Participatory provenance framework audits AI in public consultation
A novel measurement framework known as participatory provenance, which is based on optimal transport theory, causal inference, and semantic analysis, monitors the transformation, filtration, or loss of individual public submissions during AI-driven summarization. When applied to the consultation for Canada's 2025-2026 national AI Strategy, involving 5,253 participants across two distinct policy areas, the framework indicates that government summaries fall short of a random-participant baseline by -9.1% in both cases. This research highlights the accountability issues in AI-facilitated public consultations, as current methods for AI explainability, grounding, and hallucination detection prioritize output quality over the fidelity of input. No established framework has existed to evaluate whether AI summaries accurately reflect the source population.
Key facts
- Participatory provenance is introduced as a measurement framework for AI-mediated public consultation.
- The framework uses optimal transport theory, causal inference, and semantic analysis.
- Applied to Canada's 2025-2026 national AI Strategy consultation with 5,253 respondents.
- Official government summaries underperform a random-participant baseline by -9.1%.
- Existing AI explainability approaches do not address input fidelity.
- The framework tracks transformation, filtering, and loss of individual submissions.
- Two independent policy topics were analyzed in the consultation.
- The study highlights an accountability gap in AI-mediated public input synthesis.
Entities
Institutions
- arXiv
Locations
- Canada