{"id":"75d06d02-03c8-4d0e-a876-59f1c709a025","arxiv_id":"2606.01171","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AIM: a methodological stance with seven preconditions for centering minoritized lived experience before technical AI framing begins, illustrated in a Dutch healthcare engagement with 13 women and non-binary people of color.","lead":"An AI-ethics research group proposes 'AI From the Margins' (AIM): seven standing conditions for running the early, pre-design phase of participatory AI with minoritized communities, illustrated in eight sessions with 13 women and non-binary people of color in Dutch healthcare. The work targets a widely acknowledged failure mode — community input usually arriving only after the key choices about what an AI is for are already made — and offers a structured way to move that inpu","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical claim that AIM's preconditions are necessary and mutually constitutive is undercut by circular analysis: the sessions are coded through AIM's own preconditions, so they cannot demonstrate the preconditions' necessity.","rationale":"After reviewing the paper and the reader's verdict, I find the most load-bearing threat to the central claim is the evidential circularity in the qualitative analysis. The paper claims that its seven preconditions are necessary and mutually constitutive, and that the sessions demonstrate this. But the analytical approach uses the preconditions as the coding lens, so the 'finding' of interdependence is constructed rather than discovered. This is compounded by the paper's own admission that the empirical application is a worked example, not a test/evaluation. The abstract's 'demonstrating' therefore overstates what the data can show. The reader's weakest assumption—that the premise 'problem definitions ... already set' is taken from citation—is also important: if current participatory practice already allows frame reshaping, AIM's niche weakens. However, even if that premise is granted, the central claim about necessity remains unsupported by a self-confirming analysis. Thus I partially agree with the reader: both concerns matter, but the circularity/evidence gap is more directly load-bearing for the 'necessary and mutually constitutive' formulation. The proposed test—an independent blind re-coding of the transcripts—would settle whether the preconditions genuinely emerge from the material. Given the paper's status as a methodological stance, the appropriate verdict remains conditional: the framework is coherent and theoretically grounded, but the empirical demonstration should be downgraded to illustration.","tokens_in":17087,"tokens_out":6594,"duration_ms":67465,"concrete_test":"Pre-register an independent qualitative analysis of the session transcripts in which coders are blind to AIM and its preconditions, using open/inductive coding. Then map the emergent themes onto the seven preconditions. If the preconditions and their interdependence do not emerge without the AIM lens, the empirical support for their necessity and mutual constitution is an artifact of the coding framework.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the seven preconditions are 'necessary and mutually constitutive conditions' for any participatory AI that centers lived experience. The empirical support is a single worked example (8 sessions, 13 participants) whose analysis is circular. In the 'Analytical approach' section, the authors state that they 'read session notes and transcripts iteratively alongside AIM's seven preconditions, identifying moments that were especially illustrative of how those preconditions manifested.' This means the identification of the preconditions in the data is guided by the preconditions themselves. The 'central observation' in the Discussion that 'each precondition created the conditions under which others could be meaningfully enacted' is thus an interpretive product of the coding framework, not an independent empirical finding. The paper also explicitly disclaims that the study is designed to test the method: 'the empirical application reported here serves as a worked example ... rather than as a study designed to test or evaluate the method's general effectiveness.' Yet the abstract claims the sessions demonstrate how preparatory orientation shapes what participatory AI is for. Without a non-circular analysis or a comparative design (e.g., a process lacking one or more preconditions), the claim that the preconditions are necessary and mutually constitutive cannot be sustained. The 'already set' premise about current practice is also load-bearing, but even granting it, the universal normative claim outruns the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'AI From the Margins' (AIM), a methodological stance for participatory AI design that defines a preparatory stage prior to technical framing. AIM is operationalized as seven preconditions — reciprocity, decolonizing stance, centering minoritized standpoints, methodological flexibility, structural availability of refusal, multidimensional accessibility, and co-ownership — which the authors argue are necessary and mutually constitutive for any participatory AI process that genuinely centers the lived experiences of minoritized communities. The paper grounds these preconditions in existing literature (standpoint theory, Design from the Margins, decolonizing methodologies) and illustrates them through eight Lived Experience Sessions with 13 women and non-binary people of color and five municipal policy workers in a Dutch healthcare context. The sessions used narrative elicitation, co-constructed rule-making, participant-led decisions about AI's role, and a dialogue with policymakers. The authors explicitly frame the empirical work as a worked example rather than a test of the method, yet the abstract and parts of the discussion present the results as demonstrating how the preparatory orientation shapes participatory AI design.","tokens_in":17256,"tokens_out":5342,"duration_ms":55858,"significance":"If the conceptual framework is accepted, AIM provides a useful synthesis of critical participatory AI scholarship into a named, actionable stance. Its emphasis on a preparatory stage before technical framing responds to a real gap identified in the literature, and the seven preconditions offer a checklist that practitioners and researchers could use to reflect on their own processes. The paper also makes a methodological contribution by demonstrating how BNIM and related qualitative techniques can be adapted for this preparatory work. The explicit attention to material conditions of participation and the treatment of refusal as a legitimate outcome are valuable. However, the paper's empirical support is limited to a single illustrative case, and the strength of the claims in the abstract and discussion exceeds what a worked example can establish. As a conceptual proposal, the paper is thought-provoking; as a demonstration, it overreaches.","major_comments":[{"comment":"The analysis reads session notes and transcripts 'alongside AIM's seven preconditions, identifying moments that were especially illustrative of how those preconditions manifested.' This is explicitly a theory-driven interpretive process. The Discussion then elevates the outcome to a 'central observation' that 'each precondition created the conditions under which others could be meaningfully enacted.' This observation is not an independent empirical finding; it is a product of the coding lens, because the analysis was structured to find manifestations of the preconditions. Since the Method also disclaims that the study is designed to test or evaluate the method, the claim of mutual constitutiveness cannot be supported by this design. The authors should either soften the Discussion claim to an interpretive suggestion, or add a non-circular validation, e.g., a comparative case in which one","section":"Analytical approach; Discussion — Preconditions are Mutually Constitutive"},{"comment":"The abstract's final sentence states that participants' reflections 'demonstrat[e] how preparatory orientation fundamentally grounded in lived experience shapes what participatory AI design is for.' The Method section, however, explicitly says the empirical application 'serves as a worked example through which AIM's preconditions are enacted and illustrated, rather than as a study designed to test or evaluate the method's general effectiveness.' These two statements are in tension. A worked example with 13 participants, analyzed through the framework's own lens, illustrates but does not demonstrate. The Conclusion similarly asserts that 'the instantiation also shows that the preconditions, not the techniques, are what travel,' which no single case can show. Please align the framing by replacing 'demonstrating'/'shows' with 'illustrates'/'suggests' throughout, or provide a comparative emp","section":"Abstract; Conclusion"},{"comment":"The paper's motivation rests on the premise that 'stakeholders typically enter a process whose problem definitions, success criteria, and the design of the AI applications are already set' (citing Delgado et al. 2023). This is a load-bearing empirical claim about the current state of participatory AI, but it is accepted from a single source rather than demonstrated in this manuscript. If a substantial share of participatory AI already allows participants to reshape problem framing, the claimed 'methodological need' that AIM fills loses force. Please either provide a more systematic review of current practice, or frame this as a characterization of a dominant tendency rather than a universal fact. This is important because the novelty of AIM rests on this gap.","section":"Introduction, paragraph 3"}],"minor_comments":[{"comment":"In the sentence 'Building on Delgado et al. (2023)) demonstration...', there is a missing opening parenthesis or an extra closing parenthesis. Please fix to 'Delgado et al.'s (2023) demonstration'.","section":"Theoretical Implications"},{"comment":"The mention of a 'major data breach involving women's cervical cancer screening data' is a contextual event with no citation. Adding a reference or a short explanation would help readers outside the Netherlands understand its relevance.","section":"Results, AIM Session 3"},{"comment":"'Lived Experience Sessions' is capitalized throughout as if it were a proper noun. Consider using lowercase or introducing it as a defined term.","section":"General"},{"comment":"The paper states that 'the specific techniques are substitutable, the preconditions themselves are not.' This is a strong claim that is not tested. Consider flagging it as a conjecture to be examined in future work, or at least provide a brief argument for why techniques can vary without undermining the preconditions.","section":"Method, Design of AIM sessions"}],"recommendation":"major_revision","confidential_remarks":"This paper is well-positioned in the critical participatory AI literature and has a clear conceptual contribution. The main issue is the gap between the qualified framing in the Method/Limitations and the stronger claims in the Abstract and Discussion. If the authors are willing to reframe the contribution as a proposal with an illustrative case—rather than a demonstrated empirical result—the paper could be publishable. The circularity of the analysis is a genuine concern, but it is mitigated by the explicit 'worked example' disclaimer; however, that disclaimer needs to be consistently reflected throughout, especially in the abstract. I would encourage the editor to ask for a revised version that tightens the connection between evidence and claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, this is a synthesis paper: all seven preconditions are explicitly traced to prior work, and the authors say they \"systematize into a cohesive whole.\" The genuinely new bit is the packaging as a named stance (AIM) and the sequencing claim that these preconditions must be in place before any technical framing, not woven into an existing pipeline. Second, the body is more careful than the abstract. The abstract says the sessions \"demonstrate\" something; the body calls them a worked example that \"illustrates\" the preconditions.\n\nCredit where due: the paper is clearly written, the limitation section is honest, and the table of preconditions with grounding is useful. The session design is concrete: SQUIN, Rich Picture, co-constructed rules, and a policy-worker dialogue. The discussion of Session 4's failure to produce policy guidelines, reframed as surfacing asymmetry, is thoughtful. The facilitation choice (a woman of color leading, not the white male first author) is well-motivated.\n\nSoft spots, in order of size. First, the abstract overclaims. \"Demonstrating\" implies evidence; the paper's own method section says the study is not designed to test or evaluate the method's general effectiveness. That mismatch should be fixed. Second, the empirical support for \"mutually constitutive\" is weaker than the discussion suggests. The analysis reads transcripts through AIM's preconditions and selects illustrative moments. With that method, you can show the preconditions are compatible, but you cannot observe that they are necessary or mutually constitutive. That claim is a theoretical commitment, not an empirical finding. It should be presented as such. Third, the niche depends on a premise taken from Delgado et al. and Corbett et al.: that current participatory practice fixes problem framing before participants enter. I think that premise is defensible, but the paper doesn't demonstrate it, and if it's false for a large share of practice, AIM's claim to fill a need loses force. Fourth, the sample is small and domain-specific, with no downstream AI outcome. That is fine for a worked example, not for a demonstration.\n\nThe circularity flagged in the stress-test is real but I'd keep it proportionate. For an illustrative worked example, reading data through the framework is less damaging than it would be for a test. The heavier problem is the abstract/body mismatch and the strength of the \"necessary and mutually constitutive\" wording. If those are fixed, this is a solid contribution for people working on participatory AI and AI governance. It gives them a named set of preconditions they can adopt or argue with. I'd send it to referees. I'd also bring it to a reading group, because the sequencing claim is worth debating.","headline":"A careful synthesis of known preconditions for participatory AI with an honestly scoped worked example; the abstract's 'demonstration' and the 'necessary and mutually constitutive' claim outrun the evidence.","tokens_in":17920,"tokens_out":2629,"would_cite":true,"duration_ms":24646,"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":"Participatory AI cannot center minoritized communities unless it begins with their lived experiences before any technical framing.","keywords":["participatory AI","lived experience","minoritized communities","standpoint theory","healthcare AI","decolonizing methodologies","refusal","co-ownership"],"falsifier":"A representative audit of participatory AI projects that finds many allow participants to reset problem definitions, success criteria, and system purpose after joining would falsify the premise that participation is generally bounded by pre-set frames; alternatively, a close reading of transcripts showing that a process without a preparatory stage produces the same system purposes as one with AIM would weaken the claim that sequencing matters.","tokens_in":16881,"feed_emoji":"🤝","tokens_out":5436,"duration_ms":54224,"temperature":0.7,"pith_summary":"This paper argues that participatory AI, as commonly practiced, invites people in only after the questions, success criteria, and system purpose have already been fixed, so their input cannot change what the AI is for. To close that gap, it proposes AI From the Margins (AIM), a methodological stance that names seven preconditions—reciprocity, a decolonizing stance, centering minoritized standpoints, methodological flexibility, structural availability of refusal, multidimensional accessibility, and co-ownership—that must be in place before technical design begins. AIM is not a fixed protocol but a set of standing requirements that can be enacted through different techniques. The paper illustrates these preconditions in eight sessions with women and non-binary people of color about Dutch healthcare, where participants described the engagement as substantive, asked for copies of jointly authored rules, and called for continuation. The central claim is that the preconditions are mutually constitutive: material accessibility enables reciprocity, which enables co-ownership, and the whole sequence is what makes refusal of AI a real option.","feed_headline":"Participatory AI must begin with lived experience, before design goals are set","feed_subtitle":"A new methodological stance argues participation that starts after problem framing is fixed cannot genuinely center marginalized voices.","key_machinery":"The load-bearing mechanism is the AIM stance itself: a set of seven preconditions treated as standing requirements, not a fixed protocol. They are operationalized through a four-session sequence: a single open narrative question that lets participants choose significant events; collective rule-making from those narratives; a session in which participants decide whether, where, and how AI should be involved, with refusal treated as a legitimate outcome; and a final session in which policy workers enter participants' space and the burden of translation falls on them. The sequence is designed to enact the preconditions and to make visible asymmetries between institutional and experiential knowl","core_discovery":"The core discovery is an ordering claim: the order in which knowledge enters a design process is political, and lived experience must enter before any technical framing if it is to shape what an AI system is for rather than merely how an already-scoped system is refined. AIM specifies seven preconditions for that preparatory stage and argues they are not procedural steps but standing, mutually constitutive requirements. The empirical sessions are offered as a worked example, not a controlled test; the evidence is that participants recognized the process as substantive, co-authored rules and asked for their implementation, and used those rules to interrogate policymakers.","pith_inferences":["If the preconditions are mutually constitutive, then a process that implements them independently—say, paying participants but holding sessions in an inaccessible venue—should be expected to fail on all of them; this is a testable prediction.","The paper's sequencing suggests a transferable diagnostic: before any AI framing, check whether participants set the questions, success criteria, and whether AI is involved at all; this could be applied to existing co-design projects as an audit.","A natural next experiment is to run the same preparatory stance in a different domain, such as housing or social benefits, and ask participants the same whether/where/how questions; if refusal rates or agenda shifts differ, that would map how the preconditions interact with institutional context.","The account implies that long-term engagement is not an extra but intrinsic: collective agency formed in the preparatory stage only becomes accountability if participants and policymakers continue meeting, which future work could track."],"forward_implications":["Public-sector AI projects intended for wide use will need to budget a sustained preparatory phase, not a one-off workshop, because the preconditions require repeated relational work.","A participatory process conducted under AIM can legitimately end with no AI at all; refusal of AI is a success condition, not a failure.","Co-constructed rules from such sessions can serve as participant-authored standards for evaluating policy and holding officials accountable.","Evaluation of participatory AI should treat surfaced power asymmetries, such as policymakers' frustration with non-technical language, as a governance outcome rather than a deviation.","Accessibility measures like location, scheduling, and food are not add-ons; they are preconditions that enable reciprocity and co-ownership to exist."],"fun_headline_variants":["AI design should start with lived experience, not preset goals","For AI to center marginalized voices, start before problem framing","Order matters: lived experience must come first in AI design","Participatory AI needs lived experience before technical framing","AIM: let minoritized communities shape what AI is for"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"AIM rests on two premises it borrows from prior work: that participatory AI today normally starts after problem definitions and success criteria are fixed, and that marginalized standpoints make AI's harms legible in a way other positions cannot; if either is false, the claimed gap and the priority ordering lose their footing.","fun_headline_variants_meta":{"raw":{"variants":["AI design should start with lived experience, not preset goals","For AI to center marginalized voices, start before problem framing","Order matters: lived experience must come first in AI design","Participatory AI needs lived experience before technical framing","AIM: let minoritized communities shape what AI is for"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1180,"prompt_tokens":743,"completion_tokens":437,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":370}},"tokens_in":487,"tokens_out":437,"duration_ms":4988,"temperature":1.0,"reasoning_tokens":370,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T12:37:21.681749+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A representative audit of participatory AI projects that finds many allow participants to reset problem definitions, success criteria, and system purpose after joining would falsify the premise that participation is generally bounded by pre-set frames; alternatively, a close reading of transcripts showing that a process without a preparatory stage produces the same system purposes as one with AIM would weaken the claim that sequencing matters.","supporting_citations":[],"review_version":2}