{"id":"900adf18-555f-4ad6-a2c2-4c8d4738ab16","arxiv_id":"2411.09102","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Participatory structures in generative AI fail to deliver promised benefits to marginalized communities unless access, ownership, and political conditions are explicitly designed for.","lead":"A workshop paper argues that participatory AI projects often promise marginalized communities benefits that never arrive, because payments are one-time, models sit behind paywalls, and visibility in media does not automatically create political power. It uses a fictional case study of a Vietnamese cultural activist to map the barriers between community labor and community benefit.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the paper's reliance on a speculative case is an admitted limitation, and its central claim is scoped as a provocation; the remaining risk is the empirical grounding of 'dominant structures' in cited prior work.","rationale":"The reader correctly identified the speculative case study as the weakest assumption, and I agree that the paper's generalization about 'dominant structures' rests on anecdotal experience and selective citations rather than systematic evidence. However, I would sharpen the concern: the paper's headline claim is about explicitness (i.e., that current structures are not explicit enough about benefits and harms), but the case study primarily demonstrates that benefits may fail to materialize due to structural barriers. The leap from 'benefits are conditional' to 'participants and developers are not explicit about these conditions' is not directly evidenced. This is load-bearing because the paper's prescription--that researchers and industry actors 'should be more transparent'--depends on the latter premise. The concern is mitigated by the paper's explicit framing as a provocation, its clear Limitations section, and the genre conventions of a workshop paper, which allow for suggestive rather than conclusive empirical grounding. The proposed audit would resolve whether the descriptive claim is actually supported by existing literature or remains an assertion. Given the scope and the authors' honest caveats, the appropriate verdict remains ACCEPT/UNCHANGED; the concern points to a direction for future empirical work rather than a flaw that invalidates the provocation's contribution.","tokens_in":10065,"tokens_out":4482,"duration_ms":67532,"concrete_test":"Conduct a structured audit of a sample of recent participatory GenAI engagements described in the FAccT/CHI literature, including the sources cited as [27] and [45], coding each engagement for whether it explicitly communicates to participants (a) the contingencies on benefit realization (e.g., paywalls, access barriers), (b) potential harms such as likeness misuse or psychological burden, and (c) terms of ownership or control over data and models. If a substantial fraction of engagements already includes such explicit disclosure, the 'not explicit enough' claim is weakened; if few do, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim ('dominant structures ... are not explicit enough about the benefits and harms that members of socially marginalized groups may experience as a result of their participation') contains two logically distinct premises: (1) that in dominant participatory GenAI engagements, benefits to marginalized participants are contingent and may not materialize; and (2) that developers and researchers do not explicitly communicate or interrogate these contingencies. The speculative case study supports (1) by positing paywalls, access barriers, political-economy constraints, and misuse risks, but it does not by itself establish (2): the scenario simply omits any discussion of transparency, and the authors' cited sources [17,27,45] are used to assert that one-time consultation 'illustrates the reality' of participation. Thus the load-bearing empirical weight falls on an anecdote-plus-citation generalization rather than systematic evidence. The authors explicitly acknowledge this in their Limitations section, and the provocation genre lowers the evidentiary bar, so this is not a fatal flaw. The concern would land if the claim were read as a descriptive finding about current industry practice rather than a motivated call to interrogate that practice. This is the same weakest assumption the reader identified, but framed more precisely as a gap between showing that benefits are uncertain and showing that they are not made explicit.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper, framed as a provocation, argues that dominant structures of community participation in generative AI (GenAI) development and evaluation are not explicit enough about the benefits and harms that members of socially marginalized groups may experience as a result of their participation. To support this argument, the authors present a speculative case study of 'Thuy,' a Vietnamese cultural preservation activist invited to label cultural artifact images for a text-to-image company. They trace how potential benefits (improved representation, quality-of-service) may be blocked by paywalls, access barriers, and socio-political conditions, and how harms (misuse, displacement of cultural consultants) can arise. The paper concludes with implications for researchers and an appendix outlining alternative models of participation, ownership, and compensation. The authors acknowledge in their Limitations section that the analysis is based on a speculative context informed by their own experiences rather than on systematic empirical evidence.","tokens_in":10285,"tokens_out":3181,"duration_ms":38985,"significance":"If taken as a call to interrogate participatory AI practices, the paper is timely and valuable: it names concrete barriers to benefit realization and assembles relevant scholarship from media studies, data sovereignty, and responsible licensing. Its transparent use of a speculative case and its explicit limitations are strengths, and the appendix offers a useful set of existing alternatives (e.g., Te Hiku Media, DAIR, data leverage) that researchers can build on. The paper's weakness is that its central descriptive claim — that current practices are 'not explicit enough' — is supported by anecdote and inference rather than systematic evidence. Nonetheless, as a provocation, the paper successfully opens a space for further empirical investigation.","major_comments":[{"comment":"The central claim has two premises: (1) benefits to marginalized participants are contingent and may not materialize, and (2) developers and researchers do not explicitly communicate or interrogate these contingencies. The speculative case study convincingly illustrates premise (1) by detailing paywalls, access barriers, and political-economy constraints, but it does not establish premise (2). The scenario simply states that 'the company has not explored paths for participant ownership or control over data or AI models' and later asserts that this model 'illustrates the reality of how technology institutions and academic researchers often engage socially marginalized communities in AI development today' (Implications). The cited sources [17,27,45] support the prevalence of one-time consultative participation, but they do not directly document a systematic lack of explicitness about benefits and harms. To make the provocation load-bearing, the paper should either soften the claim to 'may not be explicit enough' or 'we should investigate whether,' or provide empirical evidence of actual transparency practices in participatory GenAI engagements.","section":"'Speculative Case Study' and 'Implications'"},{"comment":"The Limitations section honestly acknowledges that the case study is speculative and based on the authors' collective experiences, and that future work should analyze real-world examples. However, the 'Implications' section makes a stronger generalizing move, claiming that the one-time consultation model 'illustrates the reality' of current engagement. This gap between the acknowledged limitation and the assertion in Implications is a load-bearing tension. The paper would be strengthened by explicitly qualifying the generalizing claim as a hypothesis or a call for empirical validation, rather than presenting it as an established finding.","section":"Section 2 (Limitations)"}],"minor_comments":[{"comment":"There is a missing space in the sentence 'of her communityunless social, political, and economic conditions all align'; it should read 'of her community unless.'","section":"Section 'Harms marginalized groups can experience'"},{"comment":"The font size in the figure's dependency boxes and arrows is small; recommend increasing readability for print and reproduction.","section":"Figure 1"},{"comment":"The phrase 'non-dominant values' could be clarified with a brief definition or example, as it is central to the motivation.","section":"Abstract and Introduction"},{"comment":"The appendix does a good job of listing alternative models, but the transition from the main text to the appendix could be smoother; a forward reference in the 'Implications' section already exists, which is helpful.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"This is a workshop-style provocation with a clear and honest scope. The main concern is the empirical support for the descriptive claim about current practice. If the journal accepts provocations as a genre, the paper could be published after qualifying the central claim; if the journal expects empirical contributions, the authors should add a small case-study analysis or reframe the paper as a research agenda. The citation pattern is appropriate, and the paper does not overstate its contributions beyond the provocation framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-crafted provocation, not a research paper. It makes a useful move—reframing participatory GenAI evaluation from \"were participants included?\" to \"who actually captures the benefit, and via what pathways?\"—and it is honest about its speculative basis. The speculative case of Thuy is a good device: it makes concrete the financial, access, and sociopolitical barriers that are often gestured at. The paper is well-cited for a workshop piece, and the authors explicitly say they are not offering systematic evidence. That is the right posture for a provocation.\n\nThe soft spot is the empirical weight. The central claim is really two claims: that benefits to marginalized participants are contingent, and that dominant participation structures are not explicit about these contingencies. The case study supports the first well, but the second is asserted more than shown; the scenario just omits transparency, and the cited prior work is used to generalize from one-time consultation to \"the reality\" of practice. The authors do flag this in the Limitations section, so they are not hiding it. For a peer-reviewed version, I would want a more systematic look at actual participatory engagements, or at minimum a clearer argument that the opacity is a structural feature rather than an accidental omission.\n\nThis is genuinely a good reading-group piece for FAccT/ethics crowds. It will not change your research agenda, but it is a clean articulation of a critique that many people half-make. I would cite it if I were writing about participatory AI evaluation. And yes, it deserves a serious referee: a strong position paper, worth engaging on the merits. I would accept it at a workshop without hesitation, and for a conference position track I'd send it to review and let the authors decide whether to sharpen the empirical claim.","headline":"A well-scoped provocation that reframes participatory AI evaluation around benefit realization; the descriptive claim about 'dominant structures' is thinner than the argument wants, but the paper is honest about it.","tokens_in":10855,"tokens_out":2324,"would_cite":true,"duration_ms":29034,"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":"Dominant structures of community participation in generative AI fail to specify what marginalized participants actually gain, and a speculative case study shows the promised benefits are blocked by paywalls, access gaps, and…","keywords":["participatory AI","generative AI","representational harms","marginalized communities","community participation","data labor","inclusion","AI evaluation"],"falsifier":"Check the claim against real participatory engagements: for a documented text-to-image data enrichment program that recruited members of a marginalized community, determine whether participants received free or affordable access to the resulting model, ongoing royalties or ownership rights, and protections against impersonation or misuse. If such engagements consistently provide these benefits, the paper's claim that benefits are empty under dominant structures is weakened; if they do not, it is supported.","tokens_in":9826,"feed_emoji":"⚖️","tokens_out":6506,"duration_ms":101252,"temperature":0.7,"pith_summary":"Participatory AI is now a standard response to generative AI's representational harms: companies and researchers invite members of marginalized communities to label data, evaluate outputs, and share cultural expertise. This paper argues that the dominant form of this participation—one-time compensated consultation with no ownership over the resulting data or models—is not explicit enough about what those participants actually gain or risk. Using a speculative case study of a Vietnamese cultural preservation activist invited to label images of cultural artifacts for a text-to-image company, the authors trace the concrete barriers that block the promised benefits: paywalls, connectivity and interface gaps, and the political-economic conditions that keep improved representation from translating into material change. The paper concludes that the claim that community members will be better off from participation is empty under current structures, and urges AI developers to make the contingencies of benefit explicit and to restructure participation toward meaningful community ownership and power.","feed_headline":"Who really benefits from 'inclusive' AI?","feed_subtitle":"A provocation shows marginalized participants often end up behind paywalls, without ownership of the models they help build.","key_machinery":"The paper's central device is a speculative case study of 'Thuy,' a Vietnamese cultural preservation activist invited to label photographs of Vietnamese cultural artifacts for a text-to-image company's data enrichment initiative. The scenario is an explicit abstraction of what the authors observe as the dominant participation structure—one-time compensation, no ownership or control over data or models—and it is used to trace the flow of benefits and harms across three stakeholder groups: community members, social actors such as marketing agencies and publishers, and technology institutions. The case study functions as the argument's test bed, converting a general worry about extractive participation into a concrete accounting of where promised benefits stop and where harms begin.","core_discovery":"The central claim is that dominant structures of community participation in AI development and evaluation are not explicit enough about the benefits and harms that members of socially marginalized groups may experience as a result of their participation. The paper contends that participation is typically motivated by a trickle-down logic—improved model representations of a culture will benefit its members—but this logic fails when participants cannot access the resulting services, when representation does not alter material conditions, and when heightened visibility creates new harms such as impersonation or the displacement of community labor. Because these barriers are structural rather than incidental, the intended benefits remain contingent on reforms the standard consultation model does not address; hence the authors call for transparency about these contingencies and for restructuring participation around ownership, control, and power for participants.","pith_inferences":["The paper's logic implies a measurable criterion for whether a participatory engagement is extractive: participants should be able to use the outputs, share in their value, and retain rights over their contributions, a criterion that could be operationalized in future audits.","The speculative case study generates testable predictions—for example, that participants in current text-to-image data enrichment programs rarely receive free model access, ongoing royalties, or likeness protections—which could be checked through interviews or contract analysis.","By framing inclusion as a question of who benefits, the paper connects participatory AI to the wider discourse on data labor and platform exploitation, suggesting that 'inclusion' without redistribution may function as a source of legitimacy for AI developers."],"forward_implications":["If the claim holds, one-off compensated consultation is an inadequate model for participatory AI; engagements must specify and secure participant benefits before and after data collection.","Researchers and companies should tell participants which benefits they can and cannot realize, and under what conditions, rather than asserting that improved representation will help.","Alternatives such as community-owned models, usage licenses that redistribute benefits, and data-leverage bargaining become necessary components of ethical participation.","Participatory evaluation should include explicit discussion of potential harms and leave room for participants to refuse or withdraw.","Regulatory requirements that mandate consultation with impacted groups will not, by themselves, protect those groups unless the consultation structures grant ownership and control."],"supporting_citations":[{"why":"Documents the current state of participatory AI practice, grounding the paper's description of dominant structures.","marker":"[17]"},{"why":"Identifies how foundation-model participation is often limited to consultation with little say over deployment, the target of the critique.","marker":"[27]"},{"why":"Argues that 'artificial inclusion' can be extractive, supporting the claim that benefits to participants are not automatic.","marker":"[39]"},{"why":"Provides evidence of geographic representational harm in text-to-image models, motivating the case for community participation.","marker":"[2]"},{"why":"Exemplifies a community-sourced dataset initiative for culturally diverse food, illustrating the promise and limits of participation.","marker":"[4]"},{"why":"Media studies argument that representation in media does not by itself produce material change, backing the non-end-user barrier.","marker":"[33]"},{"why":"Critique of superficial representation in media, supporting the claim that improved depiction may not translate into benefit.","marker":"[34]"},{"why":"Shows how production cultures and political economy shape race representation, underpinning the emphasis on systemic barriers.","marker":"[35]"},{"why":"Introduces data leverage as a mechanism for contributor power, cited as a path forward for restructuring participation.","marker":"[49]"}],"fun_headline_variants":["Inclusive AI's blind spot: who pays for participation?","Marginalized communities often lose in 'inclusive' AI","The catch in community-based AI evaluation","Participation without ownership won't fix AI bias","Why 'inclusive' AI can exclude its own participants"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's analysis rests on the assumption that the speculative case study of 'Thuy' faithfully abstracts the dominant structures of participatory AI in industry and academia today, based on the authors' collective experiences rather than systematic empirical evidence.","fun_headline_variants_meta":{"raw":{"variants":["Inclusive AI's blind spot: who pays for participation?","Marginalized communities often lose in 'inclusive' AI","The catch in community-based AI evaluation","Participation without ownership won't fix AI bias","Why 'inclusive' AI can exclude its own participants"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3499,"prompt_tokens":832,"completion_tokens":2667,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":2590}},"tokens_in":448,"tokens_out":2667,"duration_ms":21308,"temperature":1.0,"reasoning_tokens":2590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:01:48.356399+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check the claim against real participatory engagements: for a documented text-to-image data enrichment program that recruited members of a marginalized community, determine whether participants received free or affordable access to the resulting model, ongoing royalties or ownership rights, and protections against impersonation or misuse. If such engagements consistently provide these benefits, the paper's claim that benefits are empty under dominant structures is weakened; if they do not, it is supported.","supporting_citations":[{"cited_title":"Beards, scarves, halal meat, terrorists, forced marriage’: television industries and the production of ‘race","cited_arxiv_id":null,"evidence_quote":"Shows how production cultures and political economy shape race representation, underpinning the emphasis on systemic barriers."},{"cited_title":"The participatory turn in ai design: Theoretical foundations and the current state of practice, 2023","cited_arxiv_id":null,"evidence_quote":"Documents the current state of participatory AI practice, grounding the paper's description of dominant structures."},{"cited_title":"Participation in the age of foundation models","cited_arxiv_id":null,"evidence_quote":"Identifies how foundation-model participation is often limited to consultation with little say over deployment, the target of the critique."},{"cited_title":"Stevie Bergman, Jennifer Chien, Mark Díaz, Seliem El-Sayed, Jaylen Pittman, Shakir Mohamed, and Kevin R","cited_arxiv_id":null,"evidence_quote":"Argues that 'artificial inclusion' can be extractive, supporting the claim that benefits to participants are not automatic."},{"cited_title":"Bell, Candace Ross, Adina Williams, Michal Drozdzal, and Adriana Romero Soriano","cited_arxiv_id":null,"evidence_quote":"Provides evidence of geographic representational harm in text-to-image models, motivating the case for community participation."},{"cited_title":"You are what you eat? feeding foundation models a regionally diverse food dataset of world wide dishes","cited_arxiv_id":null,"evidence_quote":"Exemplifies a community-sourced dataset initiative for culturally diverse food, illustrating the promise and limits of participation."},{"cited_title":"Subject (ed) to recognition","cited_arxiv_id":null,"evidence_quote":"Media studies argument that representation in media does not by itself produce material change, backing the non-end-user barrier."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Critique of superficial representation in media, supporting the claim that improved depiction may not translate into benefit."},{"cited_title":"Data leverage: A framework for empowering the public in its relationship with technology companies","cited_arxiv_id":null,"evidence_quote":"Introduces data leverage as a mechanism for contributor power, cited as a path forward for restructuring participation."}],"review_version":1}