Negotiating control and accountability among stakeholders in high-risk AI application scenarios: development and evaluation of the ALIGN method
Negotiating Control and Accountability among Stakeholders in High-Risk AI Use Cases: Development and Evaluation of the Method ALIGN
Translation in progress — untranslated text is omitted.
The researchers developed a scenario-based ALIGN method that supports governance and stakeholder participation in high-risk AI development and implementation. The method derives from human factors and AI governance literature, expert interviews and participant observation of a high-risk railway case; it can raise awareness of potential misalignment between accountability and control and its negative consequences, provide a structured path for assessing the degree of alignment between the two, and support design decisions that mitigate AI risk. The paper describes the iterative development and evaluation of ALIGN.
- AI systems are opaque and difficult to control, while safety-critical domains require controlling risk and making accountability clear; mismatches of control and accountability among stakeholders bring safety and governance risks, and existing participatory methods rarely address this problem.
- Based on a real high-risk AI use case—railway freight visual inspection—and combining expert interviews with participatory observation, the authors developed ALIGN: a facilitated workshop of about 2–2.5 hours, comprising a theory introduction, discussion of cross-industry scenario cards, and a control-accountability matching assessment on the AI stakeholder canvas.
- 5 workshops with 20 participants in total (railway human factors experts, AI ethics experts, use-case stakeholders, nuclear regulatory agency employees, etc.). Among the 12 people with quantitative ratings, "the workshop was interesting" was 5.00 and "stimulated discussion of alignment" was 4.83; each scenario in each workshop produced at least 6 design ideas.
- The authors note: the evaluation covers only two domains, railway and nuclear regulation, and needs validation in other safety-critical domains; it measured only feasibility and perceived usefulness (Kirkpatrick levels 1 and 2), not long-term behavior change and outcomes; the quantitative ratings cover only 12 people, and workshops 1 and 5 had only qualitative feedback.