As algorithms increasingly shape persuasive communication on digital platforms, users' awareness of algorithmic mediation is central to understanding how personalized and/or dynamically tailored messages are received. This research examines whether algorithmic media content awareness (AMCA) is platform-specific or reflects a generalized perception of algorithmic mediation. We report a constructive replication comprising two large-scale substudies and an integrative cross-platform analysis (N = 3013 combined) conducted in Chile across Facebook, Netflix, YouTube, Instagram, Spotify, and TikTok. Results support the scale's four-dimensional, higher-order structure and demonstrate scalar invariance across all six platforms, with generally comparable awareness levels. AMCA correlates positively with algorithmic familiarity, self-reported algorithmic knowledge, and perceived exposure to personalized brand content. Multigroup comparisons support a common specification of these three relationships across platforms. Overall, the findings support a platform-general representation of algorithmic awareness across structurally distinct communication environments. This advances algorithmic persuasion research by suggesting that users' awareness is organized beyond individual platform affordances. It also establishes a common basis for comparing awareness across platforms and for examining its role in responses to personalized communication.
Social media platforms are racing to deploy LLM-powered social bots in public threads, turning AI agents into visible participants in platform-mediated communication. But once the novelty fades, what is associated with users’ sustained engagement? Prior work answers largely at the post level, rarely following the same user over time. We draw on 557,645 posts from 304,400 users who mentioned CommentR, a platform-owned LLM bot on Sina Weibo, over 17 months. Combining LLM-assisted taxonomies of user purposes and bot response strategies, we analyze engagement at the turn, session, and lifecycle levels, using logistic regression, survival analysis, and a four-state Markov chain over Initial, Returning, Engaged, and Churned. From this we develop an engagement-stage account of sustained human-AI interaction in public social media threads. Engagement is sharply two-stage. The Initial state is highly transient, whereas users who reach the Engaged state show substantially greater persistence. Strategy associations are stage-dependent. Emotional Support has the highest observed rates for immediate continuation and 60-day return. Under the default lifecycle state definitions, the highest observed rates shift to Value Resonance for deepening and Personal Opinion for sustained engagement. The return gate marks continuation within an active streak, while the Returning-to-Engaged transition marks a key consolidation point among users who have already returned. These findings show that short- and long-horizon engagement follow distinct patterns and support a stage-dependent, observational account of public human-AI interaction. They also motivate multi-horizon evaluation of conversational strategies. More broadly, they highlight how platform-owned AI agents may reshape user engagement when conversational interaction becomes part of public social media infrastructure.
Social media has become an important platform for public to express their concerns about climate change, and it contains valuable information for stock market predictions. Based on 682,736 climate change–related posts from Sina Weibo between 2018 and 2024, this study applies text mining, matrix content analysis, and machine learning methods combined with SHapley Additive Explanations (SHAP) to innovatively identify the topic and sentiment evolution features, and to examine the informational value of these features in the stock market. The results show that: (1) Public attention centers on five topics: global cooperation and international politics, impacts of climate change, responses to climate change, health and lifestyle under climate change, and environmental protection philosophy; sentiment is mainly neutral (63.45%), followed by positive (29.19%) and negative (7.36%). (2) The topic “impacts of climate change” remains widely discussed and most negative, while “health and lifestyle” remains widely discussed and positive. (3) The volume of climate change-related posts and their emotional features can provide incremental information for stock market volatility prediction.
This study investigates how visual cues in dating app profiles affect perceived political ideology and romantic interest, and how these effects intersect with demographic characteristics, political ideology, and color-blind racial attitudes—a modern form of racism in which individuals believe that racism is not a significant societal issue. Using a stratified sample of 676 recent dating app users, participants viewed dating app profiles that varied in the hobby (i.e., knitting, fishing, gardening) and flag (i.e., Black Lives Matter, US, Dartmouth) displayed in the profile picture. Results show that profiles featuring fishing or US flags were perceived as more conservative by liberal participants, women, and those with lower color-blind racial attitudes. These effects were stronger for white profiles and men's profiles, indicating that visual cues may be especially salient for those dating white men. Liberal participants and those with lower color-blind racial attitudes also reported being less interested in profiles with US flags. While liberal participants were highly responsive to these signals and interpreted them as cues of political ideology, conservative participants did not, suggesting important differences in cue interpretation across political parties. The findings highlight the role of visual heuristics in partner selection on dating apps and suggest the interpretation of these cues may contribute to political homophily.
Generative artificial intelligence diffused rapidly after late 2022, raising the question of whether it would widen or narrow the digital divide for marginalized populations. We analyzed the Korean Digital Divide Survey across the transition, in 2019 and 2023, comparing three groups: the native-born Korean population, marriage immigrants, and North Korean migrants. The North Korean migrants provide a comparison for the language mechanism, a similarly marginalized group without a foreign-language barrier. Using weighted difference-in-differences with continuous-age adjustment, we found that the marriage-immigrant gap in AI use with the native-born Korean population closed and reversed, with marriage immigrants surpassing native-born Koreans (logit difference-in-differences 0.25, SE 0.12, p = .04, about 4 percentage points), while North Korean migrants did not converge (−0.13, SE 0.12, p = .30); the contrast between the two groups, the central test, was significant (0.38, SE 0.17, p = .02) and widened under full adjustment. Both marginalized groups became more positive about AI use and gained digital competency and social capital. The North Korean migrants gained even more competency, yet only the marriage immigrants converted these resources into use. The marriage-immigrant convergence was concentrated in the language-production domains, document writing and communication, where the language affordance is most direct, but North Korean migrants did not converge there. Mediation attributed the marriage-immigrant convergence chiefly to bridging social capital and digital competency; for North Korean migrants, the same resources predicted a convergence that did not occur. The pattern is most consistent with generative AI's language capability mattering for a group that operates across languages: the convergence was not shared by an equally marginalized native-speaking group and is not explained by a general narrowing of digital inclusion. Because the two migrant groups differ in more than language and proficiency is not measured, we interpret this as the most plausible explanation rather than a proven mechanism.
