reportAI in CourtsAI Governance
The bench's record: how courts and adjudicators use AI, what binds them, and where it went wrong
106 instruments on judges' AI use and 11 decisions where an adjudicator relied on fabricated authority: 10 countries bind judges, 22 guide, 6 are drafting.
Edited and verified by Cognesio LLP
Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓
Most of the record on AI in legal practice is about lawyers. This report is about the other side of the bench: what courts, judicial councils and legislatures in 35 countries and territories, plus the European Union and international bodies (CEPEJ and UNESCO), from India and Brazil to Delaware and New South Wales, have told their own judges, tribunal members and court staff about using AI, and the eleven decisions in the SafeLegalAI incident tracker in which the decision-maker, not the advocate, relied on authority that does not exist. As of 5 September 2026 the Regulation Tracker holds 106 instruments that address judicial or court use of AI. Ten countries bind their judges; twenty-two guide them; six are drafting.
The two halves of the record explain each other. The instruments that bind judges arrived after, and usually because of, a decision that went wrong: Colombia’s binding guidelines followed a Constitutional Court review of a judge who used ChatGPT; India’s Supreme Court ruling of July 2026 that a decision resting on hallucinated precedent is no decision in law followed two revenue and tribunal orders and one civil court order built on fake cases; the Philippines’ framework and Spain’s instruction followed reported incidents in their own courts. Every figure below is computed from the regulation, country and incident collections by a Python script over the YAML files (reproduced in the run log; the jurisdiction count is the number of distinct country codes on the 106 records, with Hong Kong counted as a territory and the EU and international bodies counted separately), and the datasets are exported at /regulation/documents.json, /regulation/map.json and /tracker/incidents.json.
Key findings
- As of 5 September 2026 the Regulation Tracker holds 106 instruments addressed to judges, tribunal members or court staff, or governing a court’s own AI deployment: 36 guidance documents, 27 policies, 12 court rules, 7 judgments that set a rule, 6 statutes, 6 regulations, 6 reports, 2 standing orders, 2 ethics opinions and 2 consultations. Ninety-three are in force. Their current versions date from 2018 (1), 2020 (1), 2022 (1), 2023 (8), 2024 (19), 2025 (44) and the first eight months of 2026 (32).
- At country level, 10 of 130 records carry a binding rule on judges’ use of AI (Brazil, China, Colombia, India, Indonesia, Israel, Italy, Nepal, the Philippines and Spain), 22 carry official guidance, 6 a proposal, 1 case law, 77 record nothing and 14 are unclear. Ten countries bar AI from the decision itself: nine of the same ten, with France in place of Indonesia.
- In the United States, 24 state-level instruments address judges or court staff and 7 of the 51 state records carry a binding judicial-use rule (Arizona, Arkansas, California, Delaware, Idaho, New York and Utah); the federal judiciary has interim guidance from the Administrative Office (31 July 2025) and no rule.
- The instruments converge on four rules: the judge decides (stated in some form in every binding instrument), confidential case material stays out of public tools (61 of 106 code confidentiality), output is verified (68 code verification), and the tool is approved or supplied by the institution (Delaware, New York, Utah, Spain, Colombia, Israel, Portugal, the Philippines).
- They diverge on disclosure. Colombia’s Acuerdo PCSJA24-12243 requires a judge who used AI to record in the decision the tool, model, provider, date, purpose and prompts; the Philippines’ framework requires the tool and version, purpose, extent, degree of oversight and a responsibility statement in adjudicatory documents; England and Wales, New Zealand, Canada and most US states require nothing of the kind.
- The incident tracker holds 11 decisions in which the adjudicator relied on fabricated authority: seven in India (six set aside or quashed by the Supreme Court or a High Court; the seventh, a tribunal order, reported recalled by the tribunal itself), two in South Africa (an acting judge referred to the Legal Practice Council; a disciplinary-hearing chairperson), one in Colombia (a superior tribunal’s ruling annulled by the Supreme Court) and one in Georgia, United States (a trial court that adopted a party’s AI-drafted order).
- In none of the eleven did a court impose a personal penalty on the adjudicator. India’s Supreme Court left discipline to the appointing authority; the Karnataka High Court directed that a trial judge’s order be placed before the Chief Justice; South Africa’s full court referred an acting judge to the practitioners’ regulator on the reasoning that acting judges remain practising advocates.
- Two courts have gone the other way and built AI into their own output: Argentina’s Supreme Court generates plain-language summaries for every judgment it publishes, and Taiwan’s Judicial Yuan drafts judgment text for defined criminal offences on an on-premises system, with facts, law and sentence reserved to the judge.
Why this question
The duty on lawyers to verify AI output is now set by rule or judgment in most of the jurisdictions the tracker covers. The duty on judges is newer, less uniform and less visible, because a judge’s use of AI leaves no filing to inspect. It becomes visible in two ways: when an institution writes a rule, and when a decision cites a case that does not exist. This report collects both.
It matters for a practical reason a litigant would recognise. A fabricated citation in a brief costs a sanction and a re-filing. A fabricated citation in a judgment costs the judgment: in India and Colombia in 2025 and 2026, appellate courts set aside or annulled orders on that ground alone, and India’s Supreme Court has said the resulting order is a nullity. The rules for judges are the attempt to prevent that; the incident rows are the count of times it has happened anyway.
Method and data
Three datasets. The regulation collection (324 records) was filtered to instruments whose appliesTo includes judges or court staff, or whose taxonomy categories include judicial use or courts’ own use: 106 records. The country collection (130 records, 76 checked, 54 provisional) supplies the per-country status of two categories, judicial use and AI-decision prohibition, on the six-value scale (binding, guidance, proposed, case-law, none, unclear). The incident collection (150 records) was filtered by hand to decisions in which the adjudicator or official, not a party, relied on non-existent authority: eleven rows, listed in Appendix A. The actor field, which would make this filter mechanical, is coded on 79 of 150 rows; the remaining rows were read.
“Binding” at country level means a statute, court rule, practice direction, regulator rule or administrative order that judges must follow; a judicial council’s guidelines are “guidance” even where compliance is expected. A jurisdiction is credited with barring AI from the decision where an instrument states that the decision, the assessment of evidence or the reasoning is reserved to the human judge, whether by prohibition or by exclusive reservation.
Limits. Provisional country records were researched by AI under the taxonomy rules and not yet re-opened by the editor; they are counted and labelled. Several instruments are recorded from official announcements or press reports where the text is not published (Korea’s guidebook, the Punjab and Haryana directive, the Mexican collegiate-court criteria) and are flagged in the tracker. The incident count is a count of verified decisions the tracker has found, not of occurrences.
What binds judges, and where
| Country-level status, judges’ use of AI | Records |
|---|---|
| Binding | 10 |
| Guidance | 22 |
| Proposed | 6 |
| Case law | 1 |
| None found | 77 |
| Unclear | 14 |
Binding (10). Brazil, through the National Council of Justice’s Resolution 615/2025 (approved 11 March 2025, in force 14 July 2025, replacing Resolution 332/2020): risk classification of court AI, mandatory human supervision of every AI-assisted judicial act, a national committee. China, through the Supreme People’s Court’s Opinions of 8 December 2022: AI is an auxiliary reference that shall not substitute for the judge in any case, and users may decline it. Colombia, through the Superior Council of the Judiciary’s Acuerdo PCSJA24-12243 (16 December 2024): no free commercial chatbots, no AI for assessing evidence or deciding, verification of every output, and disclosure in the decision of tool, model, provider, date, purpose and prompts. India, regionally: the Kerala High Court’s policy for the district judiciary (19 July 2025) bars AI from findings, reliefs, orders and judgments; the Gujarat High Court’s policy (April 2026) does the same; the Punjab and Haryana High Court’s directive (reported 6 April 2026) bans judicial officers from using AI for judgments or research outright. Indonesia, narrowly: the Supreme Court’s letter of 2 September 2026 bars uploading draft-judgment materials to AI services. Israel, through the Courts Administration’s Procedure 01-26 (26 March 2026, recorded from a law journal’s publication of the procedure; the official text has not been obtained): verification of every fact, quotation and reference against authoritative sources, no case-file material in external tools, and public disclosure of the courts’ own AI use. Italy, by statute: Law 132/2025, article 15(1), reserves every decision on interpretation, facts, evidence and measures to the magistrate. Nepal, through the Judiciary’s ICT Policy 2083 (adopted by the Supreme Court full bench): AI may support drafting and research but shall not intrude on judicial decisions. The Philippines, through A.M. No. 25-11-28-SC (18 February 2026) and the 2025 Code of Judicial Conduct: AI may never be the sole, primary or determinative basis of an adjudicatory outcome, judges write their own decisions, no tool is used without En Banc authorisation, and AI use in adjudicatory documents is disclosed. Spain, through the General Council of the Judiciary’s Instruccion 2/2026 (published 30 January 2026): effective human control, no autonomous decision-making, no unapproved tools, and no AI content without personal critical validation.
Guidance (22). The common-law judiciaries sit here. England and Wales’ judicial guidance (version 3, 31 October 2025) makes judges personally responsible for material in their name, bars confidential information from public chatbots, permits summarising and administrative uses but not research or analysis that cannot be independently verified, and tells judges they may ask parties whether AI was used. New Zealand’s guidelines (7 December 2023) have a separate version for judges, judicial officers, tribunal members and judicial support staff. The Canadian Judicial Council’s guidelines (September 2024) state that judges hold exclusive responsibility for their decisions and cannot delegate decision-making, and the Federal Court’s interim principles (29 September 2025) commit it not to use AI to determine outcomes without public consultation. Australia has guidelines from New South Wales (21 November 2024: no generative AI to formulate reasons or assess evidence, no part of a draft judgment into a generative tool), Victoria (14 May 2026) and Queensland (May 2024). Ireland’s Judicial Council guidelines (dated June 2024, published November 2025) permit routine and administrative tasks only. Hong Kong’s judiciary guidelines (July 2024) say generative AI for legal analysis, research or reasoning is not recommended. Korea’s National Court Administration distributed an AI guidebook to all judges on 24 February 2026 (recorded from press reports). Portugal’s Superior Council of the Judiciary adopted recommendations on 8 April 2026 that keep judicial data out of systems the council has not validated. France’s Conseil d’Etat charter (11 December 2025) governs administrative judges; the ordinary courts have no equivalent yet. Chile’s Judicial Ethics Code (Acta 244-2025, article 17) makes AI a support tool and decisions non-delegable. Argentina’s provincial protocols (Rio Negro, Santa Fe, Jujuy) bind their own judiciaries. The Council of Europe’s CEPEJ guidelines (19 December 2025) apply as guidance to the member states of the Council of Europe, and UNESCO’s guidelines (3 December 2025) worldwide.
Proposed (6). South Africa’s Office of the Chief Justice released a draft judiciary AI policy for comment in April 2026 that bars AI from evaluating evidence, assessing credibility, interpreting law, generating substantive reasoning and determining bail, sentence or risk; the July 2026 Judiciary Conference directed that it be finalised. Kenya’s Judiciary announced a draft AI policy and practice directions on 11 August 2025; the record shows no later stage. Ecuador’s Ethics Committee protocol (approved 16 June 2026, awaiting a Judiciary Council resolution) provides that no AI may issue, replace or determine a judicial decision and that judges may not motivate rulings with algorithms. India’s Supreme Court AI Committee published draft regulations for all courts on 3 June 2026 that would generalise the regional prohibitions. Japan’s Supreme Court announced on 28 August 2026 a budget request to pilot a generative-AI system for civil trials, limited to summarising records and organising issues (a January to February 2026 study used mock records). Norway’s court administration was finalising staff guidelines in 2025 and has published no policy.
Case law (1). Mexico’s federal collegiate-court criteria (Semanario Judicial de la Federacion, 22 August 2025): AI may assist a judge only as an auxiliary tool and the final decision rests exclusively with judges and magistrates.
The United States: guidance federally, rules in seven states
The federal judiciary has no rule. The Administrative Office’s AI Task Force issued interim guidance across the judiciary on 31 July 2025, not published in full, directing users to review and independently verify all AI-generated content, cautioning against delegating core judicial functions including decision-making and adjudication, advising extreme caution on novel legal questions, and asking each court to define approved tasks and tools. The Executive Office for Immigration Review’s Policy Memorandum 25-40 (8 August 2025) governs immigration judges and staff.
At state level the tracker holds 24 instruments addressed to judges or court staff, and the country record codes seven states as binding:
| State | Instrument | Date | What it does |
|---|---|---|---|
| Arizona | ACJA 1-509 | 30 Oct 2024, amended 29 Jan 2025 | Code of Judicial Administration section on generative AI use by court personnel: competence, confidentiality, verification |
| Arkansas | Administrative Order 25 | 5 Jun 2025 | Bars court personnel from exposing court-system data to external generative AI tools |
| California | Rule of Court 10.430 and Standard 10.80 | 1 Sep 2025 | Every court permitting generative AI use by judicial officers or staff must adopt a written policy by 15 Dec 2025; policies must bar confidential and sealed data from public tools, require review of output, and disclose public-facing content generated entirely by AI |
| Delaware | Interim Policy | 21 Oct 2024 | Approved tools only; training before use; AI may not make decisions; confidential data stays out of unapproved tools |
| Idaho | Office of Administrative Hearings guidelines | 24 Sep 2025 | Administrative law judges: prohibition, disclosure, confidentiality, verification |
| New York | Interim Policy on the Use of AI | 10 Oct 2025 | Court-approved tools only; training required; confidential, sealed and personal information barred from public tools; all output reviewed |
| Utah | Judicial Council Interim Rules | 27 Oct 2025 | An allow-list: eight named tools permitted (Adobe Firefly, ChatGPT, Claude, Gemini Pro, Microsoft Copilot, NotebookLM, Webex AI Assistant, Westlaw); any tool not listed is prohibited |
Guidance-tier states include Massachusetts (Supreme Judicial Court interim guidelines, December 2025: court personnel only, administrative tasks, public information only, no tools that retain or train on inputs, attorney conduct deliberately untouched), Connecticut (Responsible AI Policy Framework v2.1, February 2024, and the Judicial Branch committee’s final report of 10 August 2026), Louisiana (Supreme Court Technology Commission guidelines, October 2025), Georgia (Judicial Council committee report, July 2025), Indiana (trial-court AI toolkit, February 2026), Illinois (Supreme Court policy, January 2025) and Ohio (ethics guide, June 2026). West Virginia’s Judicial Investigation Commission (Advisory Opinion 2023-22, 13 October 2023) and Michigan’s judicial ethics opinion JI-155 (27 October 2023) are the earliest instruments in the US set.
The pattern across the seven binding states is administrative rather than adjudicative: they govern which tools a court may buy and what data may enter them, and leave the reasoning of the judgment to the general prohibition on delegation. Utah’s allow-list is the most specific instrument anywhere in the record; California’s is the only one that makes the court itself write a policy by a deadline.
Four rules and one divergence
Reading the 106 instruments together, four rules recur in almost every one that binds and most that guide.
The judge decides. Every binding instrument states it: Italy’s statute reserves interpretation, facts, evidence and measures to the magistrate; Spain’s instruction prohibits autonomous decision-making; Colombia’s guidelines bar AI from assessing evidence or motivating a decision; Brazil’s resolution requires human supervision of every AI-assisted act; China’s opinions make AI an auxiliary reference; the Philippines’ framework bars AI as the sole, primary or determinative basis of any outcome; Nepal’s policy bars intrusion on judicial decisions; India’s regional policies bar AI from findings, orders and judgments. The guidance instruments say the same in softer form: the CJC’s “cannot delegate”, CEPEJ’s “a human judge must always take the decision”, England and Wales’ “personally responsible”.
Confidential material stays out of public tools. Sixty-one of the 106 instruments code confidentiality. The formulations differ only in specificity: New South Wales bars any part of a draft judgment from a generative tool; Israel bars case-file material from tools outside the judiciary; Indonesia bars draft-judgment materials from AI services; Massachusetts bars tools that retain or train on inputs; Delaware, New York and Utah solve it by approving tools.
Output is verified. Sixty-eight of the 106 code verification. Israel’s procedure is the most exhaustive (every factual determination, quotation, statutory and case-law reference); the Philippines’ Code of Judicial Conduct, Canon VI section 5, permits AI only for research and requires sources always to be verified before citation; Colombia requires verification of the veracity, relevance and sufficiency of results and sources.
The institution supplies or approves the tool. Delaware (approved tools, training first), New York (court-approved tools, training first), Utah (allow-list), Spain (only systems supplied by the competent administrations), Portugal (no judicial data in systems the council has not validated), Colombia (no free commercial chatbots), the Philippines (En Banc authorisation), Israel (internal tools for case material), Taiwan (on-premises).
The divergence is disclosure. Colombia’s article 10 requires a judge who used AI to record in the decision the tool, model, provider, date, purpose, prompts and the passages reproduced. The Philippines requires, for court-issued adjudicatory documents, the tool and version, purpose, extent of use, degree of human oversight, preservation of the output and a responsibility statement. Israel’s procedure requires the courts’ own AI use to be visible to the public. CEPEJ’s guidelines say judicial use should be transparent to parties where it affects proceedings; UNESCO’s say the same. California requires disclosure only for public-facing content generated entirely by AI. England and Wales, New Zealand, the Canadian Judicial Council, the US federal guidance and every other US state require no disclosure of a judge’s AI use to the parties. A litigant in Bogota or Manila can learn from the judgment whether a model touched it; a litigant in London, Ottawa or Wilmington cannot.
Where it went wrong: eleven decisions
The incident tracker records eleven decisions, all between December 2024 and September 2026, in which the adjudicator or official relied on non-existent authority.
| Date | Decision | Who relied on fabricated authority | What the reviewing court did |
|---|---|---|---|
| 30 Dec 2024 | Buckeye Trust (ITAT Bengaluru, India) | The tribunal’s own order (the tracker records three of its four citations as non-existent, two attributed to the Supreme Court) | Recall of the order and the ChatGPT attribution are reported in the press; neither is read in a primary document; the order itself names no tool |
| 24 Mar 2025 | Sammaan Capital v Mantri (Karnataka High Court, India) | A City Civil Court judge (two Supreme Court judgments never delivered) | Revision allowed; order placed before the Chief Justice; no tool named |
| 30 Jun 2025 | Shahid v Esaam (Court of Appeals of Georgia, US) | A trial court that adopted a party’s proposed order containing two fictitious AI-generated cases | Order vacated; USD 2,500 penalty on the drafting lawyer, not the court |
| 6 Oct 2025 | KMG Wires v NFAC (Bombay High Court, India) | A faceless income-tax assessing officer (non-existent precedents; the court made no finding on how they arose) | Assessment quashed; quasi-judicial authorities warned to cross-verify AI output |
| 5 Nov 2025 | STC17832-2025 (Supreme Court of Justice, Colombia) | A Sincelejo superior tribunal (passages from two cassation decisions that do not exist) | Ruling annulled as a breach of due process; Judicial School ordered to disseminate (as reported by Infobae; the judgment text has not been obtained) |
| 2 Jul 2026 | Pooja Ramesh Singh v J&K Bank (Supreme Court of India) | The NCLT and NCLAT (fake, AI-hallucinated judgments) | Orders set aside; a decision resting even partly on fabricated authority is no decision in law; zero tolerance declared |
| 13 Jul 2026 | Molawa v Matjhabeng (Labour Court, Durban, South Africa) | The chairperson of a municipal disciplinary hearing (two non-existent Labour Appeal Court cases) | Fictitious citations recorded in the judgment |
| 17 Jul 2026 | EPFO v PCIT (Delhi High Court, India) | A Commissioner of Income Tax (Appeals) (six non-existent judgments) | Order set aside as a nullity and a fraud on the justice system; remanded |
| 31 Jul 2026 | FJL v TGO (Gauteng full court, South Africa) | An acting judge (at least eleven fictitious or garbled citations in his judgment) | Findings upheld on their merits; acting judge referred to the Legal Practice Council |
| 20 Aug 2026 | Faiz Enterprise v State Tax Officer (Gujarat High Court, India) | A state tax officer (non-existent and misattributed case law; officer admitted using AI) | Registration-cancellation order quashed; verification instructions issued to all tax officers |
| 2 Sep 2026 | Gadiya v Union of India (Supreme Court of India) | An Additional Commissioner of Customs (an INR 4.25 billion penalty order on non-existent judgments) | Order and the High Court judgment upholding it set aside; discipline left to the appointing authority |
Three things stand out. Seven of the eleven are Indian, and five of those are revenue or insolvency adjudicators rather than judges: the tax officer, the appellate commissioner, the customs commissioner, the assessing officer and the two tribunals. India’s Supreme Court has now decided the question twice in two months, first for tribunals (Pooja Ramesh Singh, 2 July) and then for executive adjudicators (Gadiya, 2 September), and the draft national regulations of June 2026 would carry the prohibition into every court. The pattern the country’s regional policies were written against is the one its reviewing courts keep finding.
Second, the consequence falls on the decision, not the decision-maker. Nine of the eleven orders were set aside, annulled, quashed or recalled; in Molawa the disciplinary hearing was stayed pending review, and in FJL v TGO the appeal was dismissed and the findings upheld on their merits. None of the reviewing courts fined or suspended the adjudicator. The Karnataka High Court sent the order to the Chief Justice; the Supreme Court of India left discipline to the appointing authority; the Gauteng full court referred the acting judge to the Legal Practice Council, holding it the most effective available forum because its jurisdiction survives the expiry of an acting appointment and runs concurrently with the Judicial Service Commission’s, the acting judge having remained an enrolled legal practitioner throughout. Shahid v Esaam is the exception that proves the rule: the fabricated cases were in the trial court’s order because the court adopted a party’s draft, and the penalty went to the lawyer who drafted it.
Third, the tool is rarely on the record. The Gujarat tax officer admitted using AI. Press attributed Buckeye Trust to ChatGPT. The concurring judgment in FJL v TGO said the most plausible explanation for the fictitious Lubbe reference was AI hallucination, adding that what it recorded “does not constitute findings by this Court”. In the other eight the reviewing court recorded that the authorities did not exist and did not say how they came to be cited. The instruments that require disclosure of a judge’s AI use (Colombia, the Philippines) would close that gap; none of the eleven arose in a jurisdiction that had one at the time.
Two courts that built AI into their output
The record is not only prohibitions. Argentina’s Supreme Court, under Acordada 10/2025 (15 October 2025), generates a plain-language summary by AI for every judgment published from the LEX100 system used by federal and national courts. Taiwan’s Judicial Yuan announced on 27 August 2023 a system that drafts judgment text for specified offence types on an on-premises deployment with fact-finding, application of law and sentencing reserved to the judge. Japan’s Supreme Court is requesting funds for a 2027 pilot limited to summarising civil records. Brazil’s Resolution 615 and Colombia’s Acuerdo both permit generative tools within the rules they set. The prohibition, where it exists, is on the decision, not on the technology.
Timeline
| Date | Instrument or decision | Jurisdiction | Position |
|---|---|---|---|
| 4 Dec 2018 | CEPEJ European Ethical Charter | Council of Europe | Ethical principles for AI in judicial systems, adopted at the 31st plenary |
| 21 Aug 2020 | CNJ Resolution 332 | Brazil | Binding national rule on court AI; barred AI from criminal sentencing recommendations |
| 8 Dec 2022 | Supreme People’s Court Opinions | China | AI auxiliary only; never substitutes for the judge |
| 27 Aug 2023 | Judicial Yuan statement | Taiwan | AI drafts judgment text for defined offences; judge decides facts, law and sentence |
| 13 Oct 2023 | JIC Advisory Opinion 2023-22 | West Virginia | Earliest US judicial-ethics opinion on AI |
| 7 Dec 2023 | Chief Justice’s guidelines (judicial version) | New Zealand | Separate versions for the judiciary, lawyers and non-lawyers |
| 2 Aug 2024 | Constitutional Court judgment T-323/2024 | Colombia | Review of a judge’s ChatGPT use; the Superior Council ordered (resolutive point Quinto) to issue guidance within four months of notification |
| Sep 2024 | Canadian Judicial Council guidelines | Canada | Judges cannot delegate decision-making |
| 21 Oct 2024 | Interim Policy | Delaware | First US state judiciary policy: approved tools, training, no AI decisions |
| 21 Nov 2024 | Guidelines for NSW Judges | New South Wales | No generative AI for reasons or evidence; no draft judgment into a tool |
| 16 Dec 2024 | Acuerdo PCSJA24-12243 | Colombia | Binding; disclosure of tool, model, prompts in the decision |
| 30 Dec 2024 | Buckeye Trust | India (ITAT) | First adjudicator row in the tracker |
| 11 Mar 2025 | CNJ Resolution 615 | Brazil | Replaces Resolution 332; risk classification, human supervision |
| 19 Jul 2025 | Kerala High Court policy | India | First Indian High Court policy for the district judiciary |
| 31 Jul 2025 | AO interim guidance | US federal | Verify all output; do not delegate adjudication |
| 1 Sep 2025 | Rule of Court 10.430 | California | Courts permitting AI must adopt a policy by 15 Dec 2025 |
| 25 Sep 2025 | Law 132/2025 | Italy | Statute reserves every decision to the magistrate |
| 27 Oct 2025 | Judicial Council interim rules | Utah | Eight-tool allow-list |
| 31 Oct 2025 | Judicial AI guidance v3 | England and Wales | Personal responsibility; may ask parties about AI |
| 5 Nov 2025 | STC17832-2025 | Colombia | Ruling built on non-existent cassation decisions annulled |
| 19 Dec 2025 | CEPEJ guidelines on generative AI | Council of Europe | A human judge must always decide |
| 30 Jan 2026 | Instruccion 2/2026 | Spain | Binding on judges; no autonomous decision-making |
| 18 Feb 2026 | A.M. No. 25-11-28-SC | Philippines | AI never the determinative basis; disclosure in adjudicatory documents |
| 26 Mar 2026 | Procedure 01-26 | Israel | Binding ethical code for judges and court staff |
| 6 Apr 2026 | Registrar General’s directive | Punjab and Haryana, India | Judicial officers barred from AI for judgments and research |
| 3 Jun 2026 | Draft national regulations | India | Would bar AI from deciding in every court |
| 2 Jul 2026 | Pooja Ramesh Singh | India (Supreme Court) | A decision on hallucinated precedent is no decision in law |
| 31 Jul 2026 | FJL v TGO | South Africa | Acting judge referred to the Legal Practice Council |
| 2 Sep 2026 | Gadiya; Supreme Court letter 112 | India; Indonesia | Executive adjudicators covered; draft judgments barred from AI services |
What to watch
India’s draft regulations are the largest pending change in the record: notified, they would convert three regional policies and two Supreme Court judgments into a binding rule for every court and tribunal in the country, with a disclosure duty on lawyers attached. South Africa’s draft judiciary policy is due for adoption on the Judiciary Conference’s July 2026 resolution, and would put a judicial-conduct rule on the subject alongside the practitioners’ regime the Gauteng full court relied on. Kenya’s and Ecuador’s drafts are at consultation and resolution stage respectively. The Civil Justice Council’s final report for England and Wales, expected in late 2026, concerns filings rather than judges, but the judicial guidance is on its third version in two years and a fourth would follow any change to the parties’ duties.
On the incident side, the question is whether an adjudicator row appears outside India and South Africa with the tool on the record, and whether any reviewing court moves from setting aside the order to disciplining the officer. The Indian Supreme Court’s “zero tolerance” was addressed to the Bar and Bench alike; the tracker will record the first disposition against a bench.
Three sentences journalists can quote
As of 5 September 2026 the SafeLegalAI Regulation Tracker records 106 instruments addressed to judges, tribunal members or court staff in 35 countries and territories plus the European Union and international bodies; ten countries bind their judges’ use of AI, twenty-two guide it and six are drafting.
The incident tracker holds eleven decisions since December 2024 in which the adjudicator, not the advocate, relied on authority that does not exist, seven of them in India; nine of the eleven orders were set aside, quashed or annulled, and no reviewing court penalised the adjudicator.
Colombia and the Philippines require a judge who used AI to say so in the decision, naming the tool; England and Wales, Canada, New Zealand and every US state require nothing of the kind.
Appendix A: the eleven adjudicator rows
| Tracker row | Date | Jurisdiction | Adjudicator | Primary document |
|---|---|---|---|---|
| buckeye-trust-itat | 30 Dec 2024 | India | Income Tax Appellate Tribunal, Bengaluru | ITAT order (archived) |
| sammaan-capital-v-mantri | 24 Mar 2025 | India | City Civil Court judge, Bengaluru | Karnataka High Court order |
| shahid-v-esaam | 30 Jun 2025 | US (Georgia) | Trial court adopting a party’s draft | Court of Appeals of Georgia opinion |
| kmg-wires-v-nfac | 6 Oct 2025 | India | Faceless assessing officer | Bombay High Court judgment |
| Regulation record co-corte-suprema-stc17832-2025 | 5 Nov 2025 | Colombia | Sincelejo Superior Tribunal | Infobae report (judgment text not obtained) |
| pooja-ramesh-singh-v-jk-bank | 2 Jul 2026 | India | NCLT and NCLAT | Supreme Court of India, 2026 INSC 668 |
| molawa-v-matjhabeng | 13 Jul 2026 | South Africa | Disciplinary-hearing chairperson | Labour Court judgment |
| epfo-v-pcit-delhi-hc | 17 Jul 2026 | India | Commissioner of Income Tax (Appeals) | Delhi High Court judgment |
| fjl-v-tgo | 31 Jul 2026 | South Africa | Acting judge, Gauteng Division | Full-court judgment |
| faiz-enterprise-v-state-tax-officer | 20 Aug 2026 | India | State tax officer | Gujarat High Court judgment |
| gadiya-v-union-of-india | 2 Sep 2026 | India | Additional Commissioner of Customs | Supreme Court of India, 2026 INSC 947 |
Appendix B: sources
Instruments, each linked from its Regulation Tracker record with the version URL and date:
- Brazil, CNJ Resolution 615/2025: atos.cnj.jus.br
- China, Supreme People’s Court Opinions (8 Dec 2022): english.court.gov.cn
- Colombia, Acuerdo PCSJA24-12243 (16 Dec 2024): ramajudicial.gov.co; Constitutional Court, Sentencia T-323 de 2024: corteconstitucional.gov.co
- India, Kerala High Court policy (19 Jul 2025): PDF; Supreme Court AI Committee draft regulations (3 Jun 2026): PDF; Pooja Ramesh Singh v J&K Bank, 2026 INSC 668: Indian Kanoon; Gadiya v Union of India, 2026 INSC 947: Supreme Court PDF; KMG Wires v NFAC: Indian Kanoon; Sammaan Capital v Mantri: Indian Kanoon
- Israel, Procedure 01-26 (26 Mar 2026), as published: lawforum.org.il
- Italy, Law 132/2025: Gazzetta Ufficiale
- Philippines, A.M. No. 25-11-28-SC: elibrary.judiciary.gov.ph
- Spain, CGPJ Instruccion 2/2026: BOE
- England and Wales, judicial AI guidance (31 Oct 2025): judiciary.uk
- Canada, CJC guidelines (Sep 2024): cjc-ccm.ca
- New Zealand, guidelines (7 Dec 2023): courtsofnz.govt.nz
- Australia, NSW judges’ guidelines (21 Nov 2024) and Victoria judicial guidelines (14 May 2026): supremecourt.vic.gov.au
- Council of Europe, CEPEJ(2025)18Final: rm.coe.int
- United States: Administrative Office interim guidance (31 Jul 2025), as described in the Director’s annual report: uscourts.gov; California Rule 10.430: courts.ca.gov; Delaware interim policy: courts.delaware.gov; New York interim policy: nycourts.gov PDF; Utah interim rules: utcourts.gov; Massachusetts SJC interim guidelines: mass.gov; Shahid v Esaam: CourtListener
- Argentina, Acordada 10/2025: csjn.gov.ar; Taiwan, Judicial Yuan statement (27 Aug 2023): judicial.gov.tw
- South Africa: FJL v TGO full-court judgment: Charlotin archive PDF; Molawa v Matjhabeng: Charlotin archive PDF
- United States, Shahid v Esaam, A25A0196 (Ga. Ct. App., 30 Jun 2025): opinion PDF, CourtListener storage
- Colombia, STC17832-2025, as reported (judgment text not obtained): Infobae
- SafeLegalAI data: Regulation Tracker · judicial-use category · Incident Tracker · related: The disclosure map, The sanctions ledger, Pooja Ramesh Singh: a ruling built on AI-hallucinated cases
Appendix C: changes to this report
None since first publication. Planned: code actor on the 71 tracker rows that predate the field so the adjudicator filter becomes mechanical; add the Molawa and Colombian rows’ primary documents to the regulation and incident records where only archives or reports are held.