Category: 3. Business

Introduction

YouTube as a Participatory Platform for Health-Compromising Discourse

Digital platforms are now central to public discourse. YouTube strongly influences opinion and social norms through its algorithmic system and creator-driven content [-]. Across domains—news, health, education, and politics—it serves as a primary information gateway, bypassing traditional media filters and enabling decentralized participation []. Yet, this openness also enables harmful discourse, including misinformation, health-compromising content, and hate speech, threatening public health and social cohesion.

A clear example is “pro-anorexia” (pro-ana) discourse, which glamorizes extreme dieting. These narratives reframe anorexia’s pathological traits as esthetic and moral ideals, creating communities around “thinspiration” images, restrictive meal plans, exercise regimens, and weight-loss stories [-]. They strongly affect adolescent females, provoking identity responses. Within these groups, starvation and bodily discipline appear as self-development and social validation tools [-]. Although members may gain emotional support and collective identity [,-], they remain vulnerable to anxiety, shame, and self-hatred when failing to meet body norms [,].

Algorithmic Structuring and Fragmentation of Pro-Ana Communities

YouTube recontextualizes pro-ana discourse differently from image-centric platforms like Instagram and Pinterest. Its mixed-genre videos—vlogs, challenges, and reviews—embed “body talk” into everyday narratives [,]. In this space, where production and consumption overlap, users develop bonds with creators via comments, likes, and subscriptions. This accelerates discourse cohesion, evolving into “video-based communities” built on shared beliefs []. These groups are defined by recurrently consumed, interacted-with videos, forming cognitive and emotional structures beyond playlists or channels—platform-based collectives driven by affect and action.

YouTube’s algorithm strengthens such collectives by analyzing behavioral data—clicks, watch time, comments, and subscriptions—to repeatedly recommend similar content [,]. This fosters “filter bubbles” and “echo chambers,” reinforcing preferences while limiting alternative exposure [,]. Content with high engagement circulates more widely, amplifying intervideo connectivity through shared audiences [,]. Prior research highlights interactivity, repetition, emotional resonance, and network structures as central to shaping these communities [,-].

While YouTube’s algorithmic curation and creator-viewer interactivity foster cohesion among similarly framed content, they hinder bridging across divergent frames [,]. Pro-ana advocacy videos often employ emotionally charged personal storytelling, eliciting empathetic engagement that is algorithmically amplified [,], which creates closed clusters []. In contrast, anti–pro-ana and recovery content, largely from public health perspectives, remains informational and authoritative, making it less effectively engaging and rarely connected to pro-ana communities [,,,]. Thus, pro-ana content forms dense but isolated clusters.

Within YouTube’s fragmented discursive ecology, pro-ana discourse functions not as a unified sphere but as heterogeneous clusters organized around competing frames. Fragmentation aligns with channel type. Institutional channels, run by news outlets or agencies, deliver authoritative critiques of disordered behaviors using credibility and infrastructure [-]. Conversely, many individual creators—neither health professionals nor public figures [,]—share autobiographical narratives that glorify thinness or describe recovery, fostering emotional immersion and parasocial bonds, and sometimes integrate monetization strategies [,,,].

Strategic Role of Recovery Frames and Meso-Level Channels

Subscriber scale influences video visibility, framing styles, and engagement. “Meso-level” creator channels—moderate in subscribers—often balance reach and intimacy. They outperform large-scale channels in emotional resonance and exceed micro-channels in visibility, making them potential “bridges” for disrupting clusters and introducing alternative frames [,]. Exploring how these channels foster frame intersection is crucial for mitigating pro-ana health risks.

Few studies have assessed the structural organization of pro-ana discourse on YouTube or strategies for transforming it into intersectable networks. Prior work has emphasized social support and emotional bonds from pro-ana messaging, showing advocacy’s appeal over prevention [,] while noting the marginalization of recovery narratives []. Yet, most analyses have focused on individual video content, neglecting intervideo ties and frame diffusion from a network perspective [,].

Accordingly, this study analyzed network-level relationships among pro-ana advocacy, anti–pro-ana, and recovery frames on YouTube. It investigated whether pro-ana discourse is structurally enclosed within platform clusters and explored strategies for more effective public health messaging. Specifically, we assessed whether recovery or anti–pro-ana narratives generate wider engagement and how they operate within the broader video network.

We emphasized the strategic value of the “recovery frame,” often conveyed through autobiographical confessions by former pro-ana individuals. Such narratives reframe eating disorders as emotionally resonant journeys, blending information and affective appeal [,]. With this hybrid structure, recovery frames can align with algorithmic preference while retaining emotional pull [,]. We tested whether recovery narratives, delivered through suitable channels, soften boundaries and enable cross-frame interaction.

This study situated the analysis in Korea, where idol-driven thinness ideals and self-discipline ethics strongly shape young women’s body norms []. It diagnosed structural isolation among pro-ana frames and evaluated recovery’s potential for diffusion. This will expand theories of platform-based health risk communication and guide youth protection strategies, including algorithm adjustment, targeted monitoring, and channel-specific messaging. Ultimately, the study will advance communication models that address emotionally driven, closed discourse communities.

Research Questions

In the digital era, health discourse is no longer expert-led and unidirectional. On participatory platforms like YouTube, algorithmic curation and user activity drive sophisticated, fragmented modes of dissemination and transformation. Within this environment, pro-ana discourse—glamorizing pathological body images—emerges as a public health threat. Understanding how it spreads and evolves structurally is urgent. This study examined how pro-ana discourse is produced and consumed through framing strategies and how these frames form distinct network structures shaped by algorithms and engagement.

As the first research question (RQ1), this study analyzed framing differences by channel operator (institutional vs individual) and subscriber scale (mega, meso, or micro). Unlike prior studies that focused on video content or anecdotal cases [,,,], this study empirically assessed how channel structure and scale influence frame selection and organization. This approach highlights discourse diversity and differentiation while offering insights for targeted public health interventions.

The second research question (RQ2) explored how pro-ana–related videos form structural networks through viewer engagement and how these networks evolve. It examined how channel attributes and framing strategies align within viewer-driven video networks. A video-level network was built using commenter overlap to define intervideo ties. Structural indicators—density, modularity, and community count—were analyzed longitudinally. This study tested the claim that pro-ana discourse comprises multiple “refracted publics” rather than a unified sphere []. It also assessed whether algorithms and user activity create increasingly enclosed, fragmented structures. As homogeneity strengthens, networks are expected to cluster tightly, with dense internal ties and sparse external links—hallmarks of echo chambers [].

The third research question (RQ3) examined how channel characteristics and frame attributes influence intervideo connections using the exponential random graph model (ERGM). Beyond descriptive engagement, the ERGM evaluates when frame-based homophily strengthens and when heterogeneous ties capable of bridging emerge. Special focus was given to meso-level individual channels, which—with moderate subscriber bases—may act as bridges when paired with anti–pro-ana or recovery frames. By identifying these roles, this study aimed to develop strategies for frame-channel collaboration in public health communication to counter the structural fragmentation and closure of pro-ana discourse on YouTube.

Methods

Ethical Considerations

This study analyzed only publicly available YouTube videos and comments; no private content was included. Public user handles were used only when necessary; however, all handles were deidentified in any shared datasets and were not publicly released. Data collection complied with the YouTube Data API Terms of Service and relevant legal guidelines. Researchers had no direct interaction with users. Since the study used noninterventional, publicly accessible, nonidentifiable data with minimal risk, no formal ethics review was required.

Data Collection

To analyze Korean-language YouTube discourse on eating disorders and extreme dieting, we used a multistage strategy in August 2024 via the YouTube Data API (v3). The procedure included (1) exploratory seed video retrieval, (2) expansion through algorithmically recommended videos, and (3) topic-based filtering and refinement.

Stage 1: Seed Video Retrieval

We identified seed videos using 4 Korean search terms linked to pro-ana and extreme dieting: “프로아나” (pro-ana), “프아 다이어트” (pro-ana diet), “뼈말라” (ppyeo-malla; “bone-thin”), and “개말라” (gae-malla; “extremely skinny”). These keywords were drawn from prior studies [,] and discourse analyses in Korean online communities. Using the YouTube Data API’s search.list endpoint, we retrieved top-ranked videos for each keyword by relevance. Videos explicitly tagged or described as pro-ana or extreme dieting formed the initial seed sample.

Stage 2: Expansion via Related Videos

To capture YouTube’s algorithmic exposure, we collected “related videos” for each seed video from the sidebar and autoplay suggestions. Because recommendations depend on viewing history, we used a clean, no-history account—free of logins, subscriptions, or prior views—to scrape related data. This reduced personalization bias and more objectively reflected default exposure chains [].

Stage 3: Sample Refinement

The initial corpus included 4174 videos. We refined the sample by (1) restricting publication to the period from January 2020 to August 2024, aligning with pandemic-linked increases in youth disordered eating []; (2) removing duplicates and videos without comment sections; and (3) excluding off-topic content after reviewing titles, descriptions, and transcripts. Videos without pro-ana advocacy, anti–pro-ana critique, or recovery promotion were discarded. The final dataset comprised 489 videos from 160 channels. Metadata included video ID, view count, comment count, channel name, and subscriber count ().

YouTube Channel Classification

To evaluate how channel traits influence frame choice and network structure, we classified channels along 2 dimensions: operational entity (individual vs institutional) and subscriber size (mega, meso, or micro). This scheme reflects institutional embeddedness and communicative reach, providing meaningful strata for analyzing framing strategies and network positions.

Operational Entity

Channels were coded as institutional or individual. Institutional channels included those managed by news outlets, broadcasters, health organizations, hospitals, public agencies, celebrity agencies, or corporations. These typically display organizational names and standardized formats. Individual channels, operated by single creators, often feature vlogs, diet routines, or testimonials, and may share personal contact details for sponsorship. This distinction captures ownership, production intent, trustworthiness, audience expectations, and algorithmic visibility [].

Subscriber Size

Subscriber count, a proxy for social influence and algorithmic visibility, was divided into 3 tiers based on influencer marketing literature [,]: mega (≥1,000,000 subscribers), meso (100,000-999,999 subscribers), and micro (10,000-99,999 subscribers).

Combining the 2 dimensions produced 6 channel types: mega-institution, mega-individual, meso-institution, meso-individual, micro-institution, and micro-individual. Meso-individual channels held the largest share (302/489, 61.8%), followed by micro-individual channels (129/489, 26.4%), showing that small- to mid-sized individual creators drive pro-ana discourse on Korean YouTube ().

Mega-individual channels, though fewer, recorded the highest average views per video, indicating disproportionately broad reach (). In contrast, meso- and micro-individual channels were central to grassroots production and engagement, making them structurally pivotal for network propagation. These classifications were later used as independent variables to explain frame distribution (RQ1), community structure (RQ2), and edge formation (RQ3).

Frame Analysis for RQ1

To address RQ1, we analyzed 489 YouTube videos to examine tensions between pro-ana narratives and public health counter-responses. Frames are defined here as interpretive structures that shape audiences’ problem recognition, emotions, and attitudes by guiding how issues are presented [,]. While traditional media once centralized frame dissemination, digital media now enables YouTubers to influence audiences through algorithmically optimized strategies [,]. Creators often use sensational or boundary-setting discourse to define “normal” bodies and acceptable behavior []. In health-risk discourses like pro-ana, frames serve not just as information vehicles but also as mechanisms for belief formation and identity performance [,]. When algorithmically reinforced, frame fragmentation intensifies polarization [,].

Based on prior studies, we categorized 3 major frames in pro-ana YouTube content. The pro-ana advocacy frame glamorizes extreme weight loss and presents thinness as ideal beauty, portraying practices as self-discipline or growth. Examples include diet vlogs, body-check challenges, and calorie tutorials, often embedded in the narratives of effort and transformation []. Creators position themselves as role models through microcelebrity strategies, embodying beauty and willpower. Their emotional bonds with followers reproduce the pro-ana frame [,].

In contrast, the anti–pro-ana frame stresses the dangers of extreme dieting and the harms of pro-ana narratives. It appears mainly in institutional channels, such as news media or expert-led accounts, using formats like news clips, interviews, and warning narratives. These videos critique sociocultural pressures that valorize thinness, reject normalization of pro-ana discourse, and emphasize the ethical need for intervention [,].

Lastly, the recovery frame centers on personal recovery from eating disorders, such as food diaries and mental health stories. These autobiographical accounts evoke empathy and identification among viewers, introducing decentered voices that disrupt pro-ana communities [,]. Expressive strategies include autobiographical narration, contrastive references to past videos, and audience feedback loops.

The frame typology was operationalized through a coding scheme as presented in , which outlines subframes under each primary frame. Subframes were coded using a multi-label strategy and aggregated under the primary frame. When multiple frames appeared in a video, dominance was determined by subframe frequency, with coder agreement resolving inconsistencies.

The coding team included 2 master’s students in media studies. After a pilot phase refining definitions and examples, independent coding was applied to all 489 videos. Intercoder reliability was assessed by double-coding 154 random videos (31% of the total), yielding Cohen κ values of 0.78-0.83 across frames, which indicate substantial agreement and coding validity.

Social Network Analysis for RQ2

To address RQ2, we used social network analysis (SNA) to examine how pro-ana videos are linked through shared commenters and how these links reflect temporal shifts in discursive cohesion and fragmentation. A video-level network was built from commenter overlap. Using the commentThreads endpoint of the YouTube Data API, we collected comments and user IDs from the 489 videos. The dataset included 1,21,991 comments (range 1‐4935; mean 254.5, SD 411.2), providing a large sample for interaction analysis ().

Stage 1: Network Construction

We built a 2-mode (video×user) affiliation matrix linking each video to its top commenters and then projected it onto a 1-mode (video×video) network. An edge between videos A and B was created if at least one user commented on both []. To reduce distortions from popular videos or high-subscriber channels [], edge weights were defined as the proportion of commenters on video A who also commented on video B, normalizing for exposure scale [].

Stage 2: Backbone Extraction

To retain only statistically meaningful ties in the skewed comment network, we applied the disparity filter, which preserves edges exceeding random expectation thresholds (P<.05) []. This extracted the network’s significant backbone, isolating structural ties indicative of discourse cohesion.

Stage 3: Community Detection and Structural Indicators

On the backbone network, we calculated global metrics, such as density, community count, and modularity, to assess cohesion. The Louvain algorithm was used for community detection []. Modularity values show how well the network is divided into modules with dense internal and sparse external ties []; higher values indicate echo chamber–like clusters resistant to frame crossover.

Stage 4: Time-Series Network Dynamics

To capture structural change over time, we split the dataset into quarterly periods and repeated backbone extraction, metric calculation, and community detection for each. This longitudinal approach tracked network shifts before and after the COVID-19 pandemic.

ERGM Analysis

To address RQ3, we used the ERGM to identify factors explaining why pro-ana YouTube videos are connected via shared commenters. Unlike descriptive SNA, which characterizes observed structures, the ERGM models the likelihood of edge formation from node attributes and structural dependencies []. This method is effective for revealing structural conditions driving public health risks, focusing not on individual content but on how information is linked, isolated, or diffused [].

The dependent variable was a binary indicator: 1 when two videos shared at least one commenter, and 0 otherwise, based on the backbone-pruned video network. This captured the presence of a narrative contagion path from a user-driven perspective.

Explanatory variables included channel type and dominant frame. Channel types were grouped into 6 categories based on ownership (institutional vs individual) and subscriber scale (mega, meso, or micro). Frames were coded as pro-ana, anti–pro-ana, or recovery. All variables were categorical.

To capture echo chamber dynamics, we added a nodematch term for frame homogeneity (whether connected videos shared the same frame). We also modeled interaction terms between channel type and frame to test whether specific combinations were more likely to produce user overlap and cross-frame ties.

Controls included (1) absolute difference in video views (log-transformed), (2) same-channel indicator (dummy variable), and (3) upload date gap in days. These accounted for exposure scale, channel ownership effects, and temporal proximity. Auxiliary models also included the GWESP (Geometrically Weighted Edgewise Shared Partners) term to capture transitivity, ensuring robustness of main effects.

The model was estimated using the ergm package in R via Markov chain Monte Carlo maximum likelihood estimation []. Convergence was checked with trace plots and autocorrelation functions, adjusting tuning parameters, such as burn-in, thinning, and step size, to stabilize standard errors. Coefficients were reported in log-odds and converted into odds ratios for interpretability.

Model fit was assessed through simulation-based goodness-of-fit tests comparing observed and simulated statistics (degree distribution, geodesic distances, and shared partners). Robustness was tested by (1) varying the backbone filter’s α level (α=.01, .05, or .10), (2) replicating the model on a binarized unfiltered weighted network, and (3) applying identical model specifications across quarterly networks to compare pre- and postpandemic structural change.

Our analytical framework () moves beyond video-level content evaluation by statistically identifying combinations of content and structural conditions that facilitate discourse diffusion. It also highlights strategic public health messaging targets within emotionally driven, algorithmically reinforced discourse environments.

‎

Results

Use of Frames by Channel Type

We cross-tabulated 8 subframes of pro-ana discourse across 6 channel types () and analyzed their distributions. Because videos may include multiple subframes—potentially violating independence assumptions of the chi-square test—and some cells had low counts, we used Monte Carlo–approximated Fisher exact tests by frame. As a sensitivity check, a Monte Carlo chi-square test was also applied to the full table, confirming a highly significant association between subframes and channel type (Monte Carlo χ²₃₅=585.12; P<.001).

Pro-ana advocacy subframes (thinness glorification, self-control, and self-harm) were concentrated in individual channels, especially meso- and micro-individual channels. Standardized residuals showed that self-control was far more frequent in meso-individual channels (z=14.95), while thinness glorification was more frequent in micro-individual channels (z=4.30). These results suggest that mid- and small-scale channels incubate messages glorifying thinness and reframing disordered eating as “self-discipline.”

Anti–pro-ana subframes (disruption of everyday life, personal responsibility, and societal responsibility) were concentrated in institutional channels. Mega- and meso-institutional channels exceeded expectations for disruption of everyday life (z=7.85 and 7.27, respectively), and both personal responsibility and societal responsibility were overrepresented in mega-institutional channels (z=3.64 and 7.52, respectively) and meso-institutional channels (z=3.74 and 5.82, respectively). These findings confirm that legacy media and professional channels drive critical discourse on eating disorder risks and societal responsibilities.

Recovery subframes, especially restoration of everyday life and self-reflection, appeared mainly in micro-individual channels. Restoration of everyday life was significantly above expectation (z=10.15; P<.001), with self-reflection also elevated (z=4.03; P<.001). This suggests that intimate, recovery-oriented narratives are most clearly articulated by a small-scale creator.

When subframes were aggregated into 3 primary frame categories and reanalyzed (), channel type and frame category again showed a significant association (χ²₁₀=397.10; P<.001). Post hoc residuals indicated strong overrepresentation of pro-ana advocacy in meso-individual channels (z=13.86); anti–pro-ana in mega- (z=10.71) and meso-institutional channels (z=8.72); and recovery in micro-individual channels (z=11.08).

These results reveal a bifurcated ecosystem: mid-scale individuals diffuse pro-ana advocacy, large institutional channels anchor anti–pro-ana critique, and small individual creators host recovery narratives. Recovery’s concentration in micro-individual channels suggests limited diffusion, underscoring the need for bridging strategies to broaden reach.

Community Structure of the Pro-Ana Video Network

After backbone filtering, we built a commenter-overlap network with 435 videos and 906 edges representing shared audiences (). Node degrees ranged from 1 to 247, with a median of 1, showing that most videos shared audiences with only a few others. Network density was low (0.96%). Louvain community detection identified 19 modules, with modularity at 0.58, suggesting closed subgroups with strong internal ties but sparse external links. Thus, pro-ana discourse resembles a fragmented ecology of isolated modules rather than a unified sphere.

To extend this static snapshot, quarterly networks were analyzed () for density, modularity, and community count. Density was higher in late 2020 but fell sharply after mid-2021 (Kwiatkowski-Phillips-Schmidt-Shin [KPSS] P=.049), showing weakened audience overlap. Modularity increased across the period (KPSS P=.046), indicating echo chamber intensification as similar-frame videos increasingly shared commenters. Community counts peaked in Q4 2022 but fluctuated without a clear trend (KPSS P=.07). Overall, the postpandemic network grew more polarized into homogeneous, self-reinforcing clusters.

‎

To analyze thematic content, we examined frame composition in the largest communities (). Community 1 (n=242) was almost entirely pro-ana, forming a dense core repeatedly drawing active commenters. Community 2 (n=79) mixed pro-ana, anti–pro-ana, and recovery frames, creating a more heterogeneous discursive space. By contrast, community 3 (n=13) was nearly all anti–pro-ana, and community 4 (n=25) was entirely recovery—both located at the periphery. These results suggest that critical and recovery messages occasionally penetrate the advocacy core but remain marginal, largely confined to peripheral enclaves. Overall, the pro-ana ecosystem is multitiered, dominated by an advocacy-centered core, with critical or recovery frames limited to sporadic bridging.

‎

ERGM Estimates

For RQ3, we estimated 2 ERGMs with commenter-overlap edges as the dependent variable (). Model 1 included main effects for channel type, video frame, and frame-based homophily, and controls for log-transformed view count difference and same-channel membership. Model 2 added interaction terms between channel type and frame. Both models converged successfully and fit significantly better than a baseline edges-only model (model 1 Δdeviance=1109.04; P<.001).

In model 1, relative to micro-individual channels, videos from mega-institutional (β=1.41; P<.001; odds ratio [OR]=4.1), mega-individual (β=1.78; P<.001; OR=5.9), meso-institutional (β=1.12; P<.001; OR=3.1), and micro-institutional channels (β=1.32; P<.001; OR=3.8) showed a significantly higher likelihood of commenter sharing. By contrast, meso-individual channels had a significant negative effect (β=−0.75; P<.001; OR=0.47). While frame main effects were nonsignificant, strong homophily emerged for anti–pro-ana (β=0.79; P<.001; OR=2.20), meaning same-frame videos were more likely to share commenters. Pro-ana and recovery homophily were positive but nonsignificant. Among controls, larger view-count gaps reduced edge probability (β=−0.77; P<.001; OR=0.46), while same-channel videos were much more connected (β=1.72; P<.001; OR=5.58).

Model 2, with interaction terms, improved fit (Δdeviance=40.99; P<.001). Pro-ana homophily became significant (β=1.14; P<.001; OR=3.12). The meso-individual×anti–pro-ana interaction was positive (β=0.44; P=.003; OR=1.55), showing that mid-scale creators of anti–pro-ana content had greater commenter overlap than expected. The meso-individual×recovery interaction was even stronger (β=0.95; P<.001; OR=2.59), indicating that recovery narratives from meso-level creators act as effective bridges across segmented networks. In contrast, the mega-institution×anti–pro-ana (β=0.40; P=.01; OR=0.67) and mega-individual×anti–pro-ana interactions (β=−0.64; P=.01; OR=0.53) were negative, suggesting that critical content from large channels generated less commenter sharing than predicted.

In summary, channel type, frame, and their interaction significantly shaped commenter-based intervideo connections. Mega-institutional channels showed high audience overlap, forming dense network regions, whereas meso-individual channels with anti–pro-ana and recovery frames acted as connectors, suggesting alternative diffusion paths. These findings indicate that mid-scale creators occupy strategic positions for exposing advocacy-dominated communities to critical and recovery narratives.

Discussion

Principal Findings

This study examined pro-ana discourse on YouTube within South Korea’s public health context, where adolescent dieting is a major concern. We analyzed how discourse is structured by channel type and framing strategy, and how overlapping commenters influence network cohesion or fragmentation.

Channels were grouped by ownership and subscriber scale into 6 types, while videos were coded into 3 frames. Analytically, we described framing variation across channel types (RQ1), used SNA to map clustering based on shared commenters (RQ2), and applied the ERGM to test edge formation by frame and channel (RQ3).

The content analysis showed that pro-ana advocacy frames were concentrated in individual-run channels, particularly meso- and micro-level creators. In contrast, institutional channels—especially those with mega or meso followings—primarily used anti–pro-ana frames stressing health risks and social responsibility. Recovery framing appeared most often in micro-level individual channels, underscoring the need to examine how such narratives spread and connect within the broader video network.

Network analysis revealed that the video network consisted of multiple tightly knit clusters. Over time, density declined while modularity increased, indicating growing fragmentation. The largest cluster was dominated by pro-ana advocacy videos, whereas anti–pro-ana and recovery content remained peripheral. Cross-frame bridges existed but were limited in both frequency and strength.

ERGM results showed that institutional channels with mega- or meso-scale audiences were more likely to be linked through shared commenters. Meso-level individual channels, though less connected overall, had significantly higher tie formation when paired with anti–pro-ana or recovery frames. This indicates the potential for bridging discourse clusters, though not causal transmission. Overall, channel type, framing, and their interaction significantly shaped the probability of intervideo connections, highlighting mid-scale individual channels as strategic for spreading health-oriented counter-narratives.

Together, the 3 analyses provide a multidimensional account of how pro-ana discourse is organized and sustained on YouTube. Videos using critical or recovery frames, when disseminated through meso-level individual channels, emerged as key cross-frame junctures. These channels represent strategic nodes for advancing public health communication.

Implications

This study shows that pro-ana discourse on YouTube does not exist within a single open sphere but rather within a fragmented, multilayered ecology shaped by algorithmic curation and selective viewing—a “refracted public” []. Meso- and micro-level individual channels use microcelebrity tactics [] and platform affordances to promote and normalize extreme dieting. These creators generate concentrated volumes of pro-ana content and foster repeated consumption within niche groups, reinforcing discursive insularity. Conversely, institutional channels emphasize recovery narratives and health warnings, but user preferences and recommendation algorithms restrict their diffusion, limiting them to pre-existing audiences.

Our network analysis using overlapping commenters quantified this segmentation. Rising modularity and declining intercluster connectivity indicate structural limits in linking recovery or critical content with dominant pro-ana clusters. Still, meso-level individual channels emerged as prolific recovery storytellers and bridges for cross-frame commenter overlaps. Their position underscores mid-scale influencers as strategic leverage points for amplifying recovery frame dissemination.

This study advances health communication research by shifting the focus from institutional campaigns or message content alone to the interplay among user participation, network structure, and channel typology. Prior studies have shown that social media reduces stigma and spreads eating disorder information [,], and influencers can drive positive change [,,]. Our work adds empirical, network-level analysis of how users engage with and circulate pro-ana frames, offering both theoretical and methodological contributions to platform-based health risk communication.

The Korean context—where thinness ideals are reinforced through fandom, beauty, and fashion content—amplifies the algorithmic visibility of pro-ana discourse. This phenomenon is not unique; similar patterns of esthetic normalization and clustering appear in Western contexts [,], suggesting broader applicability of our findings.

Practically, this study offers 3 recommendations for public health authorities and digital health communicators. First, rather than relying on blocking harmful content or top-down criticism, emotionally resonant recovery narratives are needed that align with engagement logics inside pro-ana echo chambers. Collaborating with mid-scale creators—especially those sharing recovery experiences—can reduce resistance and increase receptivity [,]. Second, interactive formats that align with YouTube’s affective interface, such as Q&A live streams, recovery-themed challenges, and first-person storytelling, can boost engagement and algorithmic visibility. Third, these strategies extend beyond pro-ana discourse to high-risk contexts like mental health and online extremism, helping transform echo chambers into more interconnected deliberative spaces.

Limitations

This study focused only on videos with publicly available comment data to build the commenter-overlap network. As a result, it excluded silent viewing patterns not captured through comments, which may overlook less interactive but influential content. Future research should integrate exposure metrics (eg, views and watch time) with survey or interview data to capture the full spectrum of engagement.

Another limitation is the lack of reliable demographic data on commenters, restricting assessment of engagement heterogeneity by age, gender, or location and limiting analysis of minority group vulnerability. We also could not confirm whether repeated commenter appearances reflected multiple accounts or bot activity, creating potential bias in interpreting edge density or centrality. This may produce an illusion of inflated interaction around certain channels or frame types, reducing the accuracy of network interpretations. Future research should combine network analysis with qualitative content analysis or targeted interviews to better capture user motivations, identities, and engagement authenticity.

Conclusions

This study analyzed the multilayered discursive architecture of pro-ana content on YouTube by examining interactions among channels, frames, and networks. Content analysis showed that advocacy, criticism, and recovery frames varied systematically by channel ownership and subscriber scale. SNA revealed frame-based clustering and growing fragmentation over time. The ERGM estimated conditional probabilities of video-to-video connections based on shared commenters, offering insights into who spreads which messages and how.

A key finding is that meso-level individual channels, when delivering recovery or critical frames, foster cross-frame overlaps in participation that may disrupt pro-ana echo chambers. Theoretical contributions include: (1) conceptualizing pro-ana discourse as a “fragmented, refracted public” rather than a unified echo chamber, (2) advancing understanding of boundary construction and bridging mechanisms in digital public spheres, and (3) demonstrating the value of combining commenter-overlap network analysis with the ERGM for studying risk communication on digital platforms.

This study highlights the importance of network-aware health communication strategies that engage mid-scale individual channels as partners in disrupting harmful discourses and amplifying recovery frames. Public health institutions should collaborate with creators to increase message trust and reach while leveraging data-driven targeting to identify and intervene with at-risk groups. Commenter-based network analysis also provides a framework for understanding content flows and designing structural interventions that can open closed communities and promote exposure to health-supportive perspectives.

Following the Ninth Circuit’s decision to uphold a series of draconian remedies against Google in the long-running Epic v. Google litigation, Google is now seeking to take its case before the Supreme Court. In a petition filed last week, Google raised a number of important legal questions ripe for the Supreme Court’s consideration—most notably: What should the legal standard be for assessing whether its series of revenue sharing, preinstallation, and distribution agreements were anticompetitive? And did the relief imposed, including a heavy-handed catalog sharing remedy that gives third-party app stores access to Google Play Store’s extensive network of apps, go beyond the scope of proper antitrust relief? These questions are not only critical to resolving Epic v. Google but also implicate similar errors in Judge Mehta’s liability and remedy decisions in the concurrent DOJ v. Google search case.

The first and most important issue Google raises concerns the rule that was applied to determine whether it acted anticompetitively. Specifically, Google’s practices were evaluated under the rule of reason, which, in its standard formulation as set forth by the Supreme Court in cases like NCAA v. Alston, involves a three-step test: first, the plaintiff presents evidence that the conduct resulted in anticompetitive harm; second, the burden shifts to the defendant to provide a procompetitive justification for its practices; and third, the burden goes back to the plaintiff to show that those benefits could have been achieved through alternatives less restrictive of competition. If the plaintiff can meet its burden at step one and, if necessary, step three, the behavior is anticompetitive and illegal. If not, the defendant wins.

As Google explains, that’s not what happened here. Rather, the District Court adopted a test in which anticompetitive harms were balanced directly against procompetitive effects, without assessing whether less restrictive alternatives existed. To be sure, courts may allow plaintiffs to prevail under the rule of reason even if they fail the third step of demonstrating the existence of a less restrictive alternative—provided they can prove that anticompetitive harms outweigh procompetitive gains. But this four-step rule of reason is typically applied where the focus is on contractual tying, such as the Ninth Circuit’s County of Tuolumne decision. And while the practice of Google Play Store requiring the use of Google Play Billing for in-app purchases could fall into that bucket, at its core Epic v. Google concerns intrabrand restrictions on Android.

An analogous mistake with applying the rule of reason can be found in Judge Mehta’s decision in the search case. In holding that Google’s allegedly exclusive default search distribution agreements with third-party browsers, Android OEMs, and wireless carriers were anticompetitive, Judge Mehta laid out the four-step rule of reason described above: first, a plaintiff shows anticompetitive harm; next, a defendant responds by showing procompetitive benefits; and then the burden returns to the plaintiff to show either that there were less restrictive means to achieve those benefits or that they are outweighed by the anticompetitive harms. However, this was the wrong test. Under the U.S. v. Microsoft standard that Judge Mehta applied, there is no room for discounting procompetitive justifications on the grounds that less restrictive alternatives might exist. Indeed, for exclusive dealing generally, a least restrictive alternative analysis is not usually conducted; courts instead simply balance harms against benefits.

In addition to its concerns with the legal standard applied at the liability phase, Google’s Supreme Court petition in Epic v. Google takes major issue with the catalog sharing remedy imposed upon Google. In general, antitrust remedies—which can take the form of prohibitory injunctions preventing a company from engaging in certain behavior, affirmative obligations requiring a company to take proactive measures, and, in exceptional circumstances, breakups or other structural relief—can serve three purposes: terminating the illegal monopolization, undoing the fruits of the violation, and preventing future anticompetitive practices. Within this scheme, the catalog sharing remedy represents an affirmative obligation for Google to undo the fruits of its statutory violation by giving third-party app stores access to Google Play Store’s catalog of apps. This effectively results in Google losing a key network advantage that makes its Play Store more attractive to users: a greater catalog of apps.

But in upholding this remedy as a “‘reasonable method’ of counteracting the Play Store’s dominance and reducing the network effects it enjoys by temporarily lowering barriers to entry,” the Ninth Circuit seems to have erred. Specifically, the “reasonable method” standard set forth by the Supreme Court in Nat’l Soc’y Professional Engineers applies either to, as in that case, prohibitory injunctions to undo the fruits of anticompetitive behavior or, as the Massachusetts v. Microsoft case made clear, affirmative obligations designed to terminate the anticompetitive effects of the illegal monopoly. It should not apply to affirmative obligations intended to deny the fruits of anticompetitive behavior, which, as the latter court explained, require a higher standard mandating that “the fruits of a violation must be identified before they may be denied.” Yet the catalog sharing remedy makes no effort to distinguish between app network effects achieved through anticompetitive versus procompetitive means.

This error is repeated in the relief approved by Judge Mehta in the Google search case. Specifically, while rightly rejecting the DOJ’s radical proposal to force Google to divest Chrome and potentially Android, Judge Mehta similarly imposed a series of data sharing remedies that, as he made clear, “are designed primarily to deny Google a key fruit of its anticompetitive conduct—scale—and to help rivals overcome that deficit.” In particular, Judge Mehta required Google to share certain search index and user-interaction data with competitors to help improve their own search services. However, like the Ninth Circuit, Judge Mehta merely asked whether this relief was a “reasonable method of eliminating the consequences of the illegal conduct,” rather than precisely identifying which data constituted the fruits of Google’s anticompetitive behavior, as opposed to data Google obtained through the normal, procompetitive operation of its search service.

The Supreme Court doesn’t take many cases a year, and major antitrust decisions from the Court, as this one would be, are always quite rare. However, amidst the number of landmark antitrust cases against Big Tech companies that will, whichever way they are decided, have huge implications both for antitrust law and the American economy, the Epic v. Google case presents a unique opportunity for the Supreme Court to head off potential legal confusion by providing necessary guidance in two key areas where Judge Mehta in the Google search case also appears to have erred. Specifically, by clarifying which version of the rule of reason applies to different forms of conduct and what level of scrutiny should govern affirmative obligation remedies intended to divest the fruits of anticompetitive behavior, the Court can lay out a much-needed framework to guide lower courts as they adjudicate these once-in-a-generation antitrust actions against Big Tech.

Engine to provide complete digital banking platform for 2+ million Tangerine clients in Canada

TORONTO and LONDON, Nov. 4, 2025 /PRNewswire/ — Tangerine Bank (Canada’s award-winning digital bank and wholly-owned subsidiary of Scotiabank, one of the “Big 5” banks in Canada with assets of approximately $1.4 trillion) and Engine by Starling (the Starling Group’s banking Software-as-a-Service (SaaS) business) today announced an agreement to deliver a next-generation banking platform for more than 2 million Tangerine clients in Canada.

Under the terms of the 10-year agreement, Tangerine will upgrade its core digital banking system to Engine’s cloud-native banking platform, enabling the digital bank to supercharge its client experience and embark on an ambitious new phase of growth.

With Engine’s SaaS platform, Tangerine’s clients will experience best in class digital onboarding, chequing accounts, instant access savings, overdrafts, debit cards and smart money management features such as card controls and spending insights, delivered through an intuitive mobile app. Behind the scenes, Engine’s end-to-end platform will provide a simplified account view and consolidate the capabilities and support tools Tangerine needs to reduce operational cost and complexity for employees.

Tangerine becomes Engine’s first North American client after the British firm announced offices in New York and Toronto earlier this year. Born of the UK’s Starling Bank in 2022, the company currently supports Salt Bank in Romania and AMP Bank GO in Australia.

Terri-Lee Weeks, President and CEO of Tangerine, said: “Tangerine chose Engine to help build the future of banking services for our clients – delivering a premier banking experience with intuitive, personalized features that evolve with client needs. Engine’s modern core banking system uniquely provides an end-to-end platform on which Tangerine can innovate quickly and continuously, reducing the time-to-market for new products and features, and delivering world-class experiences for our clients – all while staying true to the client-first design that Tangerine is known for in Canada.”

Sam Everington, CEO of Engine by Starling, added: “Engine’s technology and operating model is a tried and tested blueprint for building market-leading digitally-native banks. It is a true fintech success story as we see our software enabling ambitious, innovative and customer-centric banks all over the world. This agreement with Tangerine is a major milestone and the largest deal we have signed to date, showing just how scalable and adaptable Engine is.”

This announcement follows Engine’s expansion into the North American market to support its global growth and to develop new capabilities. Tangerine will benefit from a dedicated Engine team in Toronto consisting of product, delivery and technical specialists, who will now collaborate to deliver a refreshed suite of digital features and services.

About Tangerine Bank:

Tangerine is one of Canada’s leading digital banks, empowering over two million clients to reach their goals and move their finances forward. Known for a simple-to-use digital and mobile experience, Tangerine offers everyday banking products without any complicated hoops to jump through. From saving and spending to investing and borrowing, Tangerine’s products are designed to meet the unique needs of Canadians. Tangerine’s commitment to putting clients first has earned the bank recognition as the #1 Bank in Canada by Forbes in 2025 and the most awarded midsize Bank by the J.D. Power Canada Retail Banking Satisfaction Study for 14 consecutive years as of 2025**. Tangerine Bank was launched as ING DIRECT Canada in 1997. In 2012, Tangerine was acquired by Scotiabank and operates independently as a wholly owned subsidiary. Tangerine is a registered trademark of The Bank of Nova Scotia, used under license.

For more information, visit tangerine.ca or connect with us on social on Instagram, LinkedIn, or TikTok.

About Engine by Starling
Engine by Starling is a SaaS technology provider with the goal of bringing its modern banking platform to banks around the world. The Engine platform, built to power Starling in the UK, is modular, API-based, cloud-native and a proven technology at scale.

For further information about Engine by Starling, please visit: enginebystarling.com

About Starling Group
Starling Group includes Starling Bank, the fully licensed and regulated UK bank, Engine by Starling, a Software-as-a-Service (SaaS) provider, and Fleet Mortgages, a specialist Buy-to-Let mortgage lender. Headquartered in London, the Group has offices in Cardiff, Manchester and Southampton.

Photo – https://mma.prnewswire.com/media/2813535/Tangerine_Tangerine_and_Engine_by_Starling_sign_agreement_to_pro.jpg 
Logo – https://mma.prnewswire.com/media/2813537/Tangerine_Tangerine_and_Engine_by_Starling_sign_agreement_to_pro.jpg

By Joseph Adinolfi and Frances Yue

The pioneering cryptocurrency is on the cusp of a bear market, data shows

Bitcoin is down nearly 20% from its record high of $126,272.76 in early October.

Bitcoin prices retreated below $100,000 for the first time since June on Tuesday, bringing the pioneering cryptocurrency to the cusp of a bear market, Dow Jones Market Data showed.

The largest cryptocurrency (BTCUSD) briefly traded as low as $99,982 before recouping some losses and bouncing back to $101,269, according to FactSet data. It was down 19.8% from its record high of $126,272.76 on Oct. 6, but still up 8.5% year to date. An index or asset is considered to be in a bear market after a drop of 20% or more from a recent high.

Bitcoin was off by more than 5% on the day in recent trading, leaving it on track for its biggest one-day drop since April 3, Dow Jones Market Data showed.

The selloff in cryptocurrencies like bitcoin has coincided with a loss in altitude for other popular momentum trades. Gold (GC00), small-caps and quantum-computing stocks like Rigetti Computing Inc. (RGTI) were also coming under pressure on Tuesday.

“People are in the gold trade, people are in the uranium trade, people are in the quantum computing trade, people are in the small-cap trade,” said Ram Ahluwalia, chief investment officer at Lumida Wealth. “They’re all rising and falling together.”

Ahluwalia said that while bitcoin is technically on the cusp of a bear market, veteran crypto investors have endured much larger drawdowns over the years. “For people who are seasoned in this asset class, this isn’t a big deal. I think this is just a shakeout.”

The roots of the selloff can be traced back to the October Federal Reserve meeting, he said. The central bank announced an interest-rate cut on Oct. 29, its second this year. But during the press conference that followed, Fed Chair Jerome Powell expressed some uncertainty about another reduction in December. This has been bad news for bitcoin, Ahluwalia explained, since lower rates typically help juice speculative assets like cryptocurrencies.

Katie Stockton, founder and managing partner at Fairlead Strategies, pointed out earlier this week that bitcoin had broken below its 200-day moving average, suggesting that there could be more downside ahead in the near term. Her technical analysis suggested that the next reliable support level for bitcoin would be around $94,200.

If bitcoin continues to trade at or around its current price, it would mark its lowest 4 p.m. level since June 22, when it traded at $98,923.77. The crypto was on track for its worst three-day stretch since Oct. 11, a period when it fell by 9.9%.

-Joseph Adinolfi -Frances Yue

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

  (END) Dow Jones Newswires

  11-04-25 1538ET
Copyright (c) 2025 Dow Jones & Company, Inc.

Aptar Closures has been recognized with two prominent awards at the 2025 Association for Dressings & Sauces (ADS) Annual Business Forum, which was held on October 5th in Scottsdale, Arizona. Aptar Closures earned multiple awards this year, underscoring its leadership and innovation in packaging solutions in the dressings and condiments space.

Supplier Partner of the Year – Packaging Category

Aptar Closures was named Supplier Partner of the Year in the packaging category. Among other factors, this award is determined by nominations from ADS manufacturer member companies, who recognize ADS supplier members who have elevated themselves as a partner through equipment & service, ingredients, or packaging. According to an ADS nominating member, the team at Aptar Closures “actively participates in material qualification and helps define the quality parameters that are meaningful to the consumer. They continue to find solutions, improvements, and keep the consumer needs in mind.”

Supporting Vendor for Package of the Year

Aptar Closures also received the Supporting Vendor for ADS’ Package of the Year award, celebrating its contribution to McCormick’s new Frank’s RedHot Squeeze Sauces, which utilize Aptar’s Tower closure.  Tower is a flip-top dispensing closure that provides one-handed, user-friendly convenience and ensures precise, controlled dispensing for a smooth drizzle.

“Aptar Closures has consistently demonstrated its value as a trusted partner and leading vendor in the industry. This recognition is well-deserved, and we are proud to showcase Aptar’s outstanding achievements,” said Jeannie Milewski, President of The Association for Dressings & Sauces.

In addition, Aptar Closures solutions were featured on two other winning products at the 2025 ADS Annual Business Awards.

Katie Schomberg, Food Market Director for Americas at Aptar Closures, added, “Co-creating value with our customers and delighting consumers is at the forefront of everything we do, and we are honored to be recognized by our customers and industry peers for our collaboration and innovative packaging in these notable food categories.”

About The Association for Dressings & Sauces (ADS)  

Founded in 1926, the Association for Dressings & Sauces is an international trade association representing manufacturers of salad dressing, mayonnaise, and condiment sauces as well as suppliers of raw materials, packaging and equipment to the industry. Its purpose is to serve the best interests of industry members, its customers, and consumers of its products. For more information about ADS, visit the Association’s website or follow ADS on Twitter, Instagram, Facebook, or LinkedIn.

Google is hatching plans to put artificial intelligence datacentres into space, with its first trial equipment sent into orbit in early 2027.

Its scientists and engineers believe tightly packed constellations of about 80 solar-powered satellites could be arranged in orbit about 400 miles above the Earth’s surface equipped with the powerful processors required to meet rising demand for AI.

Prices of space launches are falling so quickly that by the middle of the 2030s the running costs of a space-based datacentre could be comparable to one on Earth, according to Google research released on Tuesday.

Using satellites could also minimise the impact on the land and water resources needed to cool existing datacentres.

Once in orbit, the datacentres would be powered by solar panels that can be up to eight times more productive than those on Earth. However, launching a single rocket into space emits hundreds of tonnes of CO₂.

Objections could be raised by astronomers concerned that rising numbers of satellites in low orbit are “like bugs on a windshield” when they are trying to peer into the universe.

The orbiting datacentres envisaged under Project Suncatcher would beam their results back through optical links, which typically use light or laser beams to transmit information.

Major technology companies pursuing rapid advances in AI are projected to spend $3tn (£2.3tn) on earthbound datacentres from India to Texas and from Lincolnshire to Brazil. The spending has fueled rising concern about the impact on carbon emissions if clean energy is not found to power the sites.

“In the future, space may be the best place to scale AI computers,” Google said.

“Working backward from there, our new research moonshot, Project Suncatcher, envisions compact constellations of solar-powered satellites, carrying Google TPUs and connected by free-space optical links. This approach would have tremendous potential for scale, and also minimises impact on terrestrial resources.”

TPUs are processors optimised for training and the day-to-day use of AI models. Free-space optical links deliver wireless transmission.

Elon Musk, who runs the Starlink satellite internet provider and the SpaceX rocket programme, last week said his companies would start scaling up to create datacentres in space.

skip past newsletter promotion

after newsletter promotion

Nvidia AI chips will also be launched into space later this month in partnership with the startup Starcloud.

“In space, you get almost unlimited, low-cost renewable energy,” said Philip Johnston, co-founder of the startup. “The only cost on the environment will be on the launch, then there will be 10 times carbon dioxide savings over the life of the datacentre compared with powering the datacentre terrestrially.”

Google is planning to launch two prototype satellites by early 2027 and said its research results were a “first milestone towards a scalable space-based AI”.

But it sounded a cautionary note: “Significant engineering challenges remain, such as thermal management, high-bandwidth ground communications and on-orbit system reliability.”