The Brief
Issue 16 · 12 October 2026
The judges wrote their own AI rules. Read them as your checklist.
England and Wales publishes the rules its own judges must follow when they use AI. A paralegal's ChatGPT appeal ends in two self-reports. California puts the verification duty into a statute. And two Australian courts show both ends of it: checked AI work is not misconduct, an unchecked AI rationale cannot stand.
The judiciary of England and Wales published its Position Statement on the Judicial Use of AI on 9 October, issued by Sir Colin Birss, President of the Business and Property Division and the judiciary's lead judge for AI. It lets judges use AI to support their work and draws a hard line around the rest: “AI tools must not determine the outcome of a case or displace judicial reasoning, analysis or discretion.”
The conditions are the useful part. Only approved tools may be used, and an approved tool must have UK data residency, enterprise data protection and web access disabled. Any new tool needs proper approval, evaluation and ongoing review before wider use, under the oversight of the senior judiciary. The pilots so far have tested the unglamorous uses: summarisation, transcription, redaction, chronology-building and document analysis. And the judge stays fully accountable: “Every AI output used in judicial work must be reviewed, verified and approved by the responsible judicial office holder.”
This is the courts' own procurement standard, written down. When a judge's tools must keep data in the jurisdiction, stay off the open web and pass a formal approval, a firm appearing before that judge will struggle to argue its own tools deserve less scrutiny. Read the statement as a checklist for your next vendor review.
Why InferisInferis is built to the same shape: answers grounded in your firm's own documents, each with a citation back to the passage it came from, and for EU firms the data and the model both stay inside the EU.
A paralegal's ChatGPT appeal, and two self-reports
The Upper Tribunal dismissed an asylum appeal after finding that the grounds of appeal and skeleton argument, drafted by a paralegal at Sarker Solicitors using ChatGPT, contained fake cases, misquoted cases, and quotations from the lower tribunal's decision that the judge described as “pure invention”.
The panel, Upper Tribunal Judge Keith sitting with Mr Justice Lavender, President of the Chamber, found the supervising solicitor had “not assumed any managerial responsibility” for the paralegal's work. The Home Office was invited to apply for wasted costs, two people have reported themselves to the Solicitors Regulation Authority, and the tribunal left the regulatory questions to the SRA. As reported by the Law Society Gazette, it called the case “one of the more extreme examples of the misuse of AI”.
The paralegal ran the tool. The failure the tribunal named was supervision. Every firm now has someone junior who can produce a polished document in minutes, and the person who signs it off is the one a tribunal will look to.
Why InferisInferis can record every AI query in a hash-chained audit trail, so the person who signs off has a record of what was asked and what came back.
California writes the verification duty into law
Governor Newsom signed SB 574 on 30 September. Filed as Chapter 858, Statutes of 2026, it adds a new section 6068.1 to California's Business and Professions Code, and as an ordinary bill it takes effect on 1 January 2027.
The duties are specific. An attorney “shall not delegate the practice of law to generative artificial intelligence”, must take reasonable steps to verify what it produces, including every case and statutory citation, and must correct anything fabricated. AI use must be disclosed to the court for documents submitted to it, and confidential information must not go into a system where access is not restricted to the attorney and people bound to keep it confidential. The Code of Civil Procedure now bars filings containing citations the responsible attorney “has not personally verified”, and arbitrators may not delegate any part of their decision-making to AI.
Courts have been imposing these duties case by case for two years. California has now put them in a statute with a start date. Eleven weeks is not long to be able to show that every citation in a filing was checked by a person, and that client data never went anywhere it could be read.
Why InferisEvery Inferis answer cites the passage it came from, so checking a citation is a click rather than a search, and answers can carry an AI-disclosure footer.
An Australian court declines to punish checked AI work
In Jahani v Qiu [2026] FCA 1419, decided on 25 September, Justice Cheeseman of the Federal Court of Australia dealt with summaries of more than 7,500 documents that the plaintiffs' solicitors, Norton Rose Fulbright, had prepared with generative AI and then checked by hand. The exercise had grown some fifteen-fold from about 500 documents. The hearing set for 28 September was vacated and relisted for 1 February 2027, and the plaintiffs pay the defendants' costs thrown away.
The more important part is what the court declined to do. Asked whether the Court's practice note on generative AI gave a separate ground for punishing the AI use itself, Justice Cheeseman answered plainly: “I do not accept, however, that GPN-AI provides a separate basis for an adverse costs order in this case.” The costs followed from how late the scale of the exercise was disclosed, not from the use of AI. As the judgment puts it, “the overarching purpose is better served by both parties making appropriate use of available technology.”
That is the other half of every sanctions story in this newsletter. Disclosed, verified AI work is not misconduct. It is a tool the court expects both sides to use, provided someone checks the output and tells the court what was done.
Why InferisThis is the work Inferis is designed for: AI takes the first pass over a large document set, and every summary links back to its source so the review by hand goes quickly.
A murder parole condition built on cases that do not exist
Susan Neill-Fraser served 13 years for the 2009 murder of her partner Bob Chappell before her release on parole in 2022, and she maintains she is innocent. A condition of her parole barred her from speaking to the media to claim her innocence or a wrongful conviction. On 15 September Justice Stephen Estcourt of the Supreme Court of Tasmania ruled the condition invalid: it denied her procedural fairness, was legally unreasonable and created uncertainty.
The court heard that the parole board's secretary had used AI to prepare the rationale behind the condition, that the board conceded it, and that the document cited case law that does not exist. Attorney-General Guy Barnett called the board's professional standards “unacceptable”. By the end of that week the Department of Justice had confirmed it “will undertake a review to determine the extent to which artificial intelligence may have been used to inform past Tasmanian parole board decisions.”
Every other sanction in this newsletter has landed on a lawyer. This one landed on a public decision-maker, and the person whose freedom it restricted could not check the reasoning until her own lawyers did. When an AI-drafted rationale can limit what someone is allowed to say, verifying it is not a professional courtesy. It is the safeguard.
Why InferisInferis is tested to decline when the answer is not in the documents, rather than invent an authority to fill the gap.
Our read
Five decisions, one line running through them. The judges of England and Wales have written down what an acceptable AI tool looks like for their own work: local data, no open web, vetted, and a person who answers for every output. A tribunal has shown what happens when the person in that chain is missing. California has turned the same duty into a statute with a date. And two Australian courts have shown both ends of it: checked, disclosed AI work is simply work, while an unchecked AI rationale cannot stand, even when it comes from a parole board rather than a lawyer. The standard is the same wherever it is being set. Use tools you could defend to a judge, and make sure a named person has checked what they produce before anyone relies on it.
Get the next issue in your inbox