03/01 · Essay
Participatory research has a simple premise. The people who will live with the results of a study should be able to see how it works and change it. Over the years I have watched that premise do real work: a community advisory group that rewrote a consent script until it meant something, mothers who turned a growth chart into a conversation tool because the original was an instrument of blame, frontline workers who told us which indicators they already tracked in their heads so we could stop asking for the ones they did not.
Each of those changes was possible because the method was legible. A consent script can be read aloud. A growth chart can be redrawn. A survey can be argued with.
When the method is a model
A model is not legible in that way. A community advisory board can be shown a dashboard, but it cannot be shown why the dashboard flagged one household and not another, because in most cases the people who built the model cannot show that either. The participatory literature on AI is beginning to notice this. A review of co-research in cancer found that community participation rarely spans the life cycle of an AI project: people are consulted at the start, about the problem, and at the end, about the interface, and absent from the middle, where the data is chosen and the model is built. A framework from a tribal health system in Alaska and a set of ten engagement principles from a diabetes prediction project both try to push participation into that middle. They are honest about how hard it is.
I would add a problem that is specific to the places I work. When a model is trained on routine program data, the people whose records train it were never asked. They consented to care. The participatory question is not only who sits on the advisory board. It is whether the thousands of women whose antenatal records became training data had any way to know, or to say no.
What participation could mean
I do not have a complete answer, but I have a few commitments I try to hold.
People should participate in choosing the data, not only the problem. Which registers, which fields, which years, and which populations are excluded. This is where bias enters and it is where a community’s knowledge of its own records is irreplaceable.
People should see the model’s mistakes before they see its successes. A silent-phase error review with community members and frontline workers on the panel is the most participatory moment an AI project has, because it is the one where a model is genuinely open to being wrong.
People should be able to stop it. A participatory process that cannot end in “no” is a consultation.
The deeper worry
The thing I am most wary of is participation being simulated. A model that drafts the community engagement plan, generates the personas, and summarises the focus groups has not made the research more participatory. It has replaced the participants with a plausible imitation and kept the vocabulary. The test is whether anyone outside the research team could have changed the method. If the answer is no, the method was not participatory, whatever the protocol said.
Sources
- Community engagement at the core of responsible artificial intelligence in health. Journal of Medical Internet Research, 2025. https://www.jmir.org/2025/1/e68198
- Ten principles for engaging communities in AI for diabetes prediction. Journal of Participatory Medicine, 2025. https://jopm.jmir.org/2025/1/e69497
- A framework for community-engaged AI in a tribal health system. Frontiers in Artificial Intelligence, 2025.
- Co-research and participation across the AI life cycle in cancer: a scoping review. 2026. https://www.sciencedirect.com/science/article/pii/S2666915326000909
- Kapania S, et al. Simulacrum of stories: examining large language models as qualitative research participants. CHI, 2025. https://arxiv.org/pdf/2409.19430