]. Platforms should consider proactive algorithmic demotion of borderline content or the elevation of credible sources during the first 2 weeks of a video’s lifecycle [56,57]. Rather than removing vaccine-hesitant content outright—a strategy fraught with free-speech concerns and often ineffective after saturation [58,59]—platforms could implement ranking systems that elevate credible sources in recommendation feeds, effectively competing for the attention of early audiences before echo chamber formation. Platform design choices that prioritize engagement metrics inadvertently amplify misinformation by selecting for emotionally resonant content [15,47]; reorienting algorithmic goals toward credibility rather than engagement represents a structural intervention with potential for large-scale impact [57,60].

While we found evidence of moderation, the median delay of 14.4 (IQR 4.9-35.8) days becomes critical when considered against the temporal dynamics we identified. Vaccine-hesitant videos concentrate their engagement—a pattern indicating rapid in-group mobilization characteristic of echo chambers—within approximately 18 days. Since moderation occurs 14.4 (IQR 4.9-35.8) days post-upload on median, content removal actions occur after the most intense period of audience concentration has already occurred, a finding that supports recent work on delayed moderation [50,51].

Given the complexities of content removal [58,59], one possible policy intervention could be the proactive algorithmic amplification of content from verified health authorities, especially during the early days of a video’s lifecycle [56,57,61]. Platform design choices that prioritize engagement can inadvertently amplify misinformation [15,47]; therefore, strategic interventions to elevate credible sources could significantly improve the health information environment without resorting to widespread censorship [60].

Our findings are consistent with these concerns, but we cannot speak directly about recommendation-system mechanisms. We therefore present these governance implications as hypotheses and directions for future experimental and audit studies, rather than as causal claims about the effects of specific policy levers.

According to our findings, online vaccine discourse is likely driven by two fundamentally different creator ecosystems. Provaccine content is predominantly produced by organizations (news and health institutions), whereas individual creators dominate the vaccine-hesitant space. This sourcing divergence is associated with different discourse patterns we observed: institutional videos tend to become “battlegrounds” of disagreement, while individual creators host “echo chambers” of agreement.

This dynamic may also be further modulated by thematic framing. Health-categorized videos, which attract the highest institutional participation, become the most intense sites of controversy [62,63]. In contrast, content framed around politics and society allows individual creators to cultivate highly agreeable, insulated communities [48,64]. This suggests a fundamental tension: institutional voices attempt to maintain a medical frame, but the discourse persistently escapes into social and political contexts where individual voices claim equal standing. To bridge this gap, public health bodies should consider partnerships with credible individual creators—such as doctors who vlog—who can foster dialogue without triggering immediate polarization [24,65,66].

It is important to situate these findings within the broader landscape of vaccine policy. Our data collection (November 2024 to May 2025) captures a moment of significant institutional uncertainty. The observed engagement asymmetry likely reflects not merely ambient polarization but the consequence of eroded trust, wherein health institutions themselves become sites of political contestation.

This raises theoretical questions for future research: Do engagement asymmetries intensify during periods of institutional distrust? Longitudinal monitoring across different policy contexts would clarify whether our findings represent a transient response to uncertainty or a structural feature of polarized health discourse. Additionally, without systematic auditing of recommendation systems, the degree to which these asymmetries reflect platform design versus audience-driven polarization remains an open question [67]. Comparative studies across platforms and jurisdictions would strengthen our understanding of these mechanisms.

Our study has several limitations that define the scope of our findings.

This longitudinal study provides systematic evidence that vaccine-hesitant content on YouTube benefits from a distinct engagement advantage: it attracts approximately 10-fold higher raw interaction volume (and 2.8- to 2.9-fold higher per-view engagement rates) relative to provaccine content and reaches audience saturation significantly faster (44% faster in descriptive medians and 20% faster in adjusted regression estimates) than provaccine content. We identify a specific temporal signature—an 18-day saturation window—driven by audience structure. Vaccine-hesitant videos mobilize homogeneous, self-selected in-groups that generate rapid, confirmatory “echo chambers,” whereas provaccine content accumulates engagement incrementally through contentious “battlegrounds” of debate.

The policy implications of this mechanism are direct. Because the median time to moderation (14.4 days) lags behind the rapid saturation of hesitant narratives, reactive content removal is structurally ill-suited to mitigate misinformation. An effective public health strategy must therefore shift from post hoc rebuttal to proactive intervention. Prebunking campaigns and algorithmic ranking adjustments that elevate credible sources during the critical first 2 weeks of a video’s lifecycle represent the most viable levers for competing with the rapid mobilization of hesitant communities.

While this study is limited to long-form English content and acknowledges the trade-offs inherent to scalable LLM classification, it establishes baseline temporal patterns essential for evidence-based policymaking. Future research must extend beyond single-platform dynamics to examine how recommendation algorithms specifically amplify these asymmetries and how content diffuses across platforms. Ultimately, navigating the complex digital health environment requires not just producing scientific knowledge, but understanding the precise temporal and structural mechanisms by which that knowledge competes for public attention.