Guangrui Fan

As large language model (LLM) agents become routine participants in public online conversation, understanding how their replies reshape human-to-human interaction structure is a central challenge for social computing and platform governance. Yet most post-deployment evaluations focus on toxicity or engagement volume, leaving thread-level conversational structure—a dimension critical to deliberative quality, social capital formation, and equality of voice—largely unmeasured. We address this gap through a large-scale structural evaluation of @CommentR , a production LLM agent serving millions of users on Weibo, China’s leading microblogging platform. Modeling each thread as a directed reply graph, we match over 216,000 real post-deployment threads on strictly pre-anchor covariates and estimate effects on human-only conversational structure using a doubly robust estimator. Because agent replies often arrive before any human-to-human interaction is observed, we introduce lifecycle-aware estimands that distinguish early-stage formation effects from mature-thread rewiring effects. Under conditional ignorability, agent replies reduce reciprocity, increase branching, and reduce geographic homophily in early-stage threads, while degree-corrected bridging remains unchanged—consistent with a hub-and-spoke shift from dialogic exchange toward one-off commenting around a focal reply. In mature threads, eligibility for the incumbent-rewiring analysis is itself reduced by the agent—sustained incumbent exchange becomes less likely—so we report the mature-thread rewiring estimates as bounded rather than point-identified and treat the formation regime as the one in which our structural evidence is secure. The magnitude of the reshaping depends on how the agent answers: more comprehensive, factual replies produce a larger focal shift. Reply-target analysis confirms that a substantial share of human comments redirect toward the agent, reducing lateral human-to-human exchange. These findings demonstrate that public AI agents can reshape not only what people say but how people talk to each other , with direct implications for platform governance, conversational agent design, and the structural monitoring of AI-mediated public discourse.

Algorithmically ranked feeds allocate attention in ways that can reproduce inequalities in visibility. We test whether a compact, feed-level process disclosure —a banner describing exposure-safeguard governance, paired with provenance cues and in-situ user-choice affordances—improves perceived procedural justice, distributive fairness, and legitimacy among short-video creators in mainland China. In a mobile-first online experiment with 1,154 research-verified PRC creators in the People’s Republic of China (PRC), participants viewed a realistic feed randomized to No label , Personalization notice , or Fairness disclosure ; within the fairness condition, we factorially varied frame (Safeguards/Parity/Quotas/Exposure-fairness), issuer provenance (platform vs. independent audit), and disclosure depth (minimal badge vs. brief rationale). The compound fairness disclosure increased perceived procedural justice, distributive fairness, and legitimacy relative to both controls and was accompanied by lower self-censorship and higher voice intentions. Effects were directionally stronger in civic/news contexts and among creators who self-identified as marginalized or previously down-ranked. Within-arm exploratory diagnostics suggested that quota-like parity framing may yield directionally lower legitimacy than exposure-fairness framing, while audit cues produced directionally positive but imprecise trust gains. Backfire checks found no implied-unfairness spillover for unlabeled items and no interaction with per-item AI-content labels, although AI labels reduced perceived accuracy. Together, these findings suggest that governance-aligned micro-disclosures can function as procedural visibility signals for creators, while underscoring the importance of verifiable claims and careful disclosure design.

Emotional requirements (ERs)---how users should feel and which emotional harms a system must avoid---often determine whether people adopt and keep using software, especially in sensitive domains. Yet, eliciting ERs via interviews, workshops, and focus groups is costly and hard to scale or repeat. We study whether simulated focus groups, moderated discussions among user personas generated and played by large language models (LLMs), can augment early ER elicitation. Across three domains (mental‑health journaling, personal finance, and fitness), we compare one‑shot and iterative single‑model pipelines with iterative pipelines that use two or three different LLMs, and we benchmark against two human baselines: a human focus group and Emotional Goal Modeling (EGM). Multi‑LLM pipelines generate more diverse personas and increase the share of AI‑only ERs (not seen in our human baselines) that source‑blinded raters judge relevant by 14.7 percentage points compared to an iterative single‑model workflow; same‑model controls suggest this gain is not solely due to provider differences. Compared to human methods, simulations contribute more innovative requirements, EGM contributes clearer and more feasible ones, and human focus groups provide more natural phrasing. Overall, the results support an augmentation‑oriented workflow: use multi‑LLM simulation to broaden candidate ERs, structured modeling to organize them, and human engagement to ground decisions.

This study examines whether engagement with social robots translates into improved human-directed social abilities in autistic children. We conducted an 8-week home-based randomized controlled trial with 40 children aged 5–9 using a commercial social robot (Qrobot). Families were assigned to either continued robot access or robot withdrawal. Quantitative measures and caregiver interviews assessed anxiety, social motivation, emotion inference, and empathy. Results showed that continued robot access significantly reduced anxiety, confirming strong affective benefits and high usability. However, children in the withdrawal group demonstrated greater improvements in social motivation, emotion understanding, and empathic behaviors toward caregivers and peers. Qualitative findings revealed a “handoff versus siloing” pattern: withdrawal promoted reorientation toward human social interaction, while continued access concentrated engagement within the child–robot dyad and limited transfer to real-world contexts. We interpret these results as evidence that high engagement does not guarantee social transfer.

We present a textless, compute-light pipeline for user-level mental-health screening on social media that combines salted hashed character $n$ -grams, local differential privacy (LDP), and large language model (LLM)-teacher concept distillation into a tiny, calibrated logistic classifier. The encoder stores no vocabulary or raw text; instead, all posts for a user are aggregated on device into a salted hashed presence vector, and a subsampled randomized-response mechanism is applied once per user to provide an explicit user-level LDP guarantee for this representation. To recover interpretability without online LLMs, we invoke frozen LLMs (one for Chinese and one for English) on a small, stratified subset to label clinically grounded concepts (e.g., hopelessness, self-blame, sleep disturbance, and help-seeking); compact students are then trained on hashed features to predict these concepts across the full corpus. On two large Weibo user-level corpora, our Hashed + differential privacy (DP ) + Concepts model achieves calibrated utility competitive with a dense-embedding baseline while retaining a CPU-only footprint. Relative to a hashed baseline without concepts, we observe consistent gains in macro-AUPRC and substantially lower calibration error after isotonic mapping. Privacy audits show that, at our default user-level privacy budget, LDP plus salt rotation reduces user-level membership-inference attack success, resists reconstruction compared to embedding inversion, and substantially lowers cross-deployment linkage area under the curve (AUC) toward chance in our empirical setting. We release code and audit scripts to facilitate responsible reuse.

In intimate messaging, how a difficult note is produced signals personal effort and authorship. We examine how AI assistance level (light tone rewrite versus heavy full draft) and a brief sender‑voiced co‑sign disclosure shape receiver attributions and outcomes in two scenarios: apology/repair and boundary requests. Study 1 (N = 152) instrumented authoring to create a curated message corpus with authoring telemetry. Study 2 (N = 704) used those messages in a randomized 2 × 2 receiver experiment (Help by Disclosure) with crossed random effects. Across scenarios, heavier drafting reliably reduced perceived ownership and authenticity; improvements in competence or clarity did not compensate. In apologies, co‑signing a tone rewrite increased authenticity and forgiveness; co‑signing a full draft slightly decreased both. In boundary requests, co‑signing was neutral to mildly negative. Stimulus‑level analyses linked concrete authoring traces to receiver judgments: more idiosyncratic “voice” cues and higher human contribution predicted higher ownership and authenticity; larger drafting distance predicted the reverse. Senders also overestimated the relational benefits of light help in boundary requests relative to receivers. We contribute: (i) scenario‑aware causal estimates of help and disclosure on ownership, authenticity, trust facets, and outcomes; (ii) an empirically grounded attributional account that clarifies why competence gains rarely offset authenticity losses; (iii) evidence of sender–receiver miscalibration under realistic disclosure mixtures; and (iv) scenario‑sensitive design guidance for voice‑preserving defaults, ownership‑restoring scaffolds, and CPM‑aligned disclosure.

Digital eco‑nudges are widely deployed, yet their long‑term efficacy, ethical acceptability, and net environmental impact remain unclear. We report two field studies targeting routine online behaviors under strict parity of message content and delivery budgets. Study 1 shows that minimal, factual tailoring improves compliance over generic prompts when opportunities are defined independently of delivery. Study 2 introduces a privacy‑preserving, on‑device contextual bandit that learns when to act and when to DoNothing, achieving higher compliance at comparable prompt intensity while maintaining autonomy. We operationalize an Ethical–Efficacy Frontier (EEF) to visualize autonomy–effectiveness trade‑offs, and compute an energy Return on Investment (ROI) that nets behavior‑driven savings against measured system overhead. Energy savings are estimated using literature‑calibrated proxies with sensitivity bands, and we discuss the energy trade‑offs of on‑device learning relative to a stylized cloud alternative. We probe short‑term persistence via withdrawal and a brief follow‑up; long‑term habit formation and rebound remain out of scope. We contribute design and reporting practices—ablation parity, opportunity denominators, EEF, and net‑impact accounting—that make digital sustainability interventions more rigorous, transparent, and respectful, advancing sustainable HCI beyond “small changes.”