{"id":"c95f7e65-8dcb-4f65-b93b-6a1856dd2674","arxiv_id":"2412.16355","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper argues that social science is necessary throughout foundation model development and proposes a framework to operationalize socially responsible AI.","lead":"This position paper argues that building socially responsible AI systems requires social science expertise at every stage, from data collection to deployment. It proposes a framework that maps how foundation models interact with existing systems of power and lays out incentives to encourage such collaboration.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'necessary'/'required' modal claim in Section 3 overstates the evidence: examples show social science can help, not that it is indispensable, and the Facebook case illustrates an incentive failure even with expertise present.","rationale":"This is a clearly written position paper, and much of its substance is sound: it compiles credible evidence that social science findings bear on foundation-model risks, and its incentive discussion is sensible. I read the central claim as advocacy rather than a formal theorem, but the wording 'necessary' and 'required' in Section 3 is the load-bearing assertion and is stronger than the evidence supports. The Facebook case actually undercuts the causal inference because expertise was present and ignored, shifting the explanation to incentives rather than absence of social science. The reader's weakest-assumption about journalistic sources is plausible but secondary; even if those reports are accurate, the inference to necessity still fails without a ruling out of alternative pathways. A position paper may legitimately advocate for collaboration, but if it asserts necessity, it should either defend that modality against a natural counterexample or soften it. This does not change the overall verdict: the paper should remain conditional, with the condition being a defended or weakened necessity claim.","tokens_in":14484,"tokens_out":6387,"duration_ms":59379,"concrete_test":"Perform a modal-substitution check on Section 3: replace every occurrence of 'necessary' and 'required' with 'strongly beneficial' and re-read the argument, framework, and recommendations. If no conclusion becomes false or unsupported, the strong modality is not load-bearing and the central claim should be downgraded to a weaker, better-supported claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is Section 3's 'it is necessary to involve social science expertise throughout the foundation model R&D process' and 'interdisciplinary collaboration between AI and social science is required.' The examples establish at most that social science research can identify risks and that acting on that knowledge can be suppressed or costly. That supports 'beneficial' or 'important,' not 'necessary.' The Facebook/Instagram case (Section 2) is the only concrete 'serious real-world harms' anchor, and it cuts the other way: internal studies already included social science/psychology expertise and proposed mitigations, yet the changes were not adopted because of advertising revenue (Hao 2021; Mac & Kang 2021). Because expertise was present and ineffective absent aligned incentives, the case shows an incentive failure, not an epistemic gap. A purely technical team that reads the social-science literature or consults once could plausibly reach the same outcome, so 'ongoing collaboration throughout all stages' is not proven indispensable. The reader's source-reliability worry (Wells et al. 2021, Hao 2021 are journalistic) is valid but secondary: even granting accuracy, the modal inference fails. The paper's own Appendix A concedes interdisciplinary research is costly and may dilute focus, so the recommendation is a tradeoff; labeling it 'necessary' requires ruling out cheaper substitutes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that social science expertise is necessary throughout the foundation model R&D pipeline to operationalize social responsibility. It proposes a three-component framework: understanding systems of power, designing technical interventions, and anticipating social impacts. It reviews historical examples (social media and teen mental health, Wikipedia gender bias, beauty filters) and discusses incentives for firms and researchers, ending with a list of suggestions for fostering interdisciplinary collaboration.","tokens_in":14709,"tokens_out":3743,"duration_ms":33082,"significance":"The paper offers a readable synthesis of existing literature and a structured framework for integrating social science into AI development. Its strengths include explicitly decomposing the R&D pipeline into stages, linking technical concepts (affordances, bias mitigation) to social-science concepts (systems of power, fields of visibility), and proposing concrete institutional suggestions. The paper does not present new empirical evidence, but it makes a programmatic case. However, the central 'necessity' claim is not supported by the examples; the paper is better characterized as arguing for the value and importance of social science input.","major_comments":[{"comment":"The paper states that 'it is necessary to involve social science expertise throughout the foundation model R&D process' and that 'interdisciplinary collaboration between AI and social science is required.' The evidence in Section 2 (Facebook/Instagram case) actually shows that social science expertise was present in internal research teams, yet the recommendations were not implemented because of advertising revenue incentives (Hao, 2021; Mac & Kang, 2021). This case supports an incentive/structural explanation, not a lack-of-expertise explanation. Moreover, Appendix A concedes that interdisciplinary research can be expensive and may dilute focus, implying a tradeoff. The paper does not rule out substitutes such as consulting existing social-science literature, external audits, or regulation. Without ruling out cheaper substitutes, the modal claim 'necessary/required' overstates what the evidence shows. Please either weaken the claim to 'important' or 'valuable,' or define what standard of social responsibility requires ongoing collaboration and argue why alternatives cannot meet that standard.","section":"Section 3, first paragraph"},{"comment":"The paragraph beginning 'As outlined by Gillan et al. (2021)' presents ESG/CSR financial-performance evidence and then states 'Thus, we hypothesize that tech firms prioritizing social responsibility in providing and deploying foundation models may observe similar financial benefits.' This is explicitly a speculative analogy. Since the paper's incentive argument partly relies on this alignment, the distinction between established findings and the hypothesis should be made clearer, and the limitations of the analogy (e.g., differences between measurable ESG metrics and model-specific social responsibility) should be discussed. This point does not support the 'necessity' thesis and should not be presented as if it does.","section":"Section 4, Incentives for Tech Firms"},{"comment":"The paper asserts that ignoring social-science findings 'has led to serious real-world harms' and cites Wells et al. (2021), Hao (2021), and Mac & Kang (2021) as evidence. These are journalistic accounts of internal company documents and whistleblower testimony; they are not peer-reviewed, and the internal studies' findings are not independently verified. The claim that the 'teams conducting this research were shuttered' is a strong factual assertion that relies on the same reporting. Please either use stronger sources or explicitly acknowledge the evidentiary status of these reports. This matters because the Facebook/Instagram case is the main concrete historical example of harm.","section":"Section 2, Social Media and Teen Mental Health"}],"minor_comments":[{"comment":"There is a typo: 'igredient' should be 'ingredient.'","section":"Section 3, Responsible Model Providers paragraph"},{"comment":"There is a typo: 'instace' should be 'instance.'","section":"Section 2, Social Media and Teen Mental Health paragraph"},{"comment":"Figure 1 is not referenced in the text; please add an explicit reference and a brief explanation of how the pipeline stages relate to the proposed framework.","section":"Figure 1"},{"comment":"The citation for 'Protecting Kids Online, 2021' is formatted inconsistently with the other references; it appears to be a Senate hearing and should be given a proper citation with the committee name and date.","section":"References"},{"comment":"The concept of 'fields of visibility' from Beytía & Wagner (2022) is mentioned without definition; a brief explanation would help readers unfamiliar with this framework.","section":"Section 3, Responsible Model Providers paragraph"}],"recommendation":"major_revision","confidential_remarks":"This is a well-written position paper, but the central modal claim is too strong for the evidence presented. The authors can reasonably fix this in revision by calibrating the language and making the scope of 'necessity' explicit. I would not reject, because the synthesis and suggestions have value for the intended audience; however, if the venue requires the strong claim, it needs either a new argument or a clear definition of the standard that makes ongoing collaboration indispensable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, well-referenced position paper for an ICLR workshop. The new part is modest but real: the three-component framework (understand systems of power, design interventions, anticipate impacts) plus the provider/deployer split gives people a shared vocabulary, and the incentive section is the most useful contribution. It reads as a synthesis of Selbst, Sartori and Theodorou, Birhane, and related work, and the authors do credit those sources.\n\nWhat it does well: clear writing, honest about tradeoffs—Appendix A concedes interdisciplinary research is costly and can dilute focus. That concession is more candid than most papers in this genre. The Wikipedia gender gap and beauty filter examples are well chosen, and the citation pattern looks fair; the few self-citations are peripheral.\n\nThe soft spot is the central modal claim. Section 3 says social science expertise is 'necessary' and 'required' throughout the R&D process. The examples show that social science can help identify risks and that acting on that knowledge can matter. They do not show that no substitute can work or that every pipeline stage needs a social scientist on staff. More importantly, the Facebook/Instagram case—their main concrete anchor for real-world harm—cuts the other way. By their own telling, internal studies already involved social science and psychology expertise and proposed mitigations; the changes were killed for advertising revenue. That is an incentive failure, not an epistemic gap. If expertise was present and ineffective absent aligned incentives, the case does not support 'ongoing collaboration throughout all stages is indispensable.' A purely technical team that reads the literature once might plausibly reach the same outcome, so the strong claim needs either a different evidence base or a softer modal verb.\n\nThe journalistic-source worry (Wells, Hao) is real but secondary; even if those accounts are accurate, the inference still does not go through. The ESG/CSR financial-benefit generalization is explicitly a hypothesis, which is fine, but it should be flagged as such more visibly.\n\nWho this is for: readers working on AI governance, FAccT-style accountability, or interdisciplinary team design. It is a useful framing piece, not a result. I would send it to review because the topic is important and the argument is coherent, but I would push the authors to either defend 'necessary' with a counterexample analysis or reframe as 'strongly beneficial and currently underused.' That reframing would make the paper more defensible and not much less interesting.","headline":"Useful synthesis with a load-bearing 'necessary' that the evidence doesn't support; worth a referee, but the authors should soften the modal claim.","tokens_in":15254,"tokens_out":1860,"would_cite":false,"duration_ms":17780,"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":"Socially responsible foundation models require social science at every pipeline stage, this position paper argues.","keywords":["foundation models","social responsibility","social science","sociotechnical systems","systems of power","technological affordances","interdisciplinary research","AI ethics"],"falsifier":"If internal Facebook or Instagram documents were to show that the teams were closed for reasons unrelated to advertising revenue, or that the findings were implemented at scale, the paper's central example of what happens when social science is ignored would stop supporting its conclusion. A more direct test would compare a model deployed with embedded social scientists against an otherwise identical deployment without them, measuring whether downstream social harms actually differ.","tokens_in":14271,"feed_emoji":"🤝","tokens_out":4597,"duration_ms":38839,"temperature":0.7,"pith_summary":"This position paper argues that foundation models cannot be made socially responsible by technical means alone. To understand how models reproduce or disrupt existing systems of power, to design interventions that steer models toward beneficial affordances, and to anticipate the consequences of deploying models in specific contexts, social science expertise must be involved at every stage of the research and development pipeline. The paper grounds this claim in prior cases where ignoring social science led to harm, such as social media's documented effects on teen mental health and the persistence of gender and racial bias in web-derived training data. It then proposes a three-part conceptual framework and practical incentives to make interdisciplinary collaboration the norm.","feed_headline":"AI alone cannot make foundation models socially responsible","feed_subtitle":"A position paper argues every pipeline stage, from data to deployment, needs social science expertise.","key_machinery":"The load-bearing concept is 'technological affordances'—the actions a technology enables, encourages, or constrains relative to its environment—placed within the study of 'systems of power,' the institutions and norms that distribute privilege and inequality. The paper maps these onto the foundation model pipeline, distinguishing model providers from deployers and assigning each responsibilities: providers must proactively study how training data encodes systems of power, while deployers must consult the relevant social-science disciplines to anticipate application-specific impacts and design mitigation strategies. This framework carries the argument by converting the abstract demand for 'socially responsible AI' into concrete questions at each pipeline stage.","core_discovery":"The paper's central claim is that socially responsible foundation models require the integration of social science throughout the model pipeline, from data collection and training to deployment and retirement. It decomposes this operationalization into three components—understanding systems of power, designing technical interventions, and anticipating social impacts—and argues that while AI researchers are equipped for the technical middle, the other two are properly the domain of social scientists. The paper asserts that ignoring this division of labor has already produced real harms, and that proactive interdisciplinary collaboration is therefore not optional but necessary.","pith_inferences":["If the framework is correct, model evaluation benchmarks should include measures of how well a model reproduces or disrupts systems of power, not just statistical bias metrics.","The argument implies that AI researchers cannot unilaterally define 'human values' for reinforcement learning from human feedback without social science input on cultural variability and contestation.","A testable extension would compare foundation model deployments with and without embedded social scientists, measuring whether downstream social harms actually differ.","The incentive analysis suggests that voluntary corporate action alone may be insufficient; regulatory or standards-setting pressure might be needed to break the 'race to the bottom' among providers."],"forward_implications":["Model providers would need to document and communicate learned biases to downstream deployers, for instance through model cards or dataset documentation.","Deployment decisions would require consultation with domain-appropriate social scientists before launch, not after harm occurs.","Research incentives would need to change so that interdisciplinary work is recognized in hiring, funding, and publication, rather than penalized.","Technical interventions like debiasing are not a 'silver bullet'; they must be coupled with social-science understanding of the context in which the model operates.","Firms that prioritize social responsibility in foundation models may see long-term financial benefits, consistent with evidence that high-ESG and high-CSR firms perform better."],"supporting_citations":[{"why":"Supplies the main historical case that Facebook knew Instagram harms teen mental health and chose not to act.","marker":"Wells et al., 2021"},{"why":"Corroborates the suppression of internal studies and the profit motives behind that decision.","marker":"Hao, 2021"},{"why":"Provides the canonical example of search engines encoding racial systems of power through autocompletion and search results.","marker":"Noble, 2018"},{"why":"Shows how poor data collection choices cascade into downstream social harms, anchoring the pipeline argument.","marker":"Sambasivan et al., 2021"},{"why":"Proposes model cards, the key documentation intervention the paper recommends for communicating learned biases.","marker":"Mitchell et al., 2019"},{"why":"Identifies the abstraction problem in sociotechnical systems, motivating the need for social science in AI development.","marker":"Selbst et al., 2019"}],"fun_headline_variants":["Social science is the missing piece for responsible AI","Responsible AI demands social science at every stage","AI alone can't operationalize social responsibility","Operationalizing responsible AI requires social science"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's concrete example of harm, Facebook and Instagram ignoring internal mental-health findings, rests on journalistic accounts rather than peer-reviewed evidence, and the necessity argument would be weakened if those accounts misrepresent what happened.","fun_headline_variants_meta":{"raw":{"variants":["Social science is the missing piece for responsible AI","Responsible AI demands social science at every stage","AI alone can't operationalize social responsibility","Operationalizing responsible AI requires social science"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000953,"raw_usage":{"total_tokens":3979,"prompt_tokens":771,"completion_tokens":3208,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":387,"completion_tokens_details":{"reasoning_tokens":3151}},"tokens_in":387,"tokens_out":3208,"duration_ms":20526,"temperature":1.0,"reasoning_tokens":3151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:39:11.913384+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If internal Facebook or Instagram documents were to show that the teams were closed for reasons unrelated to advertising revenue, or that the findings were implemented at scale, the paper's central example of what happens when social science is ignored would stop supporting its conclusion. A more direct test would compare a model deployed with embedded social scientists against an otherwise identical deployment without them, measuring whether downstream social harms actually differ.","supporting_citations":[{"cited_title":"Facebook knows instagram is toxic for teen girls, company documents show","cited_arxiv_id":null,"evidence_quote":"Supplies the main historical case that Facebook knew Instagram harms teen mental health and chose not to act."},{"cited_title":"The facebook whistleblower says its algorithms are dangerous","cited_arxiv_id":null,"evidence_quote":"Corroborates the suppression of internal studies and the profit motives behind that decision."},{"cited_title":"Fairness and abstraction in sociotechnical systems","cited_arxiv_id":null,"evidence_quote":"Identifies the abstraction problem in sociotechnical systems, motivating the need for social science in AI development."}],"review_version":1}