AI From the Margins (AIM): Rethinking Participatory AI Design Through the Lived Experience of Minoritized Communities

Giovanni Sileno; Laureanne Willems; Lotte Willemsen; Maaike Harbers; Mayesha Tasnim; Roland van Dierendonck; Sennay Ghebreab; Tijs Portegies

REVIEW 3 major objections 4 minor 4 references

AI From the Margins (AIM): Rethinking Participatory AI Design Through the Lived Experience of Minoritized Communities

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Participatory AI cannot center minoritized communities unless it begins with their lived experiences before any technical framing.

desk verdict 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. read the letter →

arxiv 2606.01171 v2 pith:C65QSP6O submitted 2026-05-31 cs.CY cs.AI

Tijs Portegies , Laureanne Willems , Maaike Harbers , Giovanni Sileno , Roland van Dierendonck , Mayesha Tasnim , Lotte Willemsen , Sennay Ghebreab This is my paper · ORCID

classification cs.CYcs.AI

keywords participatoryAIlivedexperienceminoritizedcommunitiesstandpointtheoryhealthcaredecolonizingmethodologiesrefusalco-ownership

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

Credit 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.

Soft 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.

The 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.

Referee Report

3 major / 4 minor

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.

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 (3)

[Analytical approach; Discussion — Preconditions are Mutually Constitutive] 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
[Abstract; Conclusion] 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
[Introduction, paragraph 3] 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.

minor comments (4)

[Theoretical Implications] 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'.
[Results, AIM Session 3] 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.
[General] 'Lived Experience Sessions' is capitalized throughout as if it were a proper noun. Consider using lowercase or introducing it as a defined term.
[Method, Design of AIM sessions] 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.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: AIM's preconditions are synthesized from external scholarship, and the worked example is explicitly not a test; only minor non-load-bearing self-citations.

full rationale

The derivation chain is not circular. AIM's seven preconditions are assembled from external, independently citable scholarship (hooks 1994; Collins 2000; Smith 2012; Udoewa 2022; Sloane et al. 2022), not from the authors' prior results. The paper's own self-citations (Portegies et al. 2026; Villalobos-Quesada et al. 2026) are used only to motivate BNIM as a narrative-elicitation technique and to supply a real-world cardiovascular-risk case; neither is invoked as an authority for the preconditions' necessity or mutual constitution, and no uniqueness claim is imported from the authors' own work. The strongest potential circularity concern is that the empirical sections read transcripts 'alongside AIM's seven preconditions' and then report that those preconditions manifested. That is a theory-driven, illustrative reading, not a test. The paper explicitly disclaims testing: '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.' Thus the abstract's 'demonstrating' language is an interpretive illustration, and the Discussion's 'central observation' of mutual constitution is a qualitative interpretation that would require comparative designs for confirmation. That weakness is an evidentiary/validity limitation, not a circular reduction by construction. No fitted parameter is renamed as a prediction, and no target result is defined in terms of the authors' own prior output. Score 2 reflects only the presence of minor self-citations and an abstract that arguably overstates what a single worked example can show; the central conceptual derivation remains independent of its inputs.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

The central claim rests on no fitted numbers. Its real inputs are: (a) five cited theoretical commitments (standpoint theory, the framing-fixity critique, BNIM validity, intersectionality, mutual constitutivity) — all domain assumptions adopted from prior work or from the authors' own interpretive reading; (b) the worked example's 13 participants and one municipality, whose representativeness is disclaimed. The one named artifact, AIM, has no external falsifiable handle yet. This is the honest cost of a synthesis-plus-case-study contribution. The qualitative analogues of free parameters — which of seven preconditions to extract from the literature, which quotes to present as typifying the sessions, and the decision to treat Session 4's non-output as a methodological outcome — are interpretive degrees of freedom rather than fitted numbers, and are flagged in the soundness and red-flag sections.

assumptions (5)

domain assumption Standpoint theory (Collins 2000; Harding 1991): marginalized standpoints make operations of power and harm legible in ways center positions cannot.
AIM's core priority — centering minoritized lived experience as a foundational input — rests on this epistemic warrant, adopted by citation in Related Literature ('Standpoint theory... provides the warrant for this claim'), not argued in the paper.
domain assumption Mainstream participatory AI practice admits stakeholders only after problem definitions, success criteria, and design artifacts are already set (Delgado et al. 2023; Corbett, Denton, and Erete 2023).
The Introduction states 'Delgado et al. (2023) show that stakeholders typically enter a process whose problem definitions... are already set.' This premise defines the gap AIM fills; if substantially false, the preparatory niche AIM claims dissolves.
domain assumption BNIM elicitation yields lived experience with minimal researcher framing and generates usable requirements for AI in healthcare (Portegies et al. 2026; Wengraf 2001).
Session 1's design and the claim that BNIM outputs can precede participatory design depend on this; the only domain-specific support cited is the authors' own in-press FAccT '26 paper.
domain assumption Intersectionality (Crenshaw 1991) warrants limiting the participant group to women and non-binary people of color and treating lived experience as multi-axis.
Experimental Procedure: 'the call for participants intentionally sought women of color and non-binary people of color... grounded in evidence that these groups have historically experienced minoritization in healthcare.' Adopted as a theoretical commitment.
domain assumption The mutual constitutivity of the preconditions is inferable from the four-session engagement (e.g., accessibility enabling reciprocity enabling co-ownership).
Discussion: 'each precondition created the conditions under which others could be meaningfully enacted.' This is the authors' interpretive reading of the same material used to illustrate the framework, with no comparative or counterfactual check.

invented entities (1)

AI From the Margins (AIM) — a named 'methodological stance' with seven preconditions
purpose: Positions a preparatory, pre-technical-framing stage as the necessary foundation for participatory AI with minoritized communities.
AIM is a synthesis of seven cited preconditions (Table 1), so it is not pulled from a hat; but as a claimed new contribution it carries no falsifiable handle outside this paper — its only evidence is the in-paper worked example and participant self-reports. The paper itself acknowledges future applications must test it (Limitations).

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI From the Margins (AIM): Rethinking Participatory AI Design Through the Lived Experience of Minoritized Communities." pith.science (2026). https://pith.science/paper/C65QSP6O

@misc{pith2026260601171,
  author       = {Pith},
  title        = {Pith review of: AI From the Margins (AIM): Rethinking Participatory AI Design Through the Lived Experience of Minoritized Communities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C65QSP6O}},
  note         = {Machine review of arXiv:2606.01171}
}

read the original abstract

Artificial intelligence (AI) can reproduce and amplify the structural inequities faced by minoritized communities. Participatory AI has been proposed as a response, but participation typically starts after problem definitions and success criteria have been set, leaving limited room for minoritized communities to reshape what an AI system is for. We propose AI From the Margins (AIM): a methodological stance that articulates the conditions under which lived experiences of minoritized communities can be elicited, centered, and carried forward to inform participatory AI design. AIM is not a fixed protocol; it articulates a set of preconditions that can be enacted through different techniques in different settings. We applied AIM in a Dutch healthcare context in eight sessions with 13 women and non-binary people of color and five municipal policy workers, namely through (1) narrative elicitation using the Biographic Narrative Interpretive Method (BNIM); (2) co-constructed rule-making; (3) participants' determination of whether, where, and how AI should be involved; and (4) translating lived experience into AI policy through dialogue with policymakers. In their reflections on the sessions, participants described the engagement as substantive and called for its continuation, demonstrating how preparatory orientation fundamentally grounded in lived experience shapes what participatory AI design is for.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references

[4]

Omar, M.; Soffer, S.; Agbareia, R.; Bragazzi, N

Dissecting racial bias in an algorithm used to manage the health of populations.Science, 366(6464): 447–453. Omar, M.; Soffer, S.; Agbareia, R.; Bragazzi, N. L.; Apakama, D. U.; Horowitz, C. R.; Charney, A. W.; Freeman, R.; Kummer, B.; Glicksberg, B. S.; Nadkarni, G. N.; and Klang, E. 2025. Sociodemographic Biases in Medical Deci- sion Making by Large Lan...

2025
[17]

Mingers, J.; and Walsham, G

Hamburg, Germany: ACM. Mingers, J.; and Walsham, G. 2010. Toward ethical infor- mation systems: The Contribution of discourse ethics.MIS Q., 34(4): 833–854. Obermeyer, Z.; Powers, B.; V ogeli, C.; and Mullainathan, S

2010
[2019]

InCHI’19 Extended Abstracts

Socratic Dialogue as a Method for Moral Inquiry in HCI. InCHI’19 Extended Abstracts. New York, NY: ACM. Harding, S. 1991.Whose Science? Whose Knowledge?: Thinking from Women’s Lives. Ithaca: Cornell University Press. hooks, b. 1994.Teaching to Transgress: Education as the Practice of Freedom. New York: Routledge. Kallina, E.; Bohn ´e, T.; and Singh, J. 20...

1991
[2024]

InProceedings of the 2024 ACM Confer- ence on Fairness, Accountability, and Transparency, FAccT ’24, 631–641

Data Agency Theory: A Precise Theory of Justice for AI Applications. InProceedings of the 2024 ACM Confer- ence on Fairness, Accountability, and Transparency, FAccT ’24, 631–641. New York, NY , USA: Association for Com- puting Machinery. Akinlade, O. 2020. Taking Black pain seriously.New Eng- land Journal of Medicine, 383(10): e68. Benjamin, R. 2019.Race ...

2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [4]

Omar, M.; Soffer, S.; Agbareia, R.; Bragazzi, N

Dissecting racial bias in an algorithm used to manage the health of populations.Science, 366(6464): 447–453. Omar, M.; Soffer, S.; Agbareia, R.; Bragazzi, N. L.; Apakama, D. U.; Horowitz, C. R.; Charney, A. W.; Freeman, R.; Kummer, B.; Glicksberg, B. S.; Nadkarni, G. N.; and Klang, E. 2025. Sociodemographic Biases in Medical Deci- sion Making by Large Lan...

2025

[2] [17]

Mingers, J.; and Walsham, G

Hamburg, Germany: ACM. Mingers, J.; and Walsham, G. 2010. Toward ethical infor- mation systems: The Contribution of discourse ethics.MIS Q., 34(4): 833–854. Obermeyer, Z.; Powers, B.; V ogeli, C.; and Mullainathan, S

2010

[3] [2019]

InCHI’19 Extended Abstracts

Socratic Dialogue as a Method for Moral Inquiry in HCI. InCHI’19 Extended Abstracts. New York, NY: ACM. Harding, S. 1991.Whose Science? Whose Knowledge?: Thinking from Women’s Lives. Ithaca: Cornell University Press. hooks, b. 1994.Teaching to Transgress: Education as the Practice of Freedom. New York: Routledge. Kallina, E.; Bohn ´e, T.; and Singh, J. 20...

1991

[4] [2024]

InProceedings of the 2024 ACM Confer- ence on Fairness, Accountability, and Transparency, FAccT ’24, 631–641

Data Agency Theory: A Precise Theory of Justice for AI Applications. InProceedings of the 2024 ACM Confer- ence on Fairness, Accountability, and Transparency, FAccT ’24, 631–641. New York, NY , USA: Association for Com- puting Machinery. Akinlade, O. 2020. Taking Black pain seriously.New Eng- land Journal of Medicine, 383(10): e68. Benjamin, R. 2019.Race ...

2024