On 28 October 2026, Chief Justice Tun Wan Ahmad Farid Wan Salleh told the International Malaysia Law Conference that the judiciary would issue a practice direction "within the next few weeks" on the use of AI by lawyers and court officers in legal proceedings. Judges, he noted, already work under their own guidelines. The courts, he said, are not going to discourage AI use. (The Star/Bernama)
This is the right instinct. A blanket ban on AI in legal practice would be unenforceable and would punish the many lawyers who use it to draft faster and cut costs. But the practice direction will be judged by what it requires, not by its tone. The harder test is the sentencing tool the courts have operated since 2020, which has had far less scrutiny than lawyers' use of chatbots.
The case for tight controls
The strongest argument for strict rules is that courts are different from other institutions. A wrong answer from a chatbot in a marketing email costs little. A wrong answer in a courtroom can cost someone their liberty. The Chief Justice has made the same point, saying in April 2026 that technology must remain "an aid – not a substitute", and warning lawyers to "master AI but don't be mastered by it" (The Star, 11 April 2026). His October warning follows the same logic: reliance on AI can erode the human reasoning that adjudication depends on. Regulators and bar councils elsewhere have also confronted lawyers filing fabricated, AI-generated citations. A court that sees such filings has good reason to demand verification.
This view is persuasive on one point. Lawyers must verify what they submit, and courts should be able to sanction those who do not. A practice direction that says so costs little and protects the process.
Where proportionality should stop
The risk is that the direction regulates the visible problem, lawyers using chatbots, while leaving the consequential one untouched. Verification duties on counsel are cheap to write. They are also the part of the system that is easiest to see.
The sentencing tool is harder to see. Sabah and Sarawak courts have used it since 2020. It is an AI system that recommends sentences from historical cases, and the final decision stays with the judge. Reporting at the time described a narrow pilot covering offences such as drug possession and sexual assault (Malay Mail/Thomson Reuters Foundation, April 2022). In 2024 the system, known as AiCOS, won the ICT category of a national competition (New Straits Times, 19 January 2024). Awards measure innovation. They do not measure whether sentences became fairer.
The public evidence on performance is thin. A 2022 report cited an analysis showing judges followed the AI recommendation in about a third of cases. The same report said the vendor had removed the "race" variable and that the training data covered 2014 to 2019. It also recorded the Bar Council's complaint that when Kuala Lumpur courts adopted the tool in mid-2021 for 20 types of crime, the Bar was "not given guidelines at all" (Malay Mail).
The Oxford Institute of Technology and Justice's AI Atlas, updated to July 2026, shows how the tool behaves in individual cases. In one Kota Kinabalu drug case the AI suggested 10 months and the court imposed 12. In a Kuala Lumpur case the AI range was RM800 to RM1,000 and the court imposed RM1,800. In a 2025 Sandakan High Court case, the 10-month recommendation was upheld. Oxford concludes that judges treat the output as "a reference point" and remain willing to depart from it (Oxford AI Atlas: Malaysia). That is encouraging. It is also anecdotal.
What the practice direction should contain
The judiciary's own 2021 media statement on guidelines for AI-assisted sentencing in the Peninsular Malaysia subordinate courts is listed on its website, but the document is not readily searchable or machine-readable (Judiciary of Malaysia, 22 July 2021). If the courts now explore AI-assisted sentencing guidelines for the Peninsula, the rules should be easy to find and read.
A proportionate package would have four parts:
- Verification duty with graduated sanctions. Counsel certify that cited authorities were checked against primary sources. Honest mistakes caught early draw a warning. Repeated or concealed fabrication draws costs or referral.
- No blanket disclosure of every AI use. Spell-check-level assistance should not need a footnote. Disclosure should attach to AI-generated legal citations and evidence, where reliability matters.
- Published outcome data for the sentencing tool. The judiciary should release anonymised figures on how often judges follow, raise or lower the recommendation, by offence and by court. Without that, no one can say whether the tool narrows disparity or entrenches old patterns.
- A right to know and to challenge. Defence counsel should be told when a recommendation was generated and should be able to argue against it on the record.
None of this requires banning anything. It requires the judiciary to apply to its own tool the evidence standard it asks of lawyers.
The bottom line
Malaysia's courts are choosing engagement over prohibition, and that is the correct call for a system with heavy caseloads and an innovation-friendly economy. The Chief Justice's concern about eroding human reasoning is real, but it is best answered with transparency, not restriction. If the new direction binds lawyers while the judiciary's own AI remains unmeasured, it will look like regulating the weakest party. Publishing the data would let Malaysia show that judicial AI can be both useful and accountable.