Twenty-four entries, read at source
AI rules for lawyers, country by country
Court rules, practice directions, government guidance and decided cases on the use of artificial intelligence in legal work, by jurisdiction
Last reviewed 5 August 2026. Twenty-four jurisdiction entries tracked, twenty-three with published pages.
Twenty-four entries across nineteen jurisdictions. Every linked entry below was read at its issuing body’s own document, not from a summary. Every row links to a page. The one exception is Saudi Arabia, where we searched and found no lawyer-facing rule at all, and we say so rather than manufacture a row.
Every jurisdiction we track
Ordered by how directly the instrument binds a lawyer filing a document, not alphabetically.
| Jurisdiction | Instrument | Kind | Status | Date |
|---|---|---|---|---|
| AustraliaAsia Pacific | Supreme Court of NSW, Practice Note SC Gen 23. Gen AI banned from affidavits and witness statements; a positive disclosure that it was not used; experts must annex a record of their prompts | Court rule | In force | 3 Feb 2025 |
| QatarMiddle East | QICDRC Practice Direction No. 1 of 2026. 27 paragraphs. Prohibits AI-generated evidence and entering privileged material into AI tools except on private secure platforms; deepfakes must be disclosed | Court rule | In force | 6 Jan 2026 |
| United StatesNorth America | Connecticut Practice Book § 4-9, a statewide rule with nonsuit on the table; nineteen verified standing orders; sixteen sanctions decisions | Court rule | In force | Ongoing |
| Canada, OntarioNorth America | Superior Court of Justice, AI provisions in all three consolidated provincial practice directions: civil, family and criminal | Court rule | In force | 17 Mar 2026 |
| SingaporeAsia Pacific | Ministry of Law, Guide for Using Generative AI in the Legal Sector, 52pp; plus the courts’ Registrar’s Circulars of 23 Sept 2024. The only instrument anywhere that addresses billing for AI tools | Government guidance | In force | 6 Mar 2026 |
| England and WalesEurope | AI Guidance for Judicial Office Holders, third edition. First official warning about hidden “white text” prompt injection in filed documents | Judicial guidance | In force | 31 Oct 2025 |
| European UnionEurope | Regulation (EU) 2024/1689, the AI Act, Article 50 transparency duties; plus the CCBE guides of 2 Oct 2025 and 27 Mar 2026, which reach every European bar | Statute and bar guidance | Applies | 2 Aug 2026 |
| GermanyEurope | No court rule. BRAK guidance (Dec 2024) and a DAV position paper (Apr 2026). Both permit AI use subject to verification of cited sources | Bar guidance | In force | Apr 2026 |
| IrelandEurope | Courts Service guidance on AI in court proceedings (12 Feb 2026); Law Society of Ireland Practice Guidance (20 May 2025). The Courts Service also published that its own judicial AI assistant was not fit for purpose | Court and bar guidance | In force | 12 Feb 2026 |
| England and WalesEurope | Ayinde v Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin). Hamid jurisdiction, wasted costs, referrals to both regulators | Case law | Decided | 6 Jun 2025 |
| Canada, British ColumbiaNorth America | Zhang v. Chen, 2024 BCSC 285. Special costs refused; personal costs ordered under Family Rule 16-1(30) instead, plus an order to review every other file within 30 days | Case law | Decided | 20 Feb 2024 |
| Canada, Federal CourtNorth America | Practice Notice and interim Principles on AI. Requires a declaration where AI generated content in a court document, and verification against trusted sources such as CanLII | Court rule | In force | 20 Dec 2023 |
| Canada, nationalNorth America | Canadian Judicial Council, Guidelines for the Use of Artificial Intelligence in Canadian Courts. AI cannot supplant judges’ exclusive responsibility for decision-making | Judicial guidance | In force | Sept 2024 |
| New ZealandAsia Pacific | Courts of New Zealand, three sets of GenAI guidelines: for lawyers, for judges and judicial officers, and for non-lawyers | Judicial guidance | In force | Dec 2023 |
| Hong KongAsia Pacific | Judiciary Guidelines on the Use of Generative AI for judges, judicial officers and support staff. Rule 1: no delegation of judicial functions to AI | Judicial guidance | In force | Jul 2024 |
| United Arab EmiratesMiddle East | DIFC Courts, Practical Guidance Note No. 2 of 2023 on large language models and generative content generators. Requires disclosure of AI-generated content, its source, and the system’s limitations | Court guidance | In force | 2023 |
| ChinaAsia Pacific | Supreme People’s Court, Opinions on Regulating and Strengthening the Applications of AI in the Judicial Fields; and the Xiamen Maritime Court’s Guidelines for Litigation Participants’ Use of AI (trial), which require full disclosure of AI use | Judicial policy and court guidance | In force | 2022, 2025 |
| BrazilLatin America | CNJ Resolution No. 615/2025, revoking Resolution 332/2020. Risk tiers for judicial AI, mandatory registration and audit via the Sinapses platform, and a National Committee for AI in the Judiciary | Judicial regulation | In force | 11 Mar 2025 |
| IndiaAsia Pacific | Kerala High Court, Policy Regarding Use of AI Tools in District Judiciary, the first binding Indian judicial policy; plus the Supreme Court of India’s draft Regulations, consultation closed 15 July 2026 | Court policy; draft rule | Kerala in force; SC draft | 19 Jul 2025 |
| South AfricaAfrica | Mavundla v MEC and the Northbound Processing matter. Seven of nine authorities fabricated in the first; referral to the Legal Practice Council; conduct called “irresponsible and unprofessional” | Case law | Decided | 2025 |
| Saudi ArabiaMiddle Eastno lawyer-facing rule found | No court rule for lawyers found. SDAIA Generative AI Guidelines (1 Jan 2024) for government employees and for the public, plus the PDPL and AI ethics principles. No dedicated stand-alone AI statute | Government guidance only | No lawyer-facing rule found | Jan 2024 |
| Canada, OntarioNorth America | Ko v. Li, 2025 ONSC 2766 / 2965 / 6785. A purged contempt reopened as criminal contempt, carriage referred to the Attorney General of Ontario | Case law | Live and undecided | 4 Dec 2025 |
| Australia, VictoriaAsia Pacific | Victorian Legal Services Board variation of a practising certificate. The only published regulator decision anywhere: two years supervised practice, quarterly reporting | Regulator decision | Imposed | 19 Aug 2025 |
| England and WalesEurope | Bandla v SRA [2025] EWHC 1167 (Admin) and Harber v HMRC [2023] UKFTT 1007 (TC). An appeal struck out as an abuse of process; and the earliest UK instance, a litigant in person | Case law | Decided | 2023, 2025 |
What every one of them agrees on
No two use the same words. None of the four duties below is satisfied by a policy document.
- VerifyIndependently, against primary sources. Qatar names “primary sources or recognised legal databases”; England and Wales names government databases, The National Archives and the Law Reports; Ontario requires citations hyperlinked to CanLII; New South Wales adds that the verification “must not be solely carried out by using a Gen AI tool”.
- SuperviseDelegation does not transfer the duty. Singapore states it in advance, expressly naming paralegals and legal secretaries. United States v. Farris is what it looks like when a court finds it breached.
- ProtectPrivileged material must not enter a tool that will not hold it. Qatar and New South Wales both write a controlled-environment test into the rule itself.
- ProveThe one nobody can meet from memory. New South Wales requires the prompts annexed; Qatar can require an affidavit; Connecticut makes filing itself the certification; British Columbia ordered counsel to review every other file and report within thirty days.
Where nothing has happened yet
Saudi Arabia is the clearest example, and worth stating plainly because the assumption runs the other way. SDAIA has issued Generative AI Guidelines for government employees and for the public, dated 1 January 2024, and the Kingdom has a Personal Data Protection Law and published AI ethics principles. There is no dedicated stand-alone AI statute and we have found no court rule or bar guidance directed at lawyers. If one exists we have not located it, and we will not invent one to fill a row.
The same honesty applies to Scotland and Northern Ireland: senior judicial statements, courts already using AI for transcription and translation, and no published practice note binding practitioners that we could find.
Common questions
Which countries have rules on lawyers using AI?
As of August 2026 we track twenty-four entries across nineteen jurisdictions. The ones with instruments that bind filers directly are Australia (New South Wales), Qatar, the United States (Connecticut and individual federal judges), Canada (Ontario and the Federal Court) and China (the Xiamen Maritime Court). Singapore, England and Wales, Hong Kong, New Zealand, Ireland, Germany, Brazil, India and the European bars have guidance rather than binding filing rules. Saudi Arabia has no lawyer-facing court rule that we have found.
Which country has the strictest AI rules for lawyers?
New South Wales, Australia. Practice Note SC Gen 23 bans generative AI from affidavits and witness statements outright, requires a positive declaration that it was not used, forbids verifying AI output with another AI tool, and requires expert witnesses to annex a record of the prompts, default values and variables they used. No other court asks for the prompts.
Do I have to tell a court I used AI?
It varies, and the pattern is consistent: rarely a pre-emptive declaration, almost always an ability to prove verification on demand. Canada's Federal Court requires a declaration where AI generated content. The DIFC Courts in Dubai require disclosure of AI-generated content and its source. Qatar requires no declaration unless asked, but you must be able to identify which portions used AI and explain how you verified it, on affidavit. New South Wales requires an affidavit to state that AI was not used.
Can I put privileged client material into an AI tool?
Qatar prohibits it unless the platform is private and secure. New South Wales prohibits entering suppressed, subpoenaed or Harman-protected material into any generative AI program unless it stays in a controlled environment and is not used for training. England and Wales tells judges that anything entered into a public chatbot should be treated as published to all the world. Singapore treats it as a professional confidentiality duty and names prompt injection as a risk.
Is AI in the courts only a United States problem?
No. The most detailed court instruments in the world are Australian and Qatari, not American. Sanctions decisions have been handed down in England and Wales, Canada and South Africa. Brazil regulates judicial AI by national resolution, and China's Supreme People's Court has had a policy since 2022.
What do all these rules have in common?
Four duties. Verify AI output independently against primary sources. Supervise anyone you delegate to, because delegation does not transfer the duty. Keep privileged material out of tools that do not hold it. And be able to prove all three afterwards, which is the duty no policy document satisfies and no lawyer can meet from memory.
The common answer
Four duties, twenty-four entries, one shape: be able to show what the machine produced, what a person checked, and when. A record sealed at the time of the work satisfies all four, everywhere on this page, and it can be made on your own machine so that nothing privileged ever leaves the building.