The Language of Approval: Identifying the Drivers of Positive Feedback Online DOI: XXXXXXX.XXXXXXX Conference: ; Under Review; 2025 CCS: Human-centered computing Empirical studies in collaborative and social computing
Agam Goyal, Charlotte Lambert, Eshwar Chandrasekharan · 2025
Read The Language of Approval: Identifying the Drivers of Positive Feedback Online DOI: XXXXXXX.XXXXXXX Conference: ; Under Review; 2025 CCS: Human-centered computing Empirical studies in collaborative and social computing by Agam Goyal, Charlotte Lambert, Eshwar Chandrasekharan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-10370.txt.
Introduction
Agam Goyal, Charlotte Lambert, and Eshwar Chandrasekharan ask what kind of writing earns likes and awards online in this 2025 computational social science paper. Positive feedback is a quiet form of governance on platforms, yet nobody knew which features of a post actually draw rewards, or how that varies by author and community.
Working with eleven million posts from one hundred subreddits, the team combines causal methods with predictive modeling, holding author reputation, timing, and community fixed. Overly complicated language, a tentative style, and toxicity all reduce rewards. Models built on their curated features spot highly upvoted posts reliably. Auditing community rules then reveals a policy-practice gap: guidelines stress civility and formatting, and say almost nothing about the traits that actually earn praise.
Catalog record
| CREATOR | Agam Goyal, Charlotte Lambert, Eshwar Chandrasekharan |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 116 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-10370
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The Language of Approval: Identifying the Drivers of Positive Feedback Online DOI: XXXXXXX.XXXXXXX Conference: ; Under Review; 2025 CCS: Human-centered computing Empirical studies in collaborative and social computing Agam Goyal, Charlotte Lambert, Eshwar Chandrasekharan Abstract . Positive feedback via likes and awards is central to online governance, yet which attributes of users’ posts elicit rewards—and how these vary across authors and communities—remains unclear. To examine this, we combine quasi-experimental causal inference with predictive modeling on 11M posts from 100 subreddits. We identify linguistic patterns and stylistic attributes causally linked to rewards, controlling for author reputation, timing, and community context. For example, overtly complicated language, tentative style, and toxicity reduce rewards. We use our set of curated features to train models that can detect highly-upvoted posts with high AUC. Our audit of community guidelines highlights a “policy-practice gap”—most rules focus primarily on civility and formatting requirements, with little emphasis on the attributes identified to drive positive feedback. These results inform the design of community guidelines, support interfaces that teach users how to craft desirable contributions, and moderation workflows that emphasize positive reinforcement over purely punitive enforcement. Figure 1. Overview of our three-part research approach for studying drivers of positive feedback online. Panel 1 (Identify Causal Effects): We apply a selection-on-observables causal inference framework to 11M posts from 100 subreddits to isolate the causal impact of linguistic attributes on three forms of positive feedback (upvotes, awards, gold) while controlling for confounding factors like author reputation and timing. Panel 2 (Predictive Modeling): We evaluate whether the linguistic features used in our causal analysis can support real-time detection of high-quality posts through both global and local predictive models, testing their ability to surface high quality contributions before community votes fully accrue. Panel 3 (Audit & Contextualize): We systematically compare our empirical findings against prior user and moderator surveys, descriptive studies, and community guidelines through a manual audit to identify where our causal estimates align with, refine, or contradict current understanding of what drives positive reception in online communities. Our work combines causal inference, predictive modeling, and comparative analysis to provide both theoretical insights and practical implications for community design and moderation. Three-panel horizontal layout showing research methodology workflow. Panel 1 displays Reddit logo with various feedback icons including thumbs up, award medal, and gold bar, connected by arrows to a causal inference diagram with branching pathways. Panel 2 shows data processing flow from feature extraction symbols to machine learning model representations, with performance metrics displayed as AUC scores of 0.654 for global and 0.726 for local models. Panel 3 contains survey rating icons, scatter plot visualization, and audit checklist symbols, with color-coded indicators showing alignment in green, refinement in orange, and gaps in red between findings and existing knowledge sources. 1. Introduction When users make contributions to online communities, subtle linguistic choices—e.g., whether to frame content as a question or statement, use specific versus general language, or adopt formal versus casual tone—can influence its reception. These decisions at an individual level aggregate into community-wide patterns that shape which voices are amplified, whose content gains visibility, and ultimately, how communities evolve ( Atcheson et al. 2024 ) . Despite the millions of posts submitted daily across social media platforms, we lack a systematic understanding of the causal mechanisms that drive positive community feedback. Social media platforms have developed sophisticated reward systems that fundamentally restructure how content is valued and distributed. These systems use collective ratings as proxies for content quality and engagement, with mechanisms that allow users to upvote, like, or award content, and then leverage that feedback to make algorithmic decisions about visibility and prominence ( Eckles 2022 ) . 1.1. Opportunities and Challenges in Positive Feedback Mechanisms The power of these reward mechanisms extends far beyond simple content ranking. Prior research demonstrates that positive feedback creates reinforcement loops that actively shape future behavior. Being highly upvoted or receiving awards on Reddit encourages users to post content that receives even more positive feedback than they would have otherwise ( Lambert et al. 2025 ) . Similarly, creator recognition features such as “creator hearts” on YouTube improve engagement and comment visibility ( Choi et al. 2025 ) . This pattern reflects broader principles of positive reinforcement validated across offline contexts, from workplace productivity to child development ( Gray 2010 ; Wei and Yazdanifard 2014 ; Athalye et al. 2018 ) . Reward systems do not merely reflect community preferences—they actively incentivize behaviors that align with perceived community values ( Weld et al. 2021 ; Goyal et al. 2026 ) . These dynamics create both opportunities and challenges for community governance. Surveys of Reddit moderators reveal strong demand for mechanisms that highlight high-quality content and desirable behavior, rather than focusing solely on removal ( Lambert et al. 2024 ) . Furthermore, research on antisocial behavior shows that undesirable content is much more prevalent than what moderators can address ( Park et al. 2022b ) , suggesting a symmetric problem: many promising contributions likely remain underexposed while platforms and moderators focus primarily on content removal ( Gilbert 2013 ) . As a result, early-stage detection and lightweight guidance systems could help surface high-quality contributions before organic voting fully occurs, potentially improving platform engagement quality, community health, and newcomer retention by making early participation feel acknowledged. However, a fundamental gap limits progress towards building such systems: we lack a causal understanding of what drives positive feedback. Current research falls into two categories, each with its own limitations. First, surveys of moderators and users identify what communities say they value ( Lambert et al. 2024 ; Weld et al. 2022 ; Weld et al. 2024 ) , but surveys may reflect aspirations rather than realized behavioral patterns in practice. Second, descriptive and correlational studies identify attributes in content that are associated with higher engagement ( Bär et al. 2023 ; Papakyriakopoulos et al. 2023 ; Voinea et al. 2024 ; Goyal et al. 2026 ) , but correlation conflates language with author reputation, timing, and visibility. We note that this distinction is crucial: knowing that popular posts tend to use specific linguistic pattern is a reflective observation, providing limited guidance for actionable proactive interventions. Without isolating causal effects, we cannot build reliable detection or guidance tools, or explain why certain interventions succeed or fail. Addressing this causal identification challenge would enable transformative applications: real-time guidance systems that help users craft more effective contributions, early-detection algorithms that surface quality content before organic and algorithmic patterns solidify, and evidence-based community guidelines that teach effective contribution strategies to users. We believe that Reddit provides an ideal setting for addressing these challenges. Its transparent feedback mechanisms such as upvotes, awards, and gold, massive scale of posts, and community diversity spanning thousands of distinct topics enable a robust causal analysis. 1.2. Research Questions To address these challenges, we pose the following research questions (RQs): RQ1: What linguistic attributes causally drive positive feedback on Reddit? (a) Which linguistic attributes causally influence the probability that a post receives positive feedback, after conditioning on author reputation and activity as well as timing- and community-level fixed effects? (b) Are these effects outcome-specific— i.e., do they differ across feedback channels (free upvotes v.s. paid awards and gold)—or are common