{"id":"20c29dde-2b5d-44ad-844f-35377c7dc6b3","arxiv_id":"2411.09751","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The first systematic walkthrough of 20 AI nudification websites finds most target women, half mention consent, and all monetize through subscriptions, crypto, and APIs.","lead":"Researchers analyzed 20 AI-based 'nudification' websites that turn clothed photos into nude images, mapping how these sites advertise, charge, and handle consent. This is the first multi-site study of this abusive software ecosystem, and it gives policymakers and platforms concrete pressure points such as payment processors and APIs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 20-app sample, drawn from NGO lists and three 'Top X' articles from a single US vantage point, may not represent the broader 'popular and easy-to-find' ecosystem; a multi-source, multi-location replication check is needed before the ecosystem-level generalizations hold.","rationale":"The reader's weakest assumption is exactly the representativeness of the 20-site sample assembled from NGO lists and three 'Top X' articles from a single US location. My analysis agrees: this is the most load-bearing point because nearly every headline finding (women-focused, weak consent, age verification gaps, commercial monetization) is a proportion over this sample, and the policy/intervention discussion generalizes from these proportions to the ecosystem. The paper acknowledges the location limitation but does not test whether the sample frame biases the observed proportions. A focused replication with an independently drawn sample, coded with the same instrument, would settle whether the findings are robust or an artifact of source selection. The verdict remains CONDITIONAL because the existing evidence is plausible and internally consistent, but the generalizability question is unresolved pending this check. No change to the reader's conditional verdict is needed; the concern strengthens the rationale for the condition rather than overturning the assessment.","tokens_in":18218,"tokens_out":1995,"duration_ms":25446,"concrete_test":"Independently construct a new sample: (1) run Google and Bing searches for nudification-related keywords from at least three vantage points (e.g., US, EU, and Asia via VPN), (2) collect apps from app-store search results and social-media advertisement links, and (3) code the first 20 newly identified, currently accessible sites using the same rubric as the paper (women-focused, consent text, age verification, payment processors, API access). Compare the new sample's proportions to the paper's reported values (e.g., 19/20 women-focused, 10/20 consent, 14/20 age verification) using Fisher's exact test for each proportion. If the new sample shows significantly lower women-focus (e.g., below 75%) or significantly higher consent/age-verification rates, the original sample is not representative and the ecosystem-level conclusions need qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim — that the nudification ecosystem is commercial, predominantly women-focused, and weakly governed by consent and age checks — depends on the 20 analyzed websites being representative of the 'popular and easy-to-find' segment. The sampling in Section 3.1 draws only from NGO-provided lists and three 'Top X' search articles, all observed from a single US location. The authors report saturation, but saturation within these sources does not establish representativeness of the wider ecosystem. The headline proportion (19/20 women-focused) could be an artifact of source selection: NGO lists and search-article authors may disproportionately surface apps that target women, while general-purpose or male-focused nudification tools may be equally easy to find through other entry points (app stores, social media ads, non-English search results, or geo-specific sites). Likewise, the US-only vantage point may miss region-specific apps or UI variants that differ in age verification and consent prompts. This matters because the paper's mitigation proposals (e.g., targeting payment processors, single sign-on providers, and APIs as intervention points) presuppose that the observed monetization and governance patterns characterize the ecosystem at large, not just a convenience sample. Without a robustness check against an independently drawn sample, the generalization from 'these 20 apps' to 'the ecosystem' is under-supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a walkthrough study of 20 AI nudification websites, based on the Light et al. application walkthrough method. It asks three questions: how the sites position themselves, what features they advertise, and how they monetize. The authors find that 19 of 20 sites explicitly specialize in the nudification of women, that half mention consent expectations in some form but few ask for affirmation during the normal user flow, that many sites lack meaningful age confirmation, and that all 20 sites are commercial, using a mix of cryptocurrency, PayPal, credit cards, affiliate programs, and API resale. The paper proposes intervention points such as payment processors, single sign-on providers, and model owners. It deliberately does not name the studied sites and uses artistic renderings of interfaces for ethical reasons.","tokens_in":18413,"tokens_out":9549,"duration_ms":93955,"significance":"If the results hold, this is a valuable first systematic characterization of a harmful and understudied ecosystem, with concrete implications for platform accountability and policy. The paper is strong in its ethical design: it discusses researcher safeguarding, includes a content warning and sensitivity read, avoids naming sites to limit traffic, and clearly frames the study as examining advertised features rather than experimentally verified capabilities. The counts are internally consistent and the first half of the sample was double-coded with Cohen's Kappa 0.88. The main value is descriptive: it provides a baseline of the ecosystem's commercial structure, consent and age-verification gaps, and monetization channels. The central limitation is external validity: the sample is a convenience sample from NGO lists and three 'Top X' articles, all from a single US location, and the paper's broader ecosystem-level claims depend on that sample being representative.","major_comments":[{"comment":"The paper describes the sample as 'representative' and generalizes from the 20 studied sites to 'the ecosystem' in Findings #1-#8 and in the Section 8 mitigation discussion. However, the sampling frame is two NGO-provided lists and three 'Top X' search articles, all observed from a single US location, and saturation is only established within those sources. This is a convenience sample: the NGO and search-article sources may systematically over-represent apps that target women and under-represent other entry points such as app stores, social media ads, non-English or geo-specific sites. Because the headline counts (e.g., 19/20 women-focused, 10/20 consent mentions, 17/20 crypto) and the proposed intervention points (payment processors, SSO providers, APIs) depend on the representativeness of those 20 sites, the authors should either add a robustness check based on an independently drawn sample from additional sources and vantage points, or explicitly restrict all ecosystem-level claims to the studied sites.","section":"Section 3.1 and Section 3.5"},{"comment":"The authors report Cohen's Kappa of 0.88 on the first 10 sites, but the remaining 10 sites were coded by a single researcher. Since the paper's contributions are exact counts out of 20 (e.g., 19/20, 10/20, 17/20), a single coder's interpretation on the second half directly affects the reported counts and the qualitative labels such as 'about half' or 'most'. I recommend releasing an anonymized codebook and coding matrix, and double-coding at least a subset of the second half to establish reliability across the full dataset, or explicitly stating this as a significant limitation of the count-based claims.","section":"Section 3.2"}],"minor_comments":[{"comment":"The text says 'seven of the 20 (33%) websites' but 7/20 is 35%; the percentage should be corrected.","section":"Section 4"},{"comment":"Finding #4 says 'Half of applications allow for images to be modified such that the image subject is put into sexual scenes,' but the preceding text reports 9 of 20 applications offering Deepnudes features; 9/20 is 'about half' under the paper's own terminology in Section 4, so the wording should be adjusted for consistency.","section":"Section 5.1"},{"comment":"The text refers to 'Appendix 6' for the payment breakdown, but the relevant appendix is labeled C; the cross-reference should be fixed.","section":"Section 6"},{"comment":"There are several typos and minor usage errors: 'nudificaiton' in Section 3, 'ubiquitious' in Section 8, 'altercations' in Section 8.1, 'Paetreon' and 'cyptocurrency' in Figure 6, and 'notification' for 'nudification' in Section 8.2.","section":"Throughout"},{"comment":"The paper does not specify how many NGO lists were used, what search queries generated the 'Top X' articles, or how many candidate sites were screened before saturation; providing these details (without naming the sites) would improve reproducibility.","section":"Section 3.1"},{"comment":"The abstract says 'most sites explicitly target the nudification of women,' but Section 4 reports 19 of 20, which the paper's own terminology would call 'almost all'; the labeling should be consistent.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a valuable and ethically careful descriptive study. The key risk is external validity: the sample is a convenience sample assembled from NGO lists and three 'Top X' articles from a single US vantage point, and the paper's ecosystem-level conclusions and intervention recommendations depend on that sample. I recommend major revision focused on either adding a robustness check or explicitly limiting the claims. The single-coder issue on half the sites is secondary but should be addressed. I do not see evidence of circular reasoning or fabrication; the counts are internally consistent apart from minor wording and labeling issues. The refusal to release data is justified by the sensitive and potentially harmful nature of the studied websites."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First systematic map of the AI nudification application ecosystem, and it is a genuinely useful baseline for policy and platform work. The walkthrough of 20 websites shows a commercial ecosystem that predominantly targets women, rarely enforces consent or meaningful age checks, and monetizes through crypto, PayPal, credit cards, and API resale. The API resale / repackaging finding (5 of 20 offer API access) is the most novel piece; it suggests the ecosystem can spawn new storefronts without retraining models.\n\nThe paper is methodologically transparent and ethically careful. Double coding the first 10 sites with Cohen's Kappa 0.88 is good practice. The authors are explicit that they did not test features, did not upload real images, and accessed the sites only from the US. They also went further than most: no site names, therapist support for researchers, artistic renderings instead of real imagery, and a sensitivity read. That should be credited.\n\nThe soft spots are mostly about representativeness. The sample is drawn from NGO lists and three 'Top X' articles, all from a single US vantage point. That could inflate the women-focused proportion and miss region-specific or app-store-distributed tools. The authors acknowledge this in Section 3.5, but the findings still report '19 of 20' as a headline number, so readers will treat it as prevalence. I read it as directional evidence, not a population estimate. The single-coder coding for the second 10 sites is a minor issue given the codebook seems straightforward, but reporting reliability for that half would strengthen the paper. The decision not to release site names is ethically justified, but the authors could still release anonymized coding sheets to support reproducibility.\n\nThe stress-test note argues that ecosystem-level generalizations are under-supported. I think that is slightly too strong. The core claims—commercial, women-focused, weak consent—are consistent with prior work on image-based sexual abuse and deepfake communities, and the paper itself calls it a case study. The exact counts are fragile; the qualitative patterns are not.\n\nThis deserves serious peer review. I would ask reviewers to push on generalizability, request a reproducibility appendix, and ask for a more careful discussion of how the sample relates to the wider ecosystem (including non-English and app-based tools). The paper is aimed at policymakers, platform safety teams, and security researchers working on abuse ecosystems, and I would bring it to reading group and cite it.","headline":"A timely, ethically careful first map of the AI nudification ecosystem, with a real but acknowledged sampling limit that makes the headline proportions directional rather than precise.","tokens_in":19009,"tokens_out":3426,"would_cite":true,"duration_ms":33831,"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":"AI nudification apps form a commercial ecosystem that predominantly targets women and is weakly governed by age and consent checks.","keywords":["AI nudification","synthetic non-consensual explicit imagery","image-based sexual abuse","deepfake","application walkthrough","monetization","consent verification","age verification"],"falsifier":"A broader sample assembled from non-English searches, app stores, social-media ads, and multiple countries that found most popular nudification sites do not specialize in women, that meaningful age or consent verification is common, or that a substantial share operate without any payment mechanism would refute the paper's central characterization.","tokens_in":17976,"feed_emoji":"🔞","tokens_out":9410,"duration_ms":89384,"temperature":0.7,"pith_summary":"This paper measures the ecosystem of websites that turn a photo of a clothed person into a nude or sexually explicit image, studying 20 popular and easy-to-find services. It argues that the ecosystem is a commercial industry whose default product is the undressing of women: 19 of 20 sites explicitly specialize in that, only half mention that users need the depicted person's consent, and six never ask users to confirm that they are adults. The paper's purpose is to replace anecdote with a systematic map of how these services present themselves, what features they sell, and how they take money, so that regulators, payment processors, and platforms have concrete points of intervention. If the characterization is right, the same ordinary infrastructure that lets these sites sell subscriptions is also what makes them reachable by policy.","feed_headline":"AI nudification apps are a commercial, women-focused market","feed_subtitle":"A 20-site study maps the payments, age gates, and consent gaps behind non-consensual AI nude images.","key_machinery":"The central machinery is the application walkthrough method [24], applied to a corpus of 20 websites and organized into entry, everyday-use, and exit stages. The paper adds double coding of the first ten sites, with inter-rater reliability of $\\kappa = 0.88$, to standardize judgments about imagery, age gates, consent text, advertised features, and payment methods. The walkthrough treats advertised functionality as the object of study rather than testing the services on real images, on the ethical ground that uploading images of people to potentially adversarial services would create risk; the working assumption is that what a site advertises is what users expect and what future versions will deliver.","core_discovery":"On the paper's own terms, the central discovery is that point-and-click AI nudification has already consolidated into a commercial, women-targeted market. Across the 20 sampled sites, the defining feature, an AI undressing tool that predicts the nude body under clothing, appears on 18 sites, and 19 sites communicate through imagery or text that their product is the nude or sexualized female form. Consent and age safeguards are largely formal: 10 of 20 sites mention in their terms of service that users should have the image subject's consent, fewer actually ask for confirmation, and the age check, where present, is usually a click-through button that can appear after nude content has already been shown. The same sites monetize through tiered subscriptions, affiliate and referral programs, and API resale, with 17 of 20 accepting cryptocurrency and the lowest-cost paid tier averaging $0.31 per generated image. The paper concludes that because every sampled site is commercial, monetization infrastructure is a viable chokepoint for reducing harm, and because the tools are positioned to create non-consensual imagery of women, consent and age verification should be required of the platforms rather than left to users.","pith_inferences":["Editorial inference: if the commercial dependency is as strong as the data suggest, a watchlist assembled from payment-processor records would likely find more of the ecosystem than search-based lists, because every sampled site must transact.","Editorial inference: the API-resale finding implies that interventions aimed only at consumer-facing websites will be bypassed; model hosts and API-key issuers are a more durable target.","Editorial inference: the paper's ethical choice not to upload images leaves open a measurement gap; testing advertised features using synthetic or fully consenting images could determine whether the sites deliver what they sell.","Editorial inference: because all observations came from a single US vantage point, a natural extension is to repeat the walkthrough from other regions; persistence of the women-focused, commercial pattern would strengthen the ecosystem-level claim."],"forward_implications":["Payment rails are a practical intervention point: 17 of 20 sampled sites accept cryptocurrency, and three rely on conventional card and online payment processors, so payment intermediaries sit inside nearly every monetization path.","Click-through age gates are not protecting minors or subjects: 6 of 20 sites never ask for an age confirmation, and 5 of the remaining sites show nude or sexual content before the confirmation appears.","Consent requirements are mostly boilerplate: only 10 of 20 sites state in their terms of service that the user needs the depicted person's consent, and only 7 present consent language before image generation.","API resale means takedown of a single storefront is unlikely to stop the capability: 5 of 20 sites sell model access for between $20 and $299 per month, letting other sites repackage the same model.","The ecosystem is not designed for user or subject control: of the 17 accounts the researchers created, only 1 could be deleted through the website."],"supporting_citations":[{"why":"Supplies the application walkthrough method that structures the study's entry, everyday-use, and exit observations.","marker":"[24]"},{"why":"Provides the foundational argument that abusive ecosystems must be understood before they can be secured.","marker":"[3]"},{"why":"Establishes the perceived harms of AI-generated non-consensual intimate imagery that motivate the study.","marker":"[6]"},{"why":"Documents how victims of image-based sexual abuse seek help, framing why the ecosystem itself must be studied.","marker":"[56]"},{"why":"Demonstrates that monetization ecosystems of abusive software can be measured and treated as intervention points.","marker":"[16]"},{"why":"Provides the prior finding that most non-consensual pornography content depicts women, supporting the gender-targeting finding.","marker":"[53]"},{"why":"Identifies the image-inpainting technique that underlies the advertised AI undressing feature.","marker":"[61]"}],"fun_headline_variants":["AI nudification study: 20 sites, few consent checks, crypto payments","Most AI nudification apps target women and skip real age gates","AI undress tools: commercial, women-focused, with $0.31 per image","Consent and age checks are mostly formal on AI nudification sites","Monetization is a chokepoint for AI nudification apps, study finds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 20 sites found through NGO lists and three 'Top X' search articles, observed from one US location, are representative enough of the popular and easy-to-find nudification ecosystem that the observed patterns belong to the ecosystem as a whole.","fun_headline_variants_meta":{"raw":{"variants":["AI nudification study: 20 sites, few consent checks, crypto payments","Most AI nudification apps target women and skip real age gates","AI undress tools: commercial, women-focused, with $0.31 per image","Consent and age checks are mostly formal on AI nudification sites","Monetization is a chokepoint for AI nudification apps, study finds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000338,"raw_usage":{"total_tokens":1921,"prompt_tokens":1052,"completion_tokens":869,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":769}},"tokens_in":668,"tokens_out":869,"duration_ms":8459,"temperature":1.0,"reasoning_tokens":769,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:20:14.616682+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A broader sample assembled from non-English searches, app stores, social-media ads, and multiple countries that found most popular nudification sites do not specialize in women, that meaningful age or consent verification is common, or that a substantial share operate without any payment mechanism would refute the paper's central characterization.","supporting_citations":[{"cited_title":"Light, J","cited_arxiv_id":null,"evidence_quote":"Supplies the application walkthrough method that structures the study's entry, everyday-use, and exit observations."},{"cited_title":"Anderson","cited_arxiv_id":null,"evidence_quote":"Provides the foundational argument that abusive ecosystems must be understood before they can be secured."},{"cited_title":"Violation of my body:","cited_arxiv_id":null,"evidence_quote":"Establishes the perceived harms of AI-generated non-consensual intimate imagery that motivate the study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents how victims of image-based sexual abuse seek help, framing why the ecosystem itself must be studied."},{"cited_title":"Gibson, V","cited_arxiv_id":null,"evidence_quote":"Demonstrates that monetization ecosystems of abusive software can be measured and treated as intervention points."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the prior finding that most non-consensual pornography content depicts women, supporting the gender-targeting finding."},{"cited_title":"confirmation","cited_arxiv_id":null,"evidence_quote":"Identifies the image-inpainting technique that underlies the advertised AI undressing feature."}],"review_version":1}