Legal AI safety: a curated reading list of primary sources

This is a curated set of external primary documents on AI safety in legal practice — the judgments, rules, and research this publication draws on. It is pointers to other people’s sources, annotated. Our own coverage lives on the incident tracker; this list is the reading behind it.

Start here: the cases that set the pattern

• Mata v. Avianca (S.D.N.Y., 2023) — the first sanction for AI-fabricated citations, and the origin of the rule that doubling down is the aggravating factor. Tracker record.
• Ayinde v Haringey; Al-Haroun v QNB [2025] EWHC 1383 (Admin) — the English Divisional Court’s anchor ruling, invoking the Hamid jurisdiction. Tracker record.
• Pooja Ramesh Singh v. J&K Bank (2026 INSC 668) — India’s Supreme Court: a ruling on fake precedent is “no decision in the eyes of the law.” Tracker record.
• Guerin v O’Doherty [2026] IECA 48 — the Irish Court of Appeal’s five principles on AI in litigation. Tracker record.

For risk officers and compliance leads

• SRA — Compliance tips for solicitors on AI and technology (updated 9 February 2026). The regulator’s own outcomes-based expectations. See our UK profession guidance history.
• Courts and Tribunals Judiciary — AI Guidance for Judicial Office Holders (v3, 31 October 2025). Written for judges, but the clearest short statement of the risks. See our version history.
• Ministry of Law, Singapore — Guide for Using Generative AI in the Legal Sector (6 March 2026). The most complete state-level sector guide. See our Singapore history.

For litigators: the court rules

• The worldwide comparison of court AI rules — prohibition (NSW), mandatory disclosure (Canada, Ireland), and silence (England, New Zealand).

For regulators and policy

• EU AI Act (Regulation (EU) 2024/1689) and its milestone timeline. The first comprehensive AI statute; note the Digital Omnibus deferrals are adopted but not yet in force.
• UNESCO — Guidelines for the Use of AI Systems in Courts and Tribunals (2025). The global-institution reference for judicial AI.

For researchers: the evidence base

• Stanford RegLab / HAI studies on legal-AI hallucination rates — the empirical measurement of how often general-purpose models and dedicated legal-research tools fabricate. (Figures compiled on our forthcoming hallucination-rate evidence page.)
• Damien Charlotin’s AI hallucination cases database and Matthew Lee’s Natural and Artificial Intelligence in Law blog — the two most comprehensive public trackers. We use both for candidate identification, then verify every case against the primary judgment before recording it; several of their “suspected” entries did not survive that check.

How to use this list

Read the case in the first section that matches your jurisdiction, then the guidance your regulator has issued, then the rule that governs the court you appear in. The through-line across all of it is one duty: verify every authority before it goes before a court. That is the failure the whole record documents.