The judiciary of England and Wales has published an institutional statement on using artificial intelligence in judicial work. Its central rule is straightforward: AI may assist, but it must not replace judicial reasoning, evaluation of evidence or the exercise of discretion. Every AI output used in judicial work must be reviewed, verified and approved by the responsible office holder, who remains fully accountable for material produced in their name. Source: Courts and Tribunals Judiciary
The document, dated 9 October, sets institutional guidance; it does not grant blanket permission to introduce any product into a court. It guides senior judiciary working with the Judicial Office, HM Courts and Tribunals Service and the Ministry of Justice. It sets out expectations for adoption, pilots, training, institutional governance, transparency and continuing review. The distinction matters: a statement of principles may shape administrative practice and tool development, but it does not itself change procedural law or authorise the delegation of case decisions.
The work so far, according to the statement, has focused on lower-risk pilots and secure tools in controlled settings. Uses examined include transcription, summarisation, plain-language drafting, document redaction, chronology-building and document analysis. The pilots have tested reliability, the amount of human checking required, the handling of sensitive information and operational safeguards. The statement publishes no accuracy figures, comparative results or names of approved tools. It therefore describes the institution’s perimeter, but does not show how much time was saved or how many errors were prevented.
The judiciary requires approved tools to have UK data residency, enterprise data protection and disabled web access. It does not list authorised systems, given how quickly their features change. New tools must pass approval, evaluation and ongoing review before wider deployment is considered. New use cases go through a central steering group chaired by the Lead AI Judge, which is to assess legal, ethical, operational, security and data-protection issues.
The underlying question is who answers if a summary omits a precedent, changes a date or merges two witness accounts. In justice, a plausible output can acquire authority because it appears inside an institution responsible for explaining coercive decisions. Requiring the judge to check every piece of content keeps the decision with an identifiable person; for the check to be real, time and verification tools must be available. If review becomes a routine box-tick, the rule may preserve the judge’s signature while moving the judgment in practice to an opaque system.
The statement also names judicial independence, impartiality, fairness, confidentiality, security and public confidence. Each criterion needs a practical test. A tool may speed transcription while introducing errors in accents or specialist terms. A redaction function may save time and leave identifying names in a public document. Useful oversight measures such failures by case type and records corrections; it does not stop at a general statement that ‘a human was in the loop’.
The British framework is deliberately limited: administrative support, controlled pilots and non-delegable judicial responsibility. Its immediate value lies in identifying who approves uses and what must be checked before an output is trusted. Its next challenge is to publish pilot results, authorised tools and correction mechanisms. The central rule can be audited: identify who approved the tool, preserve the source for each fact used, and record who checked and approved the output. If an error enters a case file, the process must allow correction before it affects a decision.