<aside> âšī¸ This presentation at PAIRS 2026 Online on 17th February 2026 09:45 UTC. Registered participants will receive zoom links to join the session via e-mail.
đī¸ Thread on PAIRS Discussion Server (Discord) (register first)
</aside>
AI-driven legal chatbots are becoming key tools for citizens seeking legal information. These systems are increasingly bridging access to justice gaps and influencing decision-making. Yet, they often operate outside formal regulatory frameworks without proper monitoring, transparency obligations, or checks for bias and error. Existing (and emerging) regulations such as the EU AI Act call for human oversight, but do not specify how it should work in practice or who should be responsible. This presentation proposes a participatory approach and design philosophy to audit legal AI, where community representatives and subject matter experts together monitor and evaluate legal advice tools as a complement to regulatory oversight.
In this presentation, the authors argue that responsible governance of legal AI cannot rely solely on top-down regulation. It views human oversight as an integral and ongoing, socio-technical process rather than a one-time compliance exercise. Drawing from experiences from a case study on building a legal assistant chatbot for underserved communities in India, the authors propose a participatory auditing framework for legal AI with two layers of oversight: first, where human stakeholders monitor AI-supported outcomes in real time, and second, where subject matter experts retrospectively audit legal AI tools. The latter involves a review of system logs, anonymised chat transcripts, and retrieval records (that reflect the reasoning or retrieval-augmented generation pathways shaping AI responses) to understand how specific responses are generated. This would enable a review of system performance, bias, and reliability.
The presentation outlines a procedural design for periodic audits of legal AI tools, supported by user feedback, qualitative surveys, and reviewer training, which can create contextual human scrutiny and cycles of accountability. A diverse, multidisciplinary pool of reviewers strengthens legitimacy and public trust, aligning with emerging global approaches to participatory AI governance.
The authors also recognise key limitations and examine challenges, such as data protection constraints, the reliability of human oversight, and difficulty in establishing shared definitions for consistent auditing. Empirical research shows that human reviewers may lack adequate training, consistency, or incentives, and that defining fairness across legal and technical contexts remains complex. However, this may be mitigated by a structured, multidisciplinary approach creating accountability between overseers including technologists, legal practitioners, and affected communities.
A participatory auditing framework can ensure that automated legal tools are not only more accountable, but also human-centred, and democratically governed.