India AI Impact Summit Insights
All sessions

From Co-Design to Courtroom: Building AI for Fairer Justice Systems

Contents

Executive Summary

This talk presents the Fair Trial Advisor, an AI-powered expert system designed to help judges and lawyers navigate the complexity of international fair trial rights by providing instant access to curated legal standards and case law. The project demonstrates a responsible approach to AI in justice systems through co-design methodology, retrieval-augmented generation (RAG) architecture, and extensive stakeholder validation, while acknowledging significant challenges around judicial autonomy, transparency, and potential misuse.

Key Takeaways

  1. RAG Architecture Matters: By grounding responses in pre-curated, authoritative legal texts rather than general training data, the Fair Trial Advisor significantly reduces hallucination risk and enables traceable citations—a design pattern applicable beyond judicial contexts.

  2. Participatory Design Surfaces Real Problems: The hackathon revealed implementation challenges (UI/UX, citation clarity, scope boundaries, misuse prevention) that developers might have missed; stakeholder co-design is essential for AI governance in sensitive domains.

  3. Judicial AI Requires Different Standards: Unlike commercial AI products, tools for justice systems must be transparent, auditable, accountable, and explicitly acknowledge limitations; the shift in framing is from "Can AI answer this?" to "Under what conditions should AI answer this?"

  4. Scaling is Bottlenecked by Institutional Factors, Not Just Technology: Even if the AI works perfectly, adoption depends on judicial training, digital infrastructure, legal text digitization, and jurisdictional political will—technical excellence alone is insufficient.

  5. Transparency & Trust Are Foundational: Users prioritized verifiability and independent oversight over feature richness; tools must document their limitations, update indicators, and submit to external auditing to earn judicial legitimacy.

Key Topics Covered

  • Problem Context: Massive judicial backlogs globally (80M pending cases in Brazil, 50M in India) and the complexity of fair trial rights across 13 component rights interpreted by 25 international bodies
  • The Fair Trial Advisor Tool: Technical design, capabilities, and architecture (RAG-based system grounded in authoritative legal texts)
  • Co-Design Methodology: Hackathon-based participatory design involving judges, magistrates, lawyers, technologists, and civil society representatives
  • User Feedback & Validation: Results from Oxford hackathon testing, including strengths and areas for improvement
  • Ethical Safeguards & Governance: Citation accuracy, transparency, limitations acknowledgment, data privacy, and prevention of misuse
  • Scaling Challenges: Technical infrastructure, digitization of legal materials, jurisdictional variation, training requirements
  • Future Development Roadmap: Domestic integration, multilingual support, offline capability, sectoral expansion
  • Technology Partnership: Collaboration with Microsoft AI for Good Lab and UNESCO frameworks

Key Points & Insights

  1. The Hallucination Problem in Legal AI: Generic LLMs create citation and accuracy risks unacceptable in judicial contexts; the Fair Trial Advisor addresses this through RAG architecture drawing only from curated, authoritative sources (the Oxford textbooks on fair trial and freedom of speech rights)

  2. Co-Design Validated the Concept: Hackathon participants (judges, magistrates, lawyers, technologists) overwhelmingly saw value in the tool for reducing legal research time and providing structured guidance, confirming demand exists

  3. Citation & Traceability Are Non-Negotiable: Participants ranked "clear, complete, and clickable citations" as the top improvement area—judges need traceable chains back to primary sources, not just commentary

  4. The Tool Is Advisory, Not Determinative: Repeated emphasis that the system should support human judgment, not replace it; misuse potential exists (e.g., using it to "play devil's advocate" and circumvent fair trial guarantees)

  5. Domestic Integration is Critical: International fair trial standards alone are insufficient; the tool must be co-designed and integrated with local criminal codes, procedural rules, and legal cultures for adoption and effectiveness

  6. Training Gap Remains Significant: UNESCO survey found 44% of judges use AI tools but only 9% receive training; technological deployment is outpacing institutional readiness and governance frameworks

  7. Accessibility Paradox: Expansion features (multilingual, mobile, voice interaction) are desired but risk over-reliance; balancing accessibility with appropriate judicial caution is unresolved

  8. Data Privacy & Judicial Logging Unresolved: Currently unclear whether interaction logs will be preserved for forensic investigation of judicial errors; this represents a significant gap for accountability

  9. Technical vs. Institutional Constraints: Building scalable systems requires not just LLM improvements but digitization of legal materials, reliable internet infrastructure, and judicial training—technical solutions cannot solve institutional gaps alone

  10. Business Model is Open-Access, Not Commercial: The tool will be free and open-source through Microsoft AI for Good Lab, removing cost barriers but raising questions about sustainability and jurisdictional support

Notable Quotes or Statements

"The technology is moving faster than the safeguards and that's the problem that we're trying to resolve." — Daisy (on the gap between AI deployment and governance frameworks in judicial systems)

"The conversation shifted from 'Can AI answer the question' to 'Under what conditions should the AI answer the question.'" — Daisy (summarizing the ethical reframing from the hackathon)

"It must be clear the whole way through how it's being developed, the things that are being used to provide the answers and where the data is coming from." — Daisy (on transparency requirements for judicial AI)

"You don't want to become too over reliant on what is essentially the commentary in the book. You want the underlying resource." — Audience member (highlighting citation depth problem)

"Participants did not reject AI in justice in principle, but in fact many saw it had strong potential as long as it's clearly done right." — Daisy (on cautious optimism from judicial stakeholders)

Speakers & Organizations Mentioned

  • Daisy Sanda – Primary speaker; appears to lead the Fair Trial Advisor project at the institute
  • Amal Clooney & Philippe Sands – Co-founders of the initiative; authors of the foundational textbooks on fair trial rights and freedom of speech in international law (Oxford University Press)
  • Microsoft AI for Good Lab – Technical partner and development collaborator
  • UNESCO – Providing ethical frameworks for AI and rule of law; conducting judicial training surveys; collaborating on MOOCs about judicial AI
  • University of Oxford – Hosted the co-design hackathon event (January)
  • Oxford University Press – Publisher of source textbooks; license holder
  • Office of Responsible AI – Mentioned as thought partner
  • African Court on Human and People's Rights – Identified as potential piloting site
  • JS Held – Digital investigations expertise mentioned in audience
  • Design Intelligence – Audience participant offering technical solutions for citation/reference challenges

Technical Concepts & Resources

  • Retrieval-Augmented Generation (RAG): Core architecture avoiding hallucination by retrieving from curated datasets rather than generating from general training data
  • Source Texts:
    • The Right to a Fair Trial in International Law (Clooney & Sands)
    • Freedom of Speech in International Law (also integrated into the system)
  • LLM Deployment: Using existing LLMs via Microsoft Azure Foundry (not proprietary models); specific model not disclosed
  • International Legal Framework: 175 states parties to International Covenant on Civil and Political Rights (ICCPR); 13 component rights of fair trial; 25 international human rights bodies; 28,000+ decisions synthesized
  • AI Justice Atlas: Institute's broader initiative mapping AI use in courts across 96 countries
  • UNESCO MOOCs: Planned collaborative educational content on judicial AI and fair trial guarantees
  • Citation System: Planned enhancement to link not just to textbook page references but to original case law sources
  • Proposed Features:
    • Offline/local deployment capability (using open-weight LLMs)
    • Multilingual versions (currently English only)
    • Domestic integration frameworks (jurisdiction-specific criminal codes and procedural rules)
    • Jurisdiction filters and treaty selection
    • Export and sharing functions
    • Mobile optimization
    • Voice interaction capability
    • Misuse prevention safeguards (scope limitation, "devil's advocate" blocking)
    • User interaction logging (status unclear; privacy implications under discussion)

Implementation Status & Roadmap

Current Phase: Prototype/early validation (institute launched October 2023)

Completed:

  • Initial prototype development and demonstration
  • Oxford hackathon (January 2024) with mixed stakeholders
  • User feedback collection and analysis

In Progress:

  • Refinement of UI/UX based on hackathon feedback
  • Citation system improvement to link primary sources
  • Development of governance and transparency frameworks
  • Planning for jurisdiction-specific pilots

Future Priorities (in order of emphasis):

  1. Domestic integration pilots in selected jurisdiction(s)
  2. Multilingual translation and localization
  3. Sectoral expansion (e.g., family law within fair trial context)
  4. Regional court and human rights body piloting (e.g., African Court)
  5. Offline/local deployment capability
  6. Integration with domestic criminal codes and procedural rules
  7. Inclusion of voice interaction and multimodal input (images, video)
  8. Formal adoption guidelines for judicial training

Unresolved Questions & Challenges

  • How to prevent tactical misuse (e.g., using the tool to circumvent fair trial guarantees)
  • Whether interaction logs will be preserved for forensic investigation of judicial errors
  • How to balance accessibility (mobile, voice, multiple languages) with preventing over-reliance
  • How to scope sectoral application (family law, civil law, tax proceedings, etc.)
  • How to ensure technical infrastructure and judicial training exist in lower-resource jurisdictions
  • How to integrate diverse legal traditions and jurisdictional contexts into a coherent system
  • Whether different LLM selection should be user-configurable (status: unclear; testing needed)
  • How to provide offline functionality without compromising accuracy and update frequency