{"id":"7f7674b2-43eb-4e1e-b083-1e8965bbc2ca","arxiv_id":"2607.02900","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Anti-misinformation COVID-19 tweets are modestly but consistently more angry, disgusted, and sad than pro-misinformation tweets and come from more established users.","lead":"Posts that push back against COVID-19 false claims on Twitter score higher on anger, disgust, and sadness than posts that support those claims, and tend to come from older, more followed accounts. The finding challenges the common moderation heuristic that negative emotion marks falsehood and matters for how platforms treat corrective speech.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Classifier error could invent the emotion gap if false-claim-supporting tweets are systematically mislabeled as opposing when they are angry.","rationale":"The reader correctly isolates residual NLI error as the weakest assumption that could reverse or invent the group-level emotion differences. That premise is load-bearing: every downstream comparison (Table 2, Figure 3, SHAP) inherits the labels produced in §3.2.2. The paper is transparent about imperfect F1 and lists classifier noise as a limitation, yet supplies no direct check that the emotion gap survives human re-labeling or is orthogonal to the model’s confusion patterns. Because the reported effects are modest, even modest systematic bias could erase or reverse them. No stronger internal contradiction appears; the observational design and modest claims are otherwise coherent. The concrete human-validation test would settle whether the concern lands without requiring new data collection beyond the existing claim–tweet pairs. Verdict therefore remains CONDITIONAL, matching the reader.","tokens_in":11854,"tokens_out":568,"duration_ms":5469,"concrete_test":"Draw a stratified random sample of ~1,000 claim–tweet pairs from the final analytical set (balanced support/oppose), obtain independent human stance labels, recompute Cliff’s δ for anger/disgust/sadness on the human-labeled subset, and also recompute δ after restricting to model-correct cases only. If the human-labeled δ reverses sign, drops below |δ|≈0.03, or is driven by the model-error cells, the headline emotion finding does not hold under cleaner labels.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (anti-misinformation tweets show higher anger/disgust/sadness, Cliff’s δ ≈ −0.10 to −0.125) rests on labels from an imperfect NLI model (F1_Con = 0.745, F1_Ent = 0.779; §3.2.1–3.2.2). The paper filters neutrals and ambiguous cross-claim cases but never reports a confusion-matrix-conditioned emotion analysis or a human-validated subsample of the final 264k tweets. If the model preferentially routes emotionally negative supporting tweets into the oppose class (or calm opposing tweets into support), the observed group difference is an artifact of labeling rather than a real stance–emotion association. Prior small-scale studies are cited as corroboration, yet they used different coding schemes; residual systematic error remains the least-secured premise for the scaled claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper profiles the counter-misinformation ecosystem on Twitter during COVID-19 by aggregating 15,374 fact-checked false claims, retrieving candidate tweets via BM25, and labeling claim–tweet pairs with a domain-specific NLI model from the authors’ prior work, yielding 264,737 tweets (195,922 supporting and 68,815 opposing false claims). It then compares 23 user- and text-level features—including Demux Ekman emotion scores, DeToxify toxicity, Botometer scores, and account metadata—using Cliff’s δ, distributional plots, and a weak feature-only classifier with SHAP. The central claim is that anti-misinformation posts are more emotionally negative than pro-misinformation posts (higher sadness, anger, and disgust; |δ|≈0.10–0.125), contrary to the common view that negative affect is a signature of falsehood, and that opposing posts tend to come from more established users (older accounts, more followers, higher listed counts). Bot scores do not separate the groups; feature-only models achieve only modest above-chance separation (RF F1≈0.632).","tokens_in":12038,"tokens_out":1424,"duration_ms":26067,"significance":"If the stance–emotion association survives tighter label validation, the result is a useful corrective to both research and platform practice: negative emotion is not a reliable proxy for falsehood, and moderation that downranks emotionally charged content risks suppressing organic correction. Strengths include scale across a broad claim set, discrete-emotion measurement rather than dictionary coding, honest reporting of modest effect sizes and weak feature-only separability, and extension of smaller prior observations on COVID-19 and crisis correction. The NLI-plus-profiling pipeline is a reusable template for studying the response side of misinformation at scale. The contribution is empirical and comparative rather than theoretical; its value hinges on whether residual labeling error can be shown not to invent the reported group differences.","major_comments":[{"comment":"§3.2.1–3.2.2, Table 2, Figure 3: The central emotion claim (higher sadness/anger/disgust in anti-misinformation tweets; Cliff’s δ ≈ −0.125/−0.098/−0.097) rests entirely on NLI labels with F1_Con = 0.745 and F1_Ent = 0.779. Neutrals and cross-claim ambiguities are filtered, but the paper reports no human-validated subsample of the final 264,737 tweets and no sensitivity analysis under systematic misclassification (e.g., angry supporting tweets preferentially labeled as opposing). Because the reported |δ| values are modest, even moderate label bias could invent or reverse the pattern. A stratified human audit of the analytical set (or confusion-matrix-conditioned emotion distributions on held-out labeled pairs) is needed before the scaled claim can be treated as secure.","section":"§3.2.1–3.2.2, Table 2, Figure 3"},{"comment":"§3.2.2, §4.2, Table 2, Figure 4: The sample comprises 264,737 tweets from 201,334 users, so user-level features (followers, listed count, account age, activity) are repeated across a user’s tweets. Cliff’s δ and the SHAP analysis treat tweets as independent observations. Dependence can inflate the apparent separation on user attributes that are constant within user. Report user-level aggregates (one observation per user, or per-user mean stance) or cluster-robust comparisons so that the “more established users” claim is not an artifact of tweet multiplicity.","section":"§3.2.2, §4.2, Table 2"},{"comment":"§3.1.2, §4.1: Candidate tweets are the BM25 top-100 per claim. Lexical retrieval can under-sample paraphrastic corrections and over-sample posts that restate claim wording, which may systematically shape emotion and length profiles (e.g., terse “that’s false” vs. elaborate supporting narratives). A brief retrieval-sensitivity check—e.g., comparing emotion δ on high- vs. low-BM25 ranks, or on a denser retrieval slice—would show whether the angry-corrector pattern is robust to the retrieval design.","section":"§3.1.2, §4.1"}],"minor_comments":[{"comment":"Table 2 reports only features with |δ|>0.05. A full appendix table of all 23 features (including toxicity and bot scores) would let readers see null results and avoid selective emphasis.","section":"Table 2"},{"comment":"Toxicity is measured (§3.2.3) but barely discussed in Results; either report the δ values or state explicitly that toxicity did not differentiate the groups.","section":"§3.2.3, §4"},{"comment":"Figure 3 KDEs would be clearer with a shared x-axis scale and a note on whether scores are calibrated probabilities or raw model outputs.","section":"Figure 3"},{"comment":"§4.3: Undersampling the majority class for the feature-only classifiers is fine for balanced accuracy, but report the original class prior and, if possible, PR-AUC on the natural imbalance so readers can judge operational utility.","section":"§4.3, Table 3"},{"comment":"The term “counter-misinformation ecosystem” is useful but introduced without a precise operational definition; a one-sentence definition in §1 would help.","section":"§1"},{"comment":"Minor: “listed more often” / “higher listed counts” is clear to Twitter researchers but could be glossed once for a broader Hypertext/Web audience.","section":"Abstract, §4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript states it is already accepted to ACM Hypertext 2026; if this review is for a journal track or special issue, scope fit is fine for cs.SI / computational social science, but novelty relative to the authors’ own NLI/CMG pipeline paper [11] should be framed as the profiling contribution rather than the detector. The labeling-validity concern is the main reason I chose major_revision rather than minor_revision; if the authors can supply even a modest human audit of the final set, the paper would be in good shape."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this is a careful observational scale-up, not a paradigm shift. Prior small-topic work already saw angrier COVID rebuttals and anger in crisis corrections. Choi and Ferrara take a broad claim set, run a domain NLI model on ~1.5M BM25 pairs, land on 265k support/oppose tweets, and show anti-misinfo posts higher on sadness, anger, and disgust (Cliff’s δ ≈ −0.10 to −0.125) plus older, more-followed, more-listed authors. Effects are modest and they say so. That honesty is a strength.\n\nWhat is new is the combination of scale, discrete-emotion scores (Demux/Ekman), joint user-attribute profiling, and a weak feature-only classifier with SHAP that still sits above chance. The pipeline is coherent: claim aggregation, retrieval, labeling, nonparametric comparison. Emotion/toxicity/bot tools are independent of the stance labels, so the comparison is not circular by construction. Citations look appropriate; self-citation of their prior NLI/CMG work is measurement reuse, not result recycling. Math and reporting are straightforward.\n\nSoft spots, in proportion: the load-bearing assumption is residual NLI error (F1_Con 0.745, F1_Ent 0.779). They filter neutrals and cross-claim ambiguities but never show a confusion-matrix-conditioned emotion check or a human-validated subsample of the final 264k. The stress-test worry—that angry supporting tweets get routed into oppose—is real and unclosed; it does not kill the claim, because smaller prior studies point the same way, but it keeps the result conditional. No code/data release, COVID/Twitter 2020–21 only, and feature-only separability is weak (RF F1 ~0.63). Those are ordinary limits for this genre, not hidden flaws.\n\nThis is for people who design moderation proxies or study organic correction. It undercuts the lazy “negative emotion = falsehood” heuristic and is worth a serious referee. I would engage, cite the distributional finding with the classifier caveat, and not treat it as settled beyond COVID Twitter.","headline":"Solid scale-up of a known pattern: anti-misinfo COVID tweets are modestly angrier/sadder/disgusted and come from more established accounts; the result is useful but rests on an imperfect NLI labeler without a human-checked subsample.","tokens_in":12680,"tokens_out":583,"would_cite":true,"duration_ms":6995,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Posts that oppose COVID-19 misinformation on Twitter are angrier, more disgusted, and sadder than posts that spread it.","keywords":["counter-misinformation","COVID-19","Twitter","stance detection","natural language inference","emotion","anger","social correction"],"falsifier":"Re-label a large random sample of the same claim–tweet pairs with independent human coders or a substantially more accurate stance model; if the anger/disgust/sadness elevation and the established-user pattern disappear or reverse under the cleaner labels, the central claim fails.","tokens_in":12709,"feed_emoji":"😠","tokens_out":880,"duration_ms":8665,"temperature":0.7,"pith_summary":"This paper asks who pushes back against false claims on social media and how they sound when they do. Using a domain-specific natural-language-inference classifier on a large COVID-19 Twitter corpus, the authors label hundreds of thousands of posts as supporting or opposing fact-checked false claims, then compare emotion, toxicity, bot scores, and account attributes across the two groups. The central finding is that corrective posts are more emotionally negative—higher in anger, disgust, and sadness—than the posts that spread the false claims, and that the people doing the correcting tend to hold more established accounts. The result challenges the common assumption that negative emotion is mainly a signature of falsehood rather than of correction, and it implies that moderation rules keyed to emotional intensity can suppress the very discourse that contests misinformation.","feed_headline":"Corrective tweets are angrier than the falsehoods they fight","feed_subtitle":"On COVID-19 Twitter, anti-misinformation posts show more anger, disgust, and sadness—and come from more established users.","key_machinery":"A domain-specific natural-language-inference (NLI) classifier, fine-tuned on COVID-19 misinformation data and synthetic pairs, that labels claim–tweet pairs as supporting, opposing, or neutral; after filtering, it yields 264,737 stance-labeled tweets whose user- and text-level features are then compared with Cliff’s delta and SHAP-interpreted classifiers.","core_discovery":"At scale on COVID-19 Twitter, posts that oppose fact-checked false claims carry higher levels of anger, disgust, and sadness than posts that support those claims. The differences are modest in size but consistent in direction across the negative emotions. The opposing posts also tend to come from more established users—older accounts, more followers, higher listed counts—while bot scores do not cleanly separate the two sides.","pith_inferences":["If the pattern holds beyond COVID-19, emotion-based downranking could systematically weaken organic correction on political and climate topics as well.","The length and surprise elevation on the pro-misinformation side may reflect narrative packaging of pseudo-evidence; testing whether shorter corrections are more persuasive would be a direct follow-on experiment.","Demographic or partisan covariates, unobserved here, may partly drive the ‘established user’ pattern; matching on those covariates would test whether establishment itself predicts corrective behavior."],"forward_implications":["Content-moderation systems that flag negative emotion as a proxy for misinformation risk downranking or silencing legitimate corrections.","Negative affect is not a reliable truth signal: corrective discourse can be angrier than the falsehoods it targets.","Counter-misinformation activity is disproportionately carried by more established accounts rather than by bots or brand-new users.","User- and text-level features alone carry a weak but real signal for separating support from oppose, even without tweet semantics.","Platforms that formalize community fact-checking should expect the corrective population to sound frustrated rather than neutral."],"fun_headline_variants":["Anti-misinfo COVID tweets show more anger and disgust","Fact-checkers on Twitter post angrier than false claims","Opposing misinformation brings higher anger, older accounts","Corrective posts carry more sadness and more followers","Anti-falsehood tweets: angrier, sadder, from established users"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That residual errors from the stance classifier and the filtering of neutral or ambiguous cases do not reverse or invent the group-level differences in emotion and account attributes between support and oppose posts.","fun_headline_variants_meta":{"raw":{"variants":["Anti-misinfo COVID tweets show more anger and disgust","Fact-checkers on Twitter post angrier than false claims","Opposing misinformation brings higher anger, older accounts","Corrective posts carry more sadness and more followers","Anti-falsehood tweets: angrier, sadder, from established users"]},"model":"grok-4.5","effort":"low","cost_usd":0.00257,"raw_usage":{"total_tokens":947,"prompt_tokens":730,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":25700000,"prompt_tokens_details":{"text_tokens":730,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":133,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":730,"tokens_out":84,"duration_ms":12657,"temperature":1.0,"reasoning_tokens":133,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T06:15:45.717964+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-label a large random sample of the same claim–tweet pairs with independent human coders or a substantially more accurate stance model; if the anger/disgust/sadness elevation and the established-user pattern disappear or reverse under the cleaner labels, the central claim fails.","supporting_citations":[],"review_version":1}