{"id":"08a102ff-e42c-4fce-b54c-112e31ac676c","arxiv_id":"2601.09117","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A 14-month audit of 4,847 paid AI-content requests on Civitai shows NSFW commissions growing to a majority of weekly bounties, deepfake requests targeting women about 9:1 among real individuals, and the platform's deepfake warning notice missing from 41.7% of NSFW deepfake bounties.","lead":"This paper measures what people pay for on Civitai, a large AI-image platform: it scraped 14 months of public 'bounty' requests and found adult (NSFW) requests growing to a majority of weekly postings, plus 323 requests to generate images of real people — most targeting female celebrities. It matters because it quantifies how a monetized, community-run AI marketplace can produce gendered harms that platform rules do not reliably catch.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NSFW majority and growth trend are contingent on Civitai's own 'R' threshold; external labels show only κ=0.52, so the central claim needs a sensitivity check before it is established.","rationale":"The reader's weakest assumption identifies the same core issue: the NSFW trend depends on Civitai's image ratings plus GPT-4.1 text classification, and the paper's own cross-check shows only moderate agreement (κ=0.52). My review agrees that this is the most load-bearing concern because it directly undermines the quantitative foundation of the paper's main contribution—the claim that paid bounties increasingly incentivize explicit content. The paper's other limitations (missing code/data, small NSFW-deepfake cell, Figure 3(a) inconsistency, the contradictory AI-usage statement in Section 6) are real but secondary: they affect reproducibility and presentation, not the validity of the core measurement. The classification concern is singular because it is a correctness risk: if the platform's 'R' label includes non-sexual content such as graphic violence, the reported NSFW majority could be an artifact of Civitai's conservative rating threshold. The proposed concrete test—recomputing the trend under OpenAI sexual flags or excluding R-only bounties—would settle this. Since the reader already assigned a CONDITIONAL verdict, my analysis does not change that verdict; it reinforces it. I therefore recommend UNCHANGED, with the explicit condition that the authors perform the sensitivity analysis and report the results before the claims are taken as established.","tokens_in":13801,"tokens_out":7229,"duration_ms":74829,"concrete_test":"Recompute the full-sample NSFW share and the weekly trend in Figure 2(d) under two alternative labeling schemes: (i) OpenAI's 'sexual' flag as the NSFW label, and (ii) Civitai labels restricted to X/XXX, with R-only bounties reclassified as SFW. If the monotonic increase and the September-2024 weekly majority persist under both alternatives, the concern is resolved. If either alternative flattens or removes the trend, the abstract and Section 3.1 should be qualified to state that the 'majority' holds under Civitai's conservative content labels, not necessarily under a stricter definition of explicit sexual content.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claims—48% NSFW bounties and a steady increase to a weekly majority (Section 3.1, Fig. 2d)—rest entirely on the coding rule in Section 2.1: a bounty is NSFW if any example image carries Civitai's R/X/XXX label, with GPT-4.1 text classification applied only to the remaining image-SFW bounties. This makes the platform's own rating system the ground truth for the central trend. But the paper's external validation shows only moderate agreement with OpenAI moderation (κ=0.52, MCC=0.53, accuracy=0.77), and the only sensitivity reported—treating R-only bounties as SFW—improves agreement to κ=0.63. That R category includes 'adult themes, partial nudity, graphic violence,' not necessarily sexual content. Counting R as NSFW may therefore inflate the 'explicit' share and make the growth-to-majority appear larger or earlier than it is. The paper acknowledges that text-only requests can be missed, which could bias in the opposite direction, but the net effect is unknown. Because the governance argument (Sections 4.2 and 4.5) depends on the trajectory of explicit demand, the central claim is not yet robust to this classification choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a longitudinal observational study of Civitai's paid 'Bounties' marketplace over the platform's first 14 months (4,847 bounties, Oct 2023–Jan 2025). The authors scrape bounty pages, record platform-provided image content labels, use GPT-4.1 for thematic classification of text/image content, and manually verify all candidate deepfake requests. They report four main findings: (1) NSFW bounties are widespread and grew to a weekly majority by September 2024; (2) 323 bounties request deepfakes of real people, with an approximately 9:1 female-to-male target ratio; (3) bounty creation is concentrated, with deepfake requests somewhat more concentrated than SFW or NSFW requests; and (4) Civitai's deepfake information notice is absent from 41.7% of NSFW deepfake bounties versus 14.4% of SFW deepfake bounties. The paper interprets these results as evidence that monetized, community-driven generative-AI platforms can normalize and incentivize explicit and gendered deepfake content, and it discusses governance and enforcement implications.","tokens_in":13906,"tokens_out":6650,"duration_ms":60987,"significance":"If the empirical claims hold, this is a valuable and timely contribution to the study of generative-AI platforms, content moderation, and gender-based harm. The paper's strengths include its complete capture of publicly listed bounties over the study window, manual verification of all 323 deepfake labels, an explicit external validation of the content classification using OpenAI moderation (reported agreement statistics), and a clearly described annotation procedure. The descriptive claims about the bounty marketplace and the gendered distribution of deepfake targets are supported by the data and are falsifiable. However, the central NSFW-growth claim depends on the platform's own image-content labels, which show only moderate agreement with an independent moderator, and some comparative claims (Gini differences, enforcement-gap temporal stability) lack uncertainty quantification or temporal controls. These gaps are addressable and do not invalidate the overall contribution, but they currently prevent the stronger causal and governance conclusions from being fully established.","major_comments":[{"comment":"The main NSFW-growth result is defined using Civitai's image labels with R treated as NSFW. The external validation against OpenAI moderation yields only κ=0.52, and the paper's only sensitivity check—treating R-only bounties as SFW—improves agreement to κ=0.63 but is not propagated to the weekly trend. Since the abstract and Section 4.2 rely on the trend crossing 50% by September 2024, please recompute the weekly NSFW series under at least three definitions: (a) the current R-as-NSFW rule, (b) an R-as-SFW rule, and (c) OpenAI moderation flags alone; report the majority-crossing dates and bootstrapped confidence intervals for each. Without this analysis, the claim that 'NSFW requests now comprise a majority of weekly bounties' is not robust to the classification choice.","section":"Section 2.1, Fig. 2d"},{"comment":"The enforcement-gap comparison records the platform notice's presence at the time of data collection, not through the bounty's lifetime. Since the study spans 14 months and the platform may have introduced or expanded the notice feature during that period, older NSFW deepfake bounties could lack the alert for temporal reasons rather than because of inconsistent enforcement. The authors should report the posting-date distribution for marked versus unmarked deepfake bounties, or restrict the comparison to bounties created after the feature was demonstrably in place. At minimum, the paper should state that this is a cross-sectional snapshot with no temporal control; as written, the 41.7% versus 14.4% gap is not sufficient to establish an enforcement inconsistency.","section":"Section 3.4, Fig. 5"},{"comment":"The Gini coefficients are reported as point estimates without confidence intervals or significance tests. The differences among deepfake (0.45), SFW (0.41), and NSFW (0.37) are modest, and the corresponding top-20% shares (56%, 53%, 49%) overlap considerably. Because the deepfake category is a subset of the other two and the categories are not independent, the authors should provide bootstrap confidence intervals for each Gini and for the differences, or a permutation test against a null of equal concentration. Without this, the claim that deepfake requests are 'somewhat more concentrated' is not statistically supported.","section":"Section 3.3, Fig. 4"}],"minor_comments":[{"comment":"The text reports LoRA n=2,936 and image n=1,450, but Figure 2(a) shows 2,984 and 1,475. The figure's values sum to the stated total of 4,847; please correct the in-text numbers or the figure.","section":"Section 3.1 vs. Fig. 2(a)"},{"comment":"The generative AI usage statement says that 'no generative AI tools' were used in analysis, but Section 2.1 describes using GPT-4o and GPT-4.1 for thematic classification. The statement should be corrected to distinguish LLM-based annotation from writing assistance.","section":"Section 6"},{"comment":"The two panels of Figure 3 are difficult to parse: the axis labels and percentage annotations suggest one panel shows the NSFW subset and the other the overall distribution, but the caption does not state this. Please clarify which panel corresponds to which subset and align the reported counts with the displayed numbers.","section":"Figure 3"},{"comment":"The paper says 2,752 SFW bounties were used for theme extraction, but 4,847 total bounties with 43% NSFW implies roughly 2,763 SFW bounties. Please reconcile the discrepancy (e.g., bounties with no images or failed classification).","section":"Section 2.1"},{"comment":"The sample is described as 'all publicly listed bounties available at the time of collection.' Bounties removed or deleted before collection would be missing, which could bias the trend and enforcement analyses. Please state this limitation explicitly and, if possible, discuss the direction of the potential bias.","section":"Section 2, data collection"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good empirical contribution, and the central claims are likely correct in direction. The main concern is the lack of sensitivity analysis for the NSFW classification rule, which is load-bearing for the headline trend. The AI-disclosure inconsistency in Section 6 should also be fixed. I would support publication after the robustness checks are added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first systematic measurement of Civitai's paid bounty market, and the deepfake counts are human-verified, which puts it ahead of most platform-scrape papers. The headline claims are plausible but not yet load-bearing; the NSFW trend rests on Civitai's own rating labels, and a few internal inconsistencies need cleanup.\n\nWhat's new: prior work looked at images, models, prompts; this is the first look at the commission mechanism itself. The 14-month scrape of 4,847 bounties, the 323 confirmed deepfake requests with a 9:1 female skew, and the enforcement-gap figure (41.7% of NSFW deepfake bounties unmarked) are genuinely new and policy-relevant. The human verification of every deepfake case is real work and gives me confidence in that specific count.\n\nWhere it's soft: the central NSFW classification uses Civitai's own explicitness labels as ground truth, with 'R or above' counting as NSFW. The external OpenAI check only gets kappa=0.52, and the paper's own sensitivity—treating R-only as SFW—raises agreement to 0.63. R includes 'adult themes' and 'partial nudity,' not necessarily sexual content. So the weekly-majority claim and its growth trajectory could shift under a different labeling rule. The paper acknowledges the direction of bias but doesn't bound it. A sensitivity analysis or robustness table using OpenAI labels would settle this. Relatedly, the Gini comparisons have no confidence intervals, and the NSFW-deepfake enforcement cell is only 24 bounties; the 41.7% is a proportion with a wide interval. Minor but worth fixing: the abstract says 'majority of all bounties' while the full-sample share is 48%; Figure 3(a) appears to show the NSFW-deepfake subset (95.8% female of 24), not the full deepfake distribution, and should be relabeled; and Section 6 says no AI tools were used in analysis, which contradicts the GPT-4.1 classification methodology. Code and data are promised but not yet available; for a time-bound scrape that's essential.\n\nNone of these look fatal. The deepfake and concentration numbers are probably robust; the NSFW trend would likely survive a more conservative label rule, but we need to see it. The paper deserves a serious referee and a revision, not a desk reject.","headline":"First systematic look at Civitai's paid bounty market, with human-verified deepfake counts; the NSFW trend rests on platform labels and needs a sensitivity check, but the core descriptive findings are likely to hold.","tokens_in":14579,"tokens_out":2517,"would_cite":true,"duration_ms":23959,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that Civitai's paid bounty marketplace, where users commission AI-generated content, has become majority-NSFW over time, and that deepfake requests—overwhelmingly targeting female celebrities—persist even where platform po","keywords":["Civitai","bounty marketplace","generative AI","deepfake","NSFW content","content moderation","gender-based harm","LoRA"],"falsifier":"Collect a fresh, complete set of bounty pages and have independent human annotators classify a random sample (say, 500 bounties) from title, description, and example-image URLs without seeing Civitai's ratings; if the human-coded NSFW share does not rise past 50% by late 2024, or the deepfake gender ratio is not near 9:1, the core claims fail. A second, sharper check: if a re-crawl after Civitai's May 2025 policy change finds deepfake-alert coverage for NSFW bounties at or near 100%, the enforcement-gap finding is a snapshot of a window rather than a durable property.","tokens_in":13510,"feed_emoji":"🎭","tokens_out":5488,"duration_ms":53636,"temperature":0.7,"pith_summary":"The paper examines all 4,847 paid bounty requests posted on Civitai in its first 14 months. It claims that this monetized marketplace is increasingly dominated by not-safe-for-work demand, growing to a majority of weekly requests, and that a small but persistent share of requests commissions deepfakes of real people—nearly nine times more women than men, mostly actresses and influencers. It also claims that the platform's own warning notice for deepfake bounties is missing from 41.7% of explicit deepfake requests while present on 85.6% of safe-for-work ones. If true, the paper shows that a mainstream, venture-backed AI content platform's commission system financially incentivizes explicit content and gendered deepfake harm, and that its governance intervention under-covers the most harmful requests. The stakes are whether monetized, community-driven generative AI platforms can produce reproducible social harms that content policies alone do not prevent.","feed_headline":"Paid AI bounty requests are now mostly explicit","feed_subtitle":"A 14-month study of Civitai finds adult content dominates paid commissions, and deepfakes disproportionately target female celebrities.","key_machinery":"The carrying mechanism is Civitai's bounty marketplace itself: supporters post a paid request in virtual 'buzz' (purchasable for dollars), attach example images, and award the reward to a winning submission. The paper treats these requests as revealed demand, and it specifically isolates LoRA bounties (Low-Rank Adaptations, lightweight fine-tunes of base diffusion models) as the dominant requested artifact—tools that let users steer models toward content they were not trained to generate, including likenesses of real people and explicit material. The measurement machinery is the platform's own explicitness ratings ('R', 'X', 'XXX') used as ground truth for NSFW classification, a large-langua","core_discovery":"On its own terms, the paper's core discovery is that Civitai's bounty system functions as a paid-commission market whose demand is visibly tilting toward adult content, and that deepfake requests concentrated on female celebrities persist despite a stated platform ban on explicit deepfakes. The authors establish this through a complete crawl of public bounties (4,847 requests from October 2023 to January 2025), classifying bounties as NSFW if the platform rated any example image 'R' or more explicit, using a large language model for thematic labeling of safe bounties, and manually verifying all putative deepfake requests. They report that NSFW requests rose steadily to a majority of weekly b","pith_inferences":["A natural extension the paper leaves implicit: other AI-content platforms that introduce paid-commission features should show a similar NSFW skew and concentration pattern, since the incentive structure, not the specific platform, drives the behavior. This is testable by replicating the crawl on comparable marketplaces.","The gender asymmetry (about 9:1 female targets) matches broader patterns of image-based sexual abuse; a testable extension is to check whether targeting of non-celebrities, such as private individuals, appears in less visible channels or after policy changes.","The authors' reliance on platform labels creates a bias; a concrete follow-up they imply is to classify bounties by text alone and compare trends, predicting that text-only classification would find an earlier or steeper NSFW majority than image-label classification.","Policy effectiveness could be studied quasi-experimentally: compare deepfake bounty volume and alert coverage before and after Civitai's May 2025 policy change; the paper's data window could reveal whether external pressure changes behavior or merely produces symbolic updates."],"forward_implications":["If the growth trend holds, adult content is not an edge case on Civitai but the marketplace's central use case, meaning the platform's economics increasingly depend on explicit generation.","Because most bounties request LoRAs, harmful behavior is encoded in reusable artifacts; a single paid request can keep producing the target content long after the bounty closes.","Deepfake demand is concentrated enough (the top 20% of requesters post about 56% of deepfake bounties) that targeting repeat requesters could be a viable governance lever.","The 41.7% missing-alert rate among explicit deepfake bounties implies the platform's stated ban is not enforced by its detection or disclosure process; the 2025 policy extension to safe-for-work deepfakes will face the same enforcement gap unless detection changes.","Because the paper relies on platform image labels as ground truth, its NSFW numbers are a lower-bound estimate of explicit intent; any error from safe-looking sample images would push the true volume higher, not lower."],"fun_headline_variants":["AI bounty requests now majority explicit","Deepfake bounty requests target female celebs","Most paid AI commissions on Civitai are explicit now"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central trend claims stand or fall on whether Civitai's own image ratings ('R' or more explicit) faithfully reflect what supporters intended to request; if supporters posted safe sample images while requesting explicit content in text, the NSFW count—and its steady rise to a majority—could be understated or artifactually shaped, and the platform's labels agree only moderately with an external moderation API (κ=0.52).","fun_headline_variants_meta":{"raw":{"variants":["AI bounty requests now majority explicit","Deepfake bounty requests target female celebs","Most paid AI commissions on Civitai are explicit now"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1357,"prompt_tokens":763,"completion_tokens":594,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":507,"tokens_out":594,"duration_ms":5433,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:41:47.488556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a fresh, complete set of bounty pages and have independent human annotators classify a random sample (say, 500 bounties) from title, description, and example-image URLs without seeing Civitai's ratings; if the human-coded NSFW share does not rise past 50% by late 2024, or the deepfake gender ratio is not near 9:1, the core claims fail. A second, sharper check: if a re-crawl after Civitai's May 2025 policy change finds deepfake-alert coverage for NSFW bounties at or near 100%, the enforcement-gap finding is a snapshot of a window rather than a durable property.","supporting_citations":[],"review_version":1}