Does participatory AI actually give stakeholders real power?

A mixed-methods study analyzing 80 research papers that described their approach as "participatory AI," combined with semi-structured interviews with 12 of those papers' authors, finds that most participatory AI projects operate in a consultative mode: they elicit stakeholder preferences and values rather than grant stakeholders power over core design decisions. All 80 papers framed their goal as improving user experience, and 52 of 80 as better aligning AI with stakeholders' preferences and values, but only 8 of 80 (10%) involved stakeholders in shaping the system's "scope and purpose." Every project in the corpus involved stakeholders in user interface design — for instance in designing explanations of AI outputs — while only 8 of 80 (10%) let stakeholders weigh in on the type of model and features used, the design of objective or loss functions, or classification decision thresholds. None of the 80 projects allowed stakeholders to rule out AI as a solution altogether, "since these projects were defined as 'AI projects' in the first place."

The paper traces this pattern to who sets the terms of participation. In 74 of 80 projects, project leads chose which stakeholders to involve, often "based upon what they presumed to be the background of that group/person," recruiting either opportunistically ("just talk to some of your friends") or through opt-in. Only 6 of 80 projects used a community-engaged approach to identify participants, and just 3 of 80 kept community stakeholders involved across the AI system's full life cycle. Methodologically, 85% of projects (68 of 80) relied on elicitation techniques — surveys, interviews, role-playing, collaborative storyboarding — built to surface preferences rather than to transfer decision authority.

This sits alongside other recent attempts to typologize human-AI decision authority, such as How should humans and AI systems divide decision-making tasks? and How do AI novelty and scope shape employee relationships?, but shifts the unit of analysis from who decides once a system is built to who has a say while the system is being designed. It also qualifies What UX principles do workplace users want in AI agents?: that framework describes what users want from a deployed agent, whereas this corpus analysis shows that current design processes rarely let stakeholders act on that same wish for control before deployment.

The corpus is limited to published computing-venue papers that self-identify with the "participatory AI" label, so the finding describes that labeled literature rather than participatory practice in AI design generally, and the paper's authors note it may miss relevant work in HCI journals or grey literature such as policy-institute white papers. What the excerpt establishes is a pattern within that labeled corpus, not a causal account of why teams default to consultation; it implies that invoking "participation" in AI design currently signals little about how much design authority stakeholders actually receive.