{"id":"1e944597-9c5e-4a15-8164-e91547f0e37e","arxiv_id":"2505.02317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Bluesky users after public launch are predominantly left-leaning, share high-credibility sources, exhibit low toxicity, and saw a Japanese-language surge and some spam-like accounts.","lead":"This paper traces Bluesky's first two months after its public opening using 114 million Firehose events. It finds the platform is mostly left-leaning, shares high-credibility news, has low toxicity, and saw a Japanese-language surge plus some likely spam accounts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 74.61% left-leaning claim may not generalize: the denominator of users with ≥5 MBFC-rated links is unreported and could be a small, unrepresentative subset.","rationale":"The reader's weakest assumption correctly identifies that shared domains may not reflect personal ideology and that the 5-link subsample may not represent the whole population. My concern sharpens this: the paper omits the sample size entirely, making it impossible to judge representativeness from the reported statistics. The 74.61% figure is central to the paper's characterization of Bluesky as a left-leaning platform, and it is repeated in the Discussion as a general statement about 'Bluesky users.' Because the paper already frames itself as descriptive and the reader's verdict is CONDITIONAL, my read does not require a harsher verdict; it reinforces the need for an explicit coverage analysis and sensitivity checks. I mark agreement as 'partial' because I go beyond the reader's general proxy concern to a specific, checkable denominator and threshold-sensitivity problem.","tokens_in":12775,"tokens_out":2655,"duration_ms":34830,"concrete_test":"Compute N = number of users with at least 5 MBFC-rated link posts and report N divided by total active users. Then recompute the distribution of mean bias scores and the 74.61% figure under three variations: (a) threshold lowered to 1 rated link; (b) user-level weighting by number of rated posts instead of unweighted averaging; (c) excluding users whose rated domains are all classified 'least biased' (score 0). If the left-leaning majority persists with N at least, say, 20% of active users and across (a)-(c), the claim is robust; if N is below 5% or the majority disappears under (a) or (c), the claim should be explicitly restricted to active news-sharers.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4.4 defines 'active users' for the political-leaning analysis as users who shared at least five posts linking to MBFC-rated websites, and reports that 74.61% of this sample have a negative mean bias score. The paper never reports how many users satisfy this criterion. Given that only ~850k of ~6.2M URL-containing posts (13.6%) have MBFC ratings, and that many users share no news links at all, the denominator could be a small, self-selected group of heavy news sharers. If so, the Discussion's stronger statement that 'Bluesky users lean predominantly left-wing' extends the finding beyond the actually measured subpopulation. A second issue is the averaging procedure itself: each rated domain contributes equally, and apolitical or 'least biased' domains contribute 0, so users who mostly share non-political content are labelled centrist, while one extreme domain among several mainstream ones can move a user left. Without reporting the sample size, the coverage rate relative to all active users, or robustness to the 5-link threshold, the platform-wide majority claim is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes 56 days of Bluesky Firehose data (January 9 to March 4, 2024) surrounding the platform's public launch, examining user activity, language use, follower-network structure, political leaning, source credibility, toxicity, communities, and moderation. The main descriptive findings are that Bluesky shows a heavy-tailed activity distribution with a high share of original posts, low overall toxicity, a left-leaning user population, very low prevalence of low-credibility content concentrated among a few 'superspreaders,' and a small fraction of accounts that were later moderated or deleted. The paper interprets these observations as evidence that Bluesky's moderation efforts have been effective and that the platform fosters a different engagement pattern than centralized counterparts.","tokens_in":13051,"tokens_out":3793,"duration_ms":44894,"significance":"If the claims hold, this is one of the first large-scale longitudinal portraits of Bluesky's early public phase and a useful descriptive baseline for decentralized social media. The study's strengths include a large Firehose-based dataset, use of external benchmarks (NewsGuard, Media Bias/Fact Check, Detoxify), explicit discussion of limitations, and a transparent account of data collection. The central claims about low toxicity and high-credibility sharing are well supported by the presented measurements. However, the headline political-leaning claim and the moderation-effectiveness inference rest on assumptions about sample representativeness and temporal interpretation that need additional support before the paper's broader conclusions can be accepted.","major_comments":[{"comment":"The manuscript reports that 74.61% of 'active users' are left-leaning, where active users are defined as users who shared at least five posts linking to MBFC-rated websites, but it never reports the number of users in this denominator or its coverage relative to all active users on the platform. Given that only 13.6% of URL-containing posts have MBFC ratings and many users share no news links at all, the denominator could be a small, self-selected group of heavy news sharers. The Discussion's statement that 'Bluesky users lean predominantly left-wing' (Section 5) therefore extends the finding beyond the actually measured subpopulation. Please report the denominator, the coverage rate relative to all active users, and the sensitivity of the 74.61% / 18.17% / 7.21% split to the five-link threshold and to the equal-weight averaging of domains.","section":"§4.4, Figure 5, and §5"},{"comment":"The conclusion that Bluesky's 'moderation efforts have been effective' is based on account statuses queried in November 2024, approximately eight months after the observation window. These statuses reflect cumulative enforcement actions and cannot by themselves establish that moderation during or shortly after the study period was effective; they could also reflect later actions or deletions unrelated to the behaviors analyzed. The paper also does not provide a baseline or a discussion of detection bias for the 0.5% moderation rate. Please either temper the effectiveness claim to what the data can support (e.g., 'a small fraction of accounts in our cohort were subsequently moderated or deleted') or add an analysis of time-to-action for the flagged mass-following and low-credibility-sharing accounts.","section":"§4.9, Figure 11, and Abstract"}],"minor_comments":[{"comment":"The Mann-Whitney U test comparing pre- and post-opening political leaning is reported only as p = 0.05; please report the exact p-value and state whether any multiple-comparison correction was applied.","section":"§4.4"},{"comment":"There is a typo in 'While the the distributions of toxicity scores'; also, the statement that no threshold was used in §4.6 should be reconciled with the 0.5 toxicity threshold introduced in §4.8.","section":"§4.6"},{"comment":"The sentence 'Fig. 11A suggests that moderation was done against users who violated the terms of service, as users who shared links from low-credibility sources remain' is unclear and appears to contradict the figure description; please rephrase to state what the comparison of account-status groups actually shows.","section":"§4.9, Figure 11"},{"comment":"The text contains a LaTeX formatting artifact: 'AccountDeactivated-emphAccountTakedown' should be 'AccountDeactivated–AccountTakedown' and similarly for the next pair.","section":"§4.9"},{"comment":"The table reports density values such as '4 .4× 10−6' with an unusual space; please format consistently and define whether density is computed on directed or undirected edges.","section":"§4.3 and Table 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is one of the first solid longitudinal looks at Bluesky's public launch, and the new empirical material—the Japanese-language surge, source-credibility trends, resharing communities, and account-status moderation check—makes it worth a referee's time. The paper is honest about its methods and limitations, and the descriptive claims are backed by a large firehose dataset. But the headline political-leaning result is softer than the abstract suggests. The 74.61% figure is computed over an unreported subset of users who shared at least five links to MBFC-rated domains. The stress-test note lands: we never learn how many users that is, or how representative they are of all active users. The averaging scheme also treats each domain equally and scores apolitical sites as zero, so heavy non-news sharers get labeled centrist. The Discussion's platform-wide phrasing overreaches.\n\nOther soft spots are minor. The moderation-effectiveness claim is inferred from account statuses fetched eight months later; it is plausible but not directly causal. The toxicity analysis is restricted to Detoxify languages, which the paper acknowledges. The 0.5 toxicity threshold and the 5-link minimum are arbitrary but standard for this literature.\n\nWhat is genuinely new: the before/after comparison around Feb 6, 2024, the Japanese influx, the finding that the five largest resharing communities cover 87% of users, and the low-credibility superspreader pattern on a fresh platform. These are measurable, reported, and useful for anyone studying decentralized platforms. The self-citation to the earlier workshop paper is appropriate; this is a substantial extension, not a republication.\n\nMy verdict: conditional accept after the authors report the denominator, show robustness to the 5-link threshold and the averaging choice, and soften the Discussion's wording. That is a fixable revision, not a fatal flaw. I would send it out.","headline":"Solid descriptive study of Bluesky's launch with genuinely new empirical findings, but the headline left-leaning majority claim rests on an unreported subset of users and needs a denominator and robustness checks before it can be taken at face value.","tokens_in":13528,"tokens_out":1771,"would_cite":true,"duration_ms":20684,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Bluesky after its public launch is predominantly left-leaning, low-toxicity, and dominated by original posts, with only 0.08% of posts linking to low-credibility sources.","keywords":["Bluesky","decentralized social media","misinformation","political leaning","toxicity","content moderation","longitudinal analysis","source credibility"],"falsifier":"Take the users classified as left-leaning by their shared domains and compare their leaning with an independent signal, such as the partisan slant of their original post text or their self-declared ideology in bios; if the two measures disagree for a substantial share of users, the 74.61% left-majority claim fails. A simpler check: recompute the user-leaning distribution using only users who wrote original posts with explicit political hashtags, and see whether the left share remains above 70%.","tokens_in":12649,"feed_emoji":"📊","tokens_out":9998,"duration_ms":101946,"temperature":0.7,"pith_summary":"The paper sets out to measure what happened to Bluesky when it opened to the public on February 6, 2024, using 56 days of activity data covering 114 million events. It argues that the platform's early community was predominantly left-leaning: 74.61% of users with at least five rated links are classified as left-of-center, while 7.21% are right-of-center. It also argues that misinformation was rare, with low-credibility links in 0.08% of all posts, and that toxicity stayed low across the languages the authors could score. A sympathetic reader would care because these are among the first large-scale measurements of whether a decentralized, user-controlled platform can avoid the misinformation and toxicity problems of centralized networks. The paper further claims that suspicious mass-following accounts appeared at launch, and that many were quickly suspended or flagged, suggesting moderation can work in a decentralized setting.","feed_headline":"Bluesky skews left and shares little misinformation","feed_subtitle":"A 56-day trace of 114M activities finds mostly original posts and quick moderation of suspicious accounts.","key_machinery":"The analysis is carried by three measurement instruments applied to the public Firehose event stream. First, every shared domain is labeled with a credibility score and a political-bias score from professional media-rating services; a user's political leaning is the average bias score of the domains they shared, computed only for users who shared at least five rated links. Second, a toxicity classifier scores posts and per-user toxicity in seven languages. Third, a directed reshare network is partitioned with a standard community-detection algorithm to identify the five largest communities, which contain 87% of users. These instruments connect raw activity data to the paper's claims: the credibility labels produce the 0.08% low-credibility figure and the superspreader concentration, the bias labels produce the 74.61% left-leaning majority, and the community partition explains toxicity differences by language and content type.","core_discovery":"The central claim is that Bluesky's first two months as a public platform looked different from mainstream social media in measurable ways, while still reproducing familiar concentration patterns. Using the Firehose stream, the authors find that original posts outnumber reposts, that the follower network's largest strongly connected component more than tripled after the launch, and that users predominantly shared sources rated high in credibility. The headline statistic is the political leaning of the active user base: averaging the political bias scores of the domains each user shared yields 74.61% left-leaning, 18.17% centrist, and 7.21% right-leaning users, with no significant shift before versus after the opening. Misinformation is concentrated in a handful of accounts, with ten users responsible for 62% of low-credibility links, and the authors interpret the quick suspension of several mass-following accounts as evidence that platform moderation was functioning. The overall picture the paper argues for is a left-skewed, comparatively civil, mostly original-content platform whose main risks are manipulation attempts rather than endemic toxicity.","pith_inferences":["The 74.61% figure may overstate the population's left lean if users who share more links are also more left-leaning; the paper only measures users with at least five rated links, and link-sharing itself is behavior, not identity.","If moderation decisions correlate with mass-following rather than content, the paper's account-status data could be re-analyzed to test whether suspension was predicted by follow rate, helping future platforms design anti-spam rules.","Extending the same pipeline to Bluesky after the 2024 U.S. election and the Brazil migration wave would test whether the left-leaning, low-toxicity profile is stable or a launch-window artifact.","Because Japanese posts became 44% of content but toxicity was scored only in seven languages, the overall low-toxicity result might change if Japanese text were scored."],"forward_implications":["If the early signal persists, Bluesky is a low-misinformation environment: one in roughly 1,250 posts contains a low-credibility link, and high-credibility domains are shared roughly 125 times more often per day.","Because only 0.5% of accounts were moderated by November 2024 and most suspicious actors at launch were caught quickly, decentralized moderation can apparently contain coordinated follow-spam without heavy-handed takedowns.","A left-skewed user base means Bluesky may become an ideologically homogeneous space, limiting cross-partisan exposure.","The concentration of low-credibility sharing in ten accounts implies that removing or demoting a tiny number of superspreaders would cut most misinformation on the platform.","Language is a primary organizing axis: Japanese speakers became the largest community after launch, so platform-level statistics may be driven by specific linguistic sub-communities rather than a global culture."],"supporting_citations":[{"why":"Provides the Firehose endpoint used to collect all posts, follows, likes, and blocks in real time.","marker":"[21]"},{"why":"Establishes the source-credibility labeling approach that grounds the misinformation analysis.","marker":"[26]"},{"why":"Supplies the threshold convention that defines low-credibility and mainstream news sources.","marker":"[29]"},{"why":"Supplies the method of averaging source bias scores to estimate a user's political leaning.","marker":"[32]"},{"why":"Provides the community-detection algorithm used to partition the reshare network into the five largest communities.","marker":"[33]"},{"why":"Provides the toxicity classifier used to score posts and per-user toxicity across languages.","marker":"[40]"},{"why":"Introduces the superspreader measurement that the paper applies to low-credibility sharing.","marker":"[37]"}],"fun_headline_variants":["Bluesky users skew left and share credible sources","Bluesky: original posts dominate, toxicity low","Bluesky's early data: left-leaning, civil, moderation works","Bluesky: 74% left, 7% right, little misinformation","Bluesky launch: spam accounts suspended quickly, left lean persists"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the average political-bias score of the news domains a user shares is a valid measure of that user's own political leaning, and that users who shared at least five rated links represent the whole Bluesky population; if shared domains do not track personal ideology, the 74.61% left-leaning majority is an artifact of the measurement.","fun_headline_variants_meta":{"raw":{"variants":["Bluesky users skew left and share credible sources","Bluesky: original posts dominate, toxicity low","Bluesky's early data: left-leaning, civil, moderation works","Bluesky: 74% left, 7% right, little misinformation","Bluesky launch: spam accounts suspended quickly, left lean persists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000308,"raw_usage":{"total_tokens":1775,"prompt_tokens":971,"completion_tokens":804,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":714}},"tokens_in":587,"tokens_out":804,"duration_ms":9811,"temperature":1.0,"reasoning_tokens":714,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:54:29.374424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the users classified as left-leaning by their shared domains and compare their leaning with an independent signal, such as the partisan slant of their original post text or their self-declared ideology in bios; if the two measures disagree for a substantial share of users, the 74.61% left-majority claim fails. A simpler check: recompute the user-leaning distribution using only users who wrote original posts with explicit political hashtags, and see whether the left share remains above 70%.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the source-credibility labeling approach that grounds the misinformation analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Firehose endpoint used to collect all posts, follows, likes, and blocks in real time."},{"cited_title":"Nogara, F","cited_arxiv_id":null,"evidence_quote":"Supplies the threshold convention that defines low-credibility and mainstream news sources."},{"cited_title":"Hanu, Unitary team, Detoxify, Github","cited_arxiv_id":null,"evidence_quote":"Provides the toxicity classifier used to score posts and per-user toxicity across languages."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the superspreader measurement that the paper applies to low-credibility sharing."}],"review_version":1}