AI in court filings: who must abstain, declare, certify or verify, and who stays silent, in 130 jurisdictions

As of 5 September 2026 the SafeLegalAI Regulation Tracker records 324 official documents on AI in legal practice from 130 countries and entities, and 114 of them are court-facing instruments: practice directions, court rules, standing orders, judgments that set a rule, and guidance addressed to filers or judges. This report codes every one of them, and every country record, on a single question: when a lawyer or a litigant puts AI-assisted material before a court, what does that court require?

The answer divides into five positions. A court can prohibit AI for a class of document, require a declaration that AI was used, require a certification that the content was verified, impose a bare duty to verify, or say nothing and rely on the duties that already bind every advocate. Six countries have a binding instrument in the first three positions; ten have a binding verification duty; the rest are guidance, case law or silence. The United States has all five positions at once, judge by judge and state by state. Every figure below is computed from the regulation and country records by the script in the run log, and the datasets are exported at /regulation/documents.json and /regulation/map.json.

Key findings

1. Of the 114 court-facing instruments, 59 are binding in form (court rules, practice directions, standing orders, statutes or regulations) and 46 of those address litigants or parties as well as lawyers; 42 of the 46 are in force. The 59 instruments’ current versions date from 2023 (8), 2024 (10), 2025 (17) and the first eight months of 2026 (24).
2. Among the 46 binding filing instruments, 39 carry a coded verification duty, 27 a disclosure requirement, 16 a certification, 8 a prohibition of some kind (5 of them for a class of document) and 7 a record-keeping duty; most carry more than one.
3. At country level, 6 of 130 records carry a binding filing-disclosure or certification position (Australia, Canada, Ireland, the Philippines, Qatar and the United States), 3 carry guidance (Israel, the United Arab Emirates, international arbitral bodies), 4 carry a proposal (Ecuador, India, Kenya, the United Kingdom), 6 are unclear and 111 record nothing.
4. Verification is the wider duty: 10 countries bind filers to verify AI output (the six above plus Poland, Singapore, Slovakia and Spain), 28 have official guidance to that effect and 11 rely on case law alone.
5. Rules for judges outnumber rules for filers. Ten countries bind their judges’ use of AI (Brazil, China, Colombia, India, Indonesia, Israel, Italy, Nepal, the Philippines and Spain) and ten bar AI from the decision itself; only six bind what a party files.
6. Two jurisdictions have chosen silence on the record. England and Wales’ Civil Justice Council reported on 30 June 2026 a consensus against any AI-specific requirement for professionally drafted documents; New Zealand’s court guidelines of 7 December 2023 impose no disclosure unless the court asks. Singapore’s Registrar’s Circular takes the same line as a binding instrument.
7. The instruments disagree on the direction of the declaration. New South Wales and Ireland require a statement that AI was not used for sworn evidence; Canada’s Federal Court, Manitoba and Yukon require a statement that it was; Montana’s Fourth District treats the absence of a certificate as a certificate that it was not.
8. Seven instruments carry more than one version and two records did not survive in their original form: New South Wales amended SC Gen 23 within ten weeks of issue, Canada’s Federal Court updated its notice in May 2024, Victoria replaced its 2024 guidelines with a practice note in May 2026, and the Fifth Circuit decided in June 2024 not to adopt its proposed circuit-wide certification rule.

Why this question

A firm that files in more than one court cannot answer “do we have to say we used AI?” with one policy. The instruments that answer it are scattered across practice directions, local rules, individual judges’ standing orders, professional-conduct rules and judgments, and the pages that collect them are mostly published by vendors, who paraphrase one law-firm tracker of United States judge-level orders and stop at the border. This report starts from the other end: the 324-document tracker and the 130-country map, coded on one schema, with the instrument’s own words behind every cell.

The question has become sharper since June 2023, when the first five instruments appeared within four weeks of each other in Texas, Illinois, Pennsylvania, Manitoba and Yukon. The incident tracker now holds 150 decisions in which fabricated authority reached a court, and the sanctions ledger records what those courts ordered. The disclosure map is the other half: what courts asked for in advance.

Method and data

Two datasets, both exported from this site. The regulation collection holds 324 YAML records as of 5 September 2026, each with an issuing body, jurisdiction, instrument type, status, the persons it applies to, a coded set of requirements (disclosure, certification, verification, prohibition, confidentiality, competence, supervision, record-keeping, consent, permissive, risk-classification), versions with dates and URLs, and taxonomy categories from /regulation#categories. The country collection holds 130 records; each scores twenty categories with one of six statuses (binding, guidance, proposed, case-law, none, unclear) and a dated note. Seventy-six country records are checked; 54 are provisional, meaning AI-researched under the taxonomy rules and not yet re-opened source by source by the editor. Provisional records are counted and labelled, never silently merged.

“Court-facing” instruments are those carrying the taxonomy category of filing disclosure, or an instrument type of court rule, practice direction, standing order or judgment guidance: 114 records. (A wider filter that also admits every document with a verification, judicial-use or evidence category returns 236 records; those are bar guidance and judicial guidance, analysed in the country map below but not counted as court instruments.) “Binding” means the instrument type is a court rule, practice direction, standing order, statute or regulation: 59 records. “Binding filing instruments” are the binding records whose scope includes litigants or parties: 46, of which 42 are in force; one is a draft under consultation (India), one a proposal (Federal Rule of Evidence 707), one withdrawn (the Fifth Circuit’s Rule 32.3) and one superseded by a statewide rule (Florida’s 17th Circuit order, per the Supreme Court of Florida’s commentary of 28 May 2026). Guidance addressed to judges, bar ethics opinions and reports are analysed but not counted as binding on filers.

The five positions used in this report are the report’s reading of each instrument’s operative text, not a field in the data. Prohibition means AI may not be used to produce a class of document. Declaration means the filer must state that AI was used, either always or when the court directs. Certification means the filer must attest that the content or citations were verified. Verification means a duty to check without any statement to the court. Silence means no AI-specific requirement for filers, with existing duties relied on. Where an instrument sits in more than one position it is counted in each.

Limits. The United States is coded at country and state level, but individual judges’ standing orders are recorded only where the tracker has read them (8 federal instruments); no census of judge-level orders is attempted, and the US standing-orders registry explains why. Four state instruments (New York Part 161, Florida Rule 2.515, Louisiana Act 250 and Oklahoma Rule 1.17) were recorded from their announcements and re-read from archived copies of the operative text on 5 September 2026; the text reads narrower than the announcements in two cases, which the report says where it matters. Florida’s 17th Circuit order has not been read and is carried as coded. Country notes were written between 3 and 5 September 2026 and carry that date.

What the 46 binding filing instruments require

Requirement coded	Instruments (of 46)
Verification	39
Disclosure	27
Certification	16
Prohibition	8
Record-keeping	7
Confidentiality	5
Supervision	3
Permissive statement	3
Competence	2

By jurisdiction: Canada 12, United States (state) 12, Australia 9, United States (federal) 7, and one each for Ireland, Singapore, Qatar’s financial-centre court, the Philippines, India (a draft) and the JAMS arbitration rules. Thirty-nine of the 46 carry a verification duty in the tracker’s coding, which makes verification the common floor of the instruments that address filers: whatever else a court asks, it asks the filer to check.

The other three requirements are where courts diverge, and the divergence is not about whether AI is allowed. None of the 46 bars AI from filings outright. It is about what the court wants to know, and what the filer must be able to swear.

The five positions, instrument by instrument

Prohibition: no AI for sworn evidence

Eight instruments carry a coded prohibition. In five the prohibited thing is a class of document, and in every one of the five the class is evidence rather than argument; the other three prohibit something else (Arkansas bars court staff from entering court data into external AI tools; the Indian draft and the Philippine framework bar AI from deciding). The Supreme Court of New South Wales’ Practice Note SC Gen 23 (issued 21 November 2024, amended 28 January 2025, in force 3 February 2025) provides that generative AI must not be used to generate the content of affidavits, witness statements or character references, that affidavits and witness statements must contain a disclosure that generative AI was not used, and that expert reports may not use generative AI without leave. Submissions and summaries may use it, provided all citations are verified. The NSW Local Court practice note (12 January 2026), the NSW Civil and Administrative Tribunal’s Procedural Direction 7 (7 March 2025) and the Personal Injury Commission’s PIC13 (1 January 2026) carry the same structure into the state’s lower courts and tribunals.

Ireland’s Practice Direction HC 142, signed by the President of the High Court on 29 July 2026 and in operation from 1 September 2026, takes the same position for substantive content: affidavits and witness statements must carry a declaration that generative AI was not used to generate substantive content, expert reports need either that declaration or a disclosure of AI use, and every AI-assisted citation or proposition must be independently verified by a human, with the direction stating expressly that checking with another AI tool is insufficient. Its sanctions clause lists rejection of the document, adverse or personal costs, strike-out and referral to the regulator.

The rationale is the same in Sydney and Dublin: a sworn document must be in the deponent’s own words. Victoria’s SC GEN 25 (14 May 2026) states the principle without the prohibition, requiring affidavits and witness statements to be in the deponent’s own words while leaving submissions to a verification-and-direction regime. Outside Australia and Ireland, the only prohibition of this kind noted anywhere in the record is a single United States judge’s standing order barring AI in the preparation of filings, listed in the US registry but not held as a regulation record; no statewide or national rule prohibits AI for any class of document.

Declaration: say that you used it

Twenty-seven instruments require a disclosure, and they split into two kinds. The first is an automatic declaration. The Federal Court of Canada’s notice (20 December 2023, updated 7 May 2024) requires a declaration in the first paragraph of any document containing AI-generated content, applies to counsel and self-represented litigants alike, and adds that the court itself will not use AI in decision-making without prior consultation. Manitoba’s King’s Bench practice direction (23 June 2023) requires materials to indicate how AI was used; Yukon’s GENERAL-29 (26 June 2023) requires the tool and its use to be identified; Nova Scotia’s Provincial Court directive (27 October 2023) and the Registrar in Bankruptcy’s directive (4 October 2024) require disclosure of AI use in filings. In the United States, Judge Baylson’s standing order (E.D. Pa., 6 June 2023) requires a clear statement that AI was used plus a certification that every citation was verified; Judge Graham’s order (S.D. Ohio, 21 October 2025) requires a declaration naming the platform and describing what content was AI-generated; Shawnee County’s Rule 3.125 (Kansas, 14 June 2024) requires a clearly marked first-page statement; Montana’s Fourth Judicial District Rule 3.G requires the tool to be named and its use explained.

The second kind is a declaration on direction: the court may ask, and the party must then answer. The Federal Court of Australia’s GPN-AI (16 April 2026) provides that the court may direct a party to disclose whether and how generative AI was used and may make case-specific orders limiting it, while its paragraph 4.10 requires disclosure without any direction, in the body of the document, where generative AI was used to summarise or analyse material a witness relies on, to create images or video, or in any other way that might affect admissibility; the Federal Circuit and Family Court’s PD-AI (1 May 2026) follows it; Victoria’s SC GEN 25 lets the court direct a party to identify which parts of a document were AI-produced; Western Australia’s guidelines make disclosure a matter of direction; the Qatar International Court’s Practice Direction No. 1 of 2026 (6 January 2026) allows the court to require disclosure of how AI was used. The DIFC Courts’ Practical Guidance Note No. 2 of 2023 (21 December 2023) sits between the two: parties must declare use or intended use at the earliest opportunity, but the note is guidance rather than a rule.

The declaration regime has an inversion problem that the instruments resolve three ways. Canada’s Federal Court wants to be told AI was used. New South Wales and Ireland want to be told it was not, for sworn evidence. Montana’s Fourth District provides that a filing without the certification is presumed to certify that no part was prepared with generative AI, so silence is itself a statement. A lawyer filing the same affidavit in Ottawa, Sydney and Missoula would make three different declarations about the same fact.

Certification: swear that you checked

Sixteen instruments carry a certification code, and the trend since late 2024 is toward certification without any declaration of use. Ontario’s Rule 4.06.1(2.1) (in force 1 December 2024) requires every factum to certify that the signer is satisfied as to the authenticity of every authority cited, with authorities from CanLII, official court sites or recognised reporters presumed authentic; it does not ask whether AI was used. Queensland’s Practice Direction 5 of 2025 (24 September 2025) requires every written submission to name the individual responsible for its contents, naming a firm being insufficient, and makes that person responsible for verifying every reference; practitioners responsible for non-existent material may be referred to the Legal Services Commissioner or required to show cause against a personal costs order. Connecticut’s Practice Book section 4-9 (effective 23 June 2026) provides that anyone who uses generative AI to create or edit a filed document must independently verify every citation, authority and item of evidence, that filing constitutes a certification of diligent compliance, and that non-compliance may be sanctioned by nonsuit or default.

Three 2026 statewide rules in the United States sit near this position, and their operative texts, read from archived copies on 5 September 2026, are in two of the three cases narrower than their announcements. New York’s 22 NYCRR Part 161 (all Unified Court System courts, civil and criminal) states as statewide policy that attorneys and parties should not be required, on submitting papers, to disclose that they used AI; the independent-review duty and the certification-by-signature that accompany it sit in a model rule in the Part’s Appendix A, which section 161.4 says an individual court may adopt in its discretion and is encouraged to adopt. Florida’s amendment to Rule 2.515(d)(2) (28 May 2026) does not mention AI at all: it adds to the representations a signer makes on filing that “the legal authorities identified exist and are accurately cited”, and the Court’s commentary says the amendment was adopted principally as a statewide replacement for the circuits’ varied AI disclosure and certification orders. Oklahoma’s Court of Criminal Appeals Rule 1.17 (18 February 2026) requires every portion of a filed document produced or modified by generative AI to be verified as accurate by a person responsible for the document, on pain of waiver of the affected issue, striking of the document or contempt.

The first instrument in the whole record is a certification: Judge Brantley Starr’s order (N.D. Tex., dated 30 May 2023 in the tracker; the text survives in a Texas bar compilation rather than on the court’s current judge page) requires every filer to certify either that no portion of a filing was drafted by generative AI or that all AI output was human-verified using print reporters or traditional databases. Judge Berger’s standing order (M.D. Fla., 2 April 2026) takes the same either-or form, requiring a certification at the end of every filing that no generative AI was used or that any AI-drafted language was personally reviewed. Missouri’s Clay County Local Rule 3.3.1 (1 July 2024) and Montana’s Thirteenth District Local Rule 35 (1 January 2026) complete the certification set in the United States. Louisiana’s Act 250 of 2025 is adjacent rather than within it: a diligence-to-verify duty and a disclosure duty for exhibits that are AI-generated or altered, with no certification. Ireland’s HC 142 belongs here too: its declaration in affidavits is a negative certification about substantive content.

Verification without a statement

Thirty-nine of the 46 instruments carry a verification duty, and for eight it is the only AI-specific duty: the Alberta, Newfoundland and Labrador, Nova Scotia Supreme Court, Nova Scotia Court of Appeal and Saskatchewan directions, Singapore’s circular, New York’s Part 161 as it binds statewide (a permission plus a review duty, with certification only under the optional model rule), and the proposed Federal Rule of Evidence 707. Alberta’s three courts issued a joint notice on 6 October 2023 requiring exclusive reliance on authoritative sources and meaningful human control, with no disclosure. Newfoundland and Labrador (12 October 2023), Nova Scotia’s Supreme Court (28 February 2025) and Court of Appeal (14 March 2025) and Saskatchewan’s General Application Practice Directive 12 (1 January 2026) took the same line. Ontario’s Superior Court practice directions (announced 12 February 2026; Part K of the consolidated civil direction) state that AI-generated references may include incorrect or fictitious authorities, require all legal information obtained with AI to be verified against trusted and authoritative sources, require factums to hyperlink authorities to public sources such as CanLII with a paragraph reference for each citation, restate the signed authenticity certification required by rule 4.06.1, and list public reprimand, costs orders, adjournment or dismissal, contempt proceedings and referral to the Law Society of Ontario as possible responses.

Singapore’s Registrar’s Circular No. 1 of 2024 (issued 23 September 2024, effective 1 October 2024, mirrored by the State Courts and Family Justice Courts) is the clearest binding statement of the verify-only position: court users may use generative AI, are fully responsible for the content, must verify it, must not fabricate evidence, need not declare AI use unless asked, and face personal costs for fabricated authorities. The High Court has since made the costs warning good in Tan Hai Peng v Tan Cheong Joo (6 March 2026), ordering SGD 5,000 each against counsel and the supervising solicitor. Norway’s Supreme Court guidance for advocates, section 19 (1 July 2026), addressed to lawyers rather than parties, takes the same line.

Silence, chosen and recorded

The tracker distinguishes silence that has been decided from silence that reflects nothing found. England and Wales is the first kind. The judiciary’s AI guidance (version 3, 31 October 2025) is addressed to judges and imposes no party-disclosure rule. The Civil Justice Council consulted on AI in court documents from 17 February to 14 April 2026 and its interim findings of 30 June 2026 recorded a strong consensus that no AI-specific formal requirements are needed for pleadings and skeleton arguments prepared by lawyers, while leaving expert evidence, AI-shaped witness statements and litigants in person for the final report expected later in 2026. The professional duty is set by Ayinde v Haringey, by the Bar Standards Board’s guidance (18 May 2026) and by the SRA’s warning notice (17 August 2026), none of which asks for a statement to the court.

New Zealand is the second decided silence. The Chief Justice’s guidelines of 7 December 2023, in separate versions for judges, lawyers and non-lawyers, require verification and confidentiality and state that there is no requirement to disclose AI use in court documents. Hong Kong’s judiciary guidelines (July 2024) are addressed to judges and support staff and impose nothing on filers.

The third kind is the 111 country records, 54 of them provisional, whose filing-disclosure cell says “none” after a check. Germany, France, the Netherlands, Sweden, Switzerland, Japan and Malaysia are checked records in this group: each has professional guidance on verification and none has a court rule on disclosure. The count should be read as “no instrument found on the stated date”, and for provisional records as “not yet re-opened by the editor”.

The country map

Category (filers unless stated)	Binding	Guidance	Proposed	Case law	None	Unclear
Disclosure or certification in filings	6	3	4	0	111	6
Duty to verify AI output	10	28	1	11	77	3
Litigants in person and AI	6	2	1	4	110	7
Judges’ and court staff’s use of AI	10	22	6	1	77	14
Prohibition on AI deciding or drafting judgments	10	18	3	1	91	7
AI-generated and deepfake evidence	1	4	1	3	112	9

All 130 records, 5 September 2026. Excluding the 54 provisional records changes the “none” column (57 instead of 111 for filing disclosure, 23 instead of 77 for verification, 25 instead of 77 for judicial use) and two judicial-use cells, because one guidance record (Rwanda) and one proposed record (Ukraine) are provisional. Every other binding, guidance, proposed and case-law cell in the table sits in a checked record.

Binding filing disclosure or certification (6). Australia, through the NSW, federal, Victorian and Queensland instruments above. Canada, through the Federal Court, Manitoba, Yukon, Nova Scotia and Ontario. Ireland, through HC 142. The Philippines, through A.M. No. 25-11-28-SC (adopted 18 February 2026), whose mandatory disclosure is imposed on judges and court staff preparing adjudicatory documents (tool and version, purpose, extent of use, degree of human oversight, preserved output and a responsibility statement) and whose scope includes court users’ submissions without a separate lawyer certification rule. Qatar, through the QICDRC direction, which reaches only the financial-centre courts. The United States, binding only where a judge or a state has acted.

Guidance (3). Israel, where Bar opinion At/60/24 (7 May 2024) requires a lawyer to tell the court where material was processed solely by AI or where AI-derived information could not be verified, and the courts’ Procedure 01-26 (26 March 2026) requires disclosure of the courts’ own AI use. The United Arab Emirates, through the DIFC note. International bodies, through arbitral soft law (SVAMC, SCC, Ciarb) that makes disclosure case-by-case.

Proposed (4). India’s Supreme Court AI Committee published Draft Regulations for Use of AI in Courts, 2026 on 3 June 2026 for the Supreme Court, all High Courts and all subordinate courts and tribunals; lawyers would have to disclose AI use in pleadings, and AI could not decide cases, bail or credibility. Comments closed 15 July 2026 and the text is not notified. Kenya’s judiciary circulated a draft AI policy and draft practice directions in May 2026 that would require a certificate of human verification with any AI-assisted filing; the drafts are not published on the judiciary’s site and the tracker records them from reporting. Ecuador’s Ethics Committee protocol (approved 16 June 2026, awaiting a Judiciary Council resolution) would require every procedural document to declare when and how AI was used, binding judicial actors rather than private lawyers. The United Kingdom’s proposal is the CJC’s, and its interim finding points away from a rule.

Verification, binding (10). The six above plus Singapore (the Registrar’s Circular), Spain (the bar’s Circular 3/2026 of 10 April 2026 makes filing AI-assisted documents without verification a serious infraction), Poland (section 23e of the advocates’ ethics code, adopted 12 and 13 June 2026) and Slovakia (Resolution 12/4/2025 of the bar). In Spain, Poland and Slovakia the binding instrument is a professional rule, not a court rule; the courts of those countries have enforced the duty through fines and inadmissibility (the Canary Islands and Navarre chambers, the Czech Constitutional Court and the Guimaraes Court of Appeal appear in the tracker as judgment guidance).

Case law only (11). Argentina, Chile, China, Colombia, Denmark, India, Italy, Kenya, Portugal, South Africa and one further checked record have no rule or guidance naming a verification duty for filers and a court decision setting it. India is the clearest example: Pooja Ramesh Singh v J&K Bank (2 July 2026) holds that Bar and Bench owe a duty to verify AI-generated authority and that citing unverified AI judgments is advocate misconduct, and Gadiya (2 September 2026) extends it to quasi-judicial officers. Colombia’s civil cassation chamber fined a lawyer fifteen minimum monthly wages for ten non-existent rulings, and the Council of State held in June 2026 that all AI output must be humanly verified.

The United States: five positions in one country

The United States record is a map of its own. At federal level the tracker holds eight instruments: five judge-level standing orders (Starr, 30 May 2023; Fuentes, N.D. Ill., 31 May 2023; Baylson, 6 June 2023; Graham, 21 October 2025; Berger, 2 April 2026), the Fifth Circuit’s proposed Rule 32.3 certification (not adopted, decision published June 2024), the Executive Office for Immigration Review’s Policy Memorandum 25-40 (8 August 2025) and the proposed Federal Rule of Evidence 707 on machine-generated evidence (published for comment 4 June 2026). No circuit and no district has adopted a court-wide filing rule; the Fifth Circuit is the only appellate body to have considered and declined one.

At state level the 51 subdivision records (50 states and the District of Columbia) divide as follows on filing disclosure or certification:

Position	States	Which
Binding rule in force (as coded in the country record)	6	Connecticut (Practice Book 4-9, 23 Jun 2026); Florida (Rule 2.515 as amended 28 May 2026, a citation-accuracy representation that replaced the circuits’ AI orders); Kansas (Shawnee County Rule 3.125, county level); Montana (Fourth and Thirteenth District local rules); New York (Part 161: a binding statewide no-disclosure policy, with certification only where a court adopts the Appendix A model rule); Oklahoma (Court of Criminal Appeals Rule 1.17, criminal appeals only)
Guidance only	11	Illinois (Supreme Court policy, 1 Jan 2025, stating AI use need not be disclosed); New Jersey; New Mexico; North Carolina; Ohio; Oregon; Pennsylvania; Rhode Island; Texas; Washington; West Virginia
Proposal	2	Hawaii (Supreme Court committee final report, 16 Dec 2025, recommending a federal-style declaration); Maryland (proposed amendment to Rule 1-311)
Nothing found	32	The remaining states

Four of the six adopted their rule in the eighteen weeks between 18 February and 23 June 2026 (Oklahoma, Florida, New York on the effective date the tracker records, Connecticut), and none of the four asks which tool was used: Connecticut chose certification-by-filing, Florida a citation-accuracy representation, Oklahoma a verification duty, and New York a statewide statement that disclosure is not required. Illinois had already taken that last position in writing: its Supreme Court policy, effective 1 January 2025, states that AI use need not be disclosed. The bar ethics opinions in the guidance column are consistent with each other and with ABA Formal Opinion 512 (29 July 2024): verify everything, and check whether the particular court has a rule. Seven of the 51 subdivision records carry a binding judicial-use position (Arizona, Arkansas, California, Delaware, Idaho, New York and Utah), rules for judges and court staff rather than filers (Arizona’s ACJA 1-509, California’s Rule 10.430, Arkansas’s Administrative Order 25).

Rules for judges outnumber rules for filers

Ten countries bind their own judges’ use of AI, against six that bind filers. Brazil’s CNJ Resolution 615/2025 (approved 11 March 2025, in force 14 July 2025) governs judges’ and staff’s use of AI with mandatory human supervision, risk classification and a national committee, replacing Resolution 332/2020. Colombia’s Acuerdo PCSJA24-12243 (16 December 2024), issued to comply with the Constitutional Court’s judgment T-323 of 2024, bars free or commercial chatbots, bars AI for evidence assessment or deciding, and requires express disclosure in the file of tool, model, date, purpose and prompts. Italy’s Law 132/2025, article 15(1), reserves every decision on interpretation, facts, evidence and measures to the magistrate. Spain’s CGPJ Instruccion 2/2026 (published 30 January 2026) requires effective human control and bars unapproved tools. Israel’s Procedure 01-26, Nepal’s ICT Policy 2083, China’s Supreme People’s Court Opinions of 8 December 2022, Indonesia’s Supreme Court letter of 2 September 2026 (a narrow bar on uploading draft-judgment materials to AI services), India’s Kerala and Gujarat High Court policies and the Philippines’ framework complete the ten.

Ten countries also bar AI from the decision itself, and the two lists overlap in nine. France is the exception in the other direction: article 47 of the 1978 loi Informatique et Libertés forbids a judicial decision assessing a person’s behaviour from being founded on automated profiling, a statutory bar that predates generative AI by four decades, while the Conseil d’État’s charter of 11 December 2025 governs administrative judges by guidance.

The asymmetry has a reason visible in the instruments. A court can bind its own judges by administrative act; binding filers requires a rule of procedure, which in most systems means a rules committee, a consultation and a gazette. The Philippines framework is the one instrument in the record that does both at once, and its filer-side obligation is the thinner of the two.

Versions and reversals

Seven of the 114 instruments have more than one recorded version, and two further records (Victoria’s 2024 guidelines and the Fifth Circuit’s proposal) were superseded or not adopted.

Instrument	First version	Change
NSW SC Gen 23	21 Nov 2024	Amended 28 Jan 2025, in force 3 Feb 2025, ten weeks after issue
Federal Court of Canada notice	20 Dec 2023	Updated 7 May 2024
Victoria litigant guidelines	May 2024	Superseded by SC GEN 25, 14 May 2026, which moved from guidance to practice note and dropped a general disclosure rule in favour of disclosure on direction
Fifth Circuit proposed Rule 32.3	late 2023	Not adopted; decision published June 2024
Kenya High Court rulings on AI-generated pleadings	Mar 2026	Second ruling 16 Apr 2026 recorded as a further version of the same record
Arizona ACJA 1-509	2024	Revised 29 Jan 2025
Rio Negro (Argentina) acordada 15/2024	Oct 2024	Extended 1 Oct 2025 to make verification mandatory for lawyers provincially
Norway Supreme Court advocates’ guidance	2025	Section 19 on AI tools, 1 Jul 2026
Ciarb arbitration guideline	Mar 2025	Updated 5 Sep 2025

The direction of revision is not uniform. New South Wales tightened within weeks. Victoria loosened, replacing an automatic-disclosure guideline with a court-directed one. The Fifth Circuit decided not to adopt certification after comment, which is the same conclusion the Civil Justice Council reached in England two years later. Illinois wrote the no-disclosure position into policy. The trend since mid-2025 in the United States and Australia is toward the filer standing behind the content rather than declaring the tool: Queensland (September 2025), Florida (May 2026), Connecticut (June 2026) and New York’s optional model rule (June 2026) all attach the duty to the signature and none asks which tool produced the text.

Litigants in person

Six countries bind self-represented litigants directly, and in every case it is because the filing instrument applies to “parties” or “court users” without distinction: Australia (SC Gen 23, the Local Court note, NCAT PD 7, GPN-AI and PD-AI), Canada (the Federal Court, Manitoba, Yukon, Saskatchewan and Ontario’s factum certification), Ireland (HC 142), the Philippines (the framework’s definition of court users), Singapore (the Registrar’s Circular, which expressly includes litigants in person) and the United States (Judge Starr’s and Judge Fuentes’s orders reach pro se filers). Four countries have case law and no rule: China, France, Kenya and Poland, each with a decision rejecting a self-represented party’s AI-generated filing. New Zealand and the United Kingdom have guidance addressed to judges about litigants in person rather than to the litigants themselves, and the CJC’s interim report anticipates a distinct approach for them.

The country records bear on this. New Zealand’s record notes that its reported AI-citation cases have involved self-represented parties, Singapore’s that a majority of its reported cases did, and Kenya’s two leading rulings (recorded from reporting, not yet verified against the judgments) arose from one litigant’s filings. Where the rules are silent for litigants in person, the courts have written the rule in the judgment.

Timeline

Date	Instrument	Position
30 May 2023	Judge Starr, N.D. Tex.	First instrument anywhere; certification
31 May 2023	Judge Fuentes, N.D. Ill.	Conditional disclosure
6 Jun 2023	Judge Baylson, E.D. Pa.	Disclosure plus certification
23 Jun 2023	Manitoba King’s Bench	First court-wide rule; declaration
26 Jun 2023	Supreme Court of Yukon	Declaration of tool and use
6 Oct 2023	Alberta courts (joint)	Verification only
7 Dec 2023	New Zealand (Chief Justice)	No disclosure unless asked
20 Dec 2023	Federal Court of Canada	Declaration in the first paragraph
21 Dec 2023	DIFC Courts	Declaration at earliest opportunity (guidance)
Jun 2024	Fifth Circuit	Proposed certification not adopted
23 Sep 2024	Singapore Registrar’s Circular	Verify; no declaration unless asked
21 Nov 2024	NSW SC Gen 23	Prohibition for sworn evidence; no-AI statement
1 Dec 2024	Ontario Rule 4.06.1	Certification of authenticity of authorities
1 Jan 2025	Illinois Supreme Court policy	AI use need not be disclosed
24 Sep 2025	Queensland PD 5 of 2025	Named responsible person
31 Oct 2025	England and Wales judicial guidance v3	No party-disclosure rule
6 Jan 2026	QICDRC PD 1 of 2026	Disclosure on direction
18 Feb 2026	Philippines A.M. No. 25-11-28-SC; Oklahoma OCCA Rule 1.17	Judicial disclosure; verification
16 Apr 2026	Federal Court of Australia GPN-AI	Disclosure on direction
14 May 2026	Victoria SC GEN 25	Direction replaces guideline
28 May 2026	Florida Rule 2.515	Signer represents that cited authorities exist and are accurately cited; replaces circuit AI orders
1 Jun 2026 (effective date per the tracker)	New York Part 161	Statewide: no disclosure required; certification only under the optional model rule
3 Jun 2026	India draft regulations	Proposed declaration in pleadings
23 Jun 2026	Connecticut Practice Book 4-9	Certification by filing
30 Jun 2026	Civil Justice Council interim findings	Consensus against a rule
1 Sep 2026	Ireland HC 142 in operation	No-AI declaration for evidence; verification

What to watch

Four instruments could change the country table before the end of 2026. The Civil Justice Council’s final report for England and Wales is expected later in the year and is the one place where a checked “proposed” cell may become a decided “none”. India’s draft regulations, if notified, would move the world’s second-largest English-speaking legal market from case law to a binding declaration rule for every court in the country. Kenya’s draft practice directions would do the same for East Africa’s largest bar. Proposed Federal Rule of Evidence 707 addresses machine-generated evidence rather than filings but would be the first federal-level AI rule in the United States.

Three things to watch in the instruments already in force. Whether Florida’s 17th Circuit order, carried as superseded on the strength of the Supreme Court’s commentary, is formally rescinded, and whether the other circuits’ orders follow. Whether Ireland’s HC 142 produces its first certification breach in the incident tracker, which would be the first row anywhere in which the sanction attaches to the declaration rather than the fabrication. And whether Victoria’s move from automatic to directed disclosure is followed by other Australian courts, which would leave New South Wales as the only jurisdiction in the record with a standing no-AI statement.

Three sentences journalists can quote

As of 5 September 2026 the SafeLegalAI Regulation Tracker records 46 binding-form court instruments in nine jurisdictions that address what a party must do about AI when filing, 42 of them in force; 39 carry a verification duty, 27 a disclosure requirement, 16 a certification and 5 prohibit AI for a class of sworn document.

Six of 130 countries bind filers to disclose or certify AI use (Australia, Canada, Ireland, the Philippines, Qatar and the United States), while ten bind their own judges’ use of AI and ten bar AI from the decision itself.

England and Wales, New Zealand, Singapore, Illinois and New York have each decided, on the record, that a party need not declare AI use unless the court asks; New South Wales and Ireland require sworn evidence to declare that AI was not used.

Appendix A: the 46 binding filing instruments

Position codes: P prohibition (class of document), D declaration (automatic), Dd declaration on direction, C certification, V verification. Dates are the current version. Codes follow the tracker’s requirements field; where the operative text read on 5 September 2026 is narrower, the row says so.

Jurisdiction	Instrument	Current version	Position
Australia (Qld)	Supreme Court PD 14 of 2024, expert evidence in criminal proceedings	15 Jul 2024	D, V
Australia (NSW)	Supreme Court PN SC Gen 23	28 Jan 2025	P, D (negative), V
Australia (NSW)	NCAT Procedural Direction 7	7 Mar 2025	P, D, V
Australia (Qld)	Supreme Court PD 5 of 2025, accuracy of references	24 Sep 2025	C, V
Australia (NSW)	Personal Injury Commission PIC13	1 Jan 2026	P, D, V
Australia (NSW)	Local Court practice note	12 Jan 2026	P, D, V
Australia (Cth)	Federal Court GPN-AI	16 Apr 2026	D (evidentiary uses, para 4.10), Dd, V
Australia (Cth)	FCFCOA PD-AI	1 May 2026	Dd, V
Australia (Vic)	Supreme Court PN SC GEN 25	14 May 2026	Dd, V
Canada (MB)	King’s Bench practice direction	23 Jun 2023	D
Canada (YT)	Supreme Court GENERAL-29	26 Jun 2023	D
Canada (AB)	Tri-court notice on large language models	6 Oct 2023	V
Canada (NL)	Supreme Court notice	12 Oct 2023	V
Canada (NS)	Provincial Court directive	27 Oct 2023	D, V
Canada (Fed)	Federal Court notice, updated	7 May 2024	D, V
Canada (NS)	Registrar in Bankruptcy directive	4 Oct 2024	D
Canada (ON)	Rules of Civil Procedure r. 4.06.1(2.1)	1 Dec 2024	C, V
Canada (NS)	Supreme Court notice	28 Feb 2025	V
Canada (NS)	Court of Appeal notice	14 Mar 2025	V
Canada (SK)	King’s Bench GA-PD 12	1 Jan 2026	V
Canada (ON)	Superior Court practice directions	12 Feb 2026	C, V
Qatar (QFC)	QICDRC PD 1 of 2026	6 Jan 2026	Dd, V
International	JAMS rules for AI disputes	14 Jun 2024	record-keeping, confidentiality
Ireland	High Court PD HC 142	29 Jul 2026 (in operation 1 Sep 2026)	P (substantive content), D (negative), C, V
Singapore	Registrar’s Circular 1 of 2024	23 Sep 2024	V
India	Supreme Court draft regulations (consultation)	3 Jun 2026	D (proposed), P (decisions), V
Philippines	A.M. No. 25-11-28-SC	18 Feb 2026	D (judges), V, P (decisions)
US (Fed, N.D. Tex.)	Judge Starr certification order	30 May 2023	C
US (Fed, N.D. Ill.)	Judge Fuentes standing order	31 May 2023	D (conditional), V
US (Fed, E.D. Pa.)	Judge Baylson standing order	6 Jun 2023	D, C
US (Fed, 5th Cir.)	Proposed Rule 32.3 (not adopted)	Jun 2024	C (not adopted)
US (Fed, S.D. Ohio)	Judge Graham standing order	21 Oct 2025	D, C
US (Fed, M.D. Fla.)	Judge Berger standing order	2 Apr 2026	C (either no AI, or reviewed), V
US (Fed)	Proposed FRE 707 (comment)	4 Jun 2026	evidence, V (proposed)
US (Kansas)	Shawnee County District Court Rule 3.125	14 Jun 2024	D, C, V
US (Missouri)	Clay County Local Rule 3.3.1	1 Jul 2024	D, C
US (Illinois)	22nd Judicial Circuit AO 2024-23	1 Oct 2024	D, V
US (Arkansas)	Supreme Court Administrative Order 25	5 Jun 2025	confidentiality, P (confidential data)
US (Louisiana)	Act 250 of 2025 (exhibits)	1 Aug 2025	D, V
US (Montana)	Fourth District Rule 3.G	31 Oct 2025	D, C, V
US (Montana)	Thirteenth District Local Rule 35	1 Jan 2026	D, C, V
US (Florida)	17th Circuit AO 2026-03-GEN (terms not read; superseded by Rule 2.515 per the Court’s commentary)	1 Jan 2026	D, V (as coded)
US (Oklahoma)	Court of Criminal Appeals Rule 1.17	18 Feb 2026	C, V (the rule is worded as a verification duty; the tracker also codes certification)
US (Florida)	Rule 2.515(d)(2) amendment (citation-accuracy representation; no AI wording)	28 May 2026	C, V
US (New York)	22 NYCRR Part 161 (statewide: no disclosure required; Appendix A model rule optional)	1 Jun 2026 (per the tracker)	no D; C, V only under the model rule
US (Connecticut)	Practice Book section 4-9	23 Jun 2026	C, V

Appendix B: sources

Primary documents, each the current version recorded in the Regulation Tracker and re-opened for this report on 5 September 2026 unless marked unverified:

• Supreme Court of New South Wales, Practice Note SC Gen 23 (amended 28 Jan 2025): PDF
• Federal Court of Australia, GPN-AI (16 Apr 2026): fedcourt.gov.au (archive, 14 Aug 2026)
• Supreme Court of Victoria, Practice Note SC GEN 25 (14 May 2026): supremecourt.vic.gov.au
• Supreme Court of Queensland, Practice Direction 5 of 2025 (24 Sep 2025): PDF
• Federal Court of Canada, Notice to the Parties and the Profession (updated 7 May 2024): PDF
• Court of King’s Bench of Manitoba, Practice Direction (23 Jun 2023): PDF
• Supreme Court of Yukon, GENERAL-29 (26 Jun 2023): PDF
• Alberta courts, Notice to the Profession and Public (6 Oct 2023): PDF
• Ontario, O. Reg. 384/24 amending the Rules of Civil Procedure (r. 4.06.1(2.1) and (2.2)): e-Laws (renders only in a browser); the certification wording is restated in Part K of the Superior Court’s civil practice direction above
• Ontario Superior Court of Justice, consolidated provincial practice directions, Part K “The Use of Artificial Intelligence (AI) for Court Proceedings” (announced 12 Feb 2026): civil, family, criminal; announcement
• High Court of Ireland, Practice Direction HC 142 (29 Jul 2026): courts.ie
• Courts of New Zealand, Guidelines for use of generative AI (7 Dec 2023): courtsofnz.govt.nz
• Supreme Court of Singapore, Registrar’s Circular No. 1 of 2024 (23 Sep 2024): PDF
• Qatar International Court and Dispute Resolution Centre, Practice Direction No. 1 of 2026 (6 Jan 2026): qicdrc.gov.qa
• Supreme Court of the Philippines, A.M. No. 25-11-28-SC (18 Feb 2026): elibrary.judiciary.gov.ph
• Supreme Court of India AI Committee, Draft Regulations for Use of AI in Courts, 2026 (3 Jun 2026): PDF
• Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd, 2026 INSC 668 (2 Jul 2026): Indian Kanoon; tracker record. Vijay Ghanshyam Gadiya v Union of India, 2026 INSC 947 (2 Sep 2026): Supreme Court of India PDF
• Ayinde v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), 6 Jun 2025: judiciary.uk PDF; Bar Standards Board, AI guidance (May 2026): PDF; Solicitors Regulation Authority, warning notice on misuse of AI (17 Aug 2026): sra.org.uk
• Illinois Supreme Court, Policy on Artificial Intelligence (effective 1 Jan 2025): PDF
• Tan Hai Peng Micheal v Tan Cheong Joo [2026] SGHC 49 (6 Mar 2026): eLitigation; Singapore Law Watch PDF
• Civil Justice Council, Update on consultation findings (30 Jun 2026): judiciary.uk
• Judge Brantley Starr, N.D. Tex., Mandatory Certification Regarding Generative AI (30 May 2023 per the tracker; the date is not on any primary page now reachable): txnd.uscourts.gov, judge page as archived 27 Aug 2026 (the live page moved after the tracker recorded it; the order is also reproduced in the Texas ENRLS sampling of AI court orders)
• Judge Michael Baylson, E.D. Pa., Standing Order (6 Jun 2023): paed.uscourts.gov judge page (the order PDF linked there returns 404 as of 5 Sep 2026; text reproduced in the Texas ENRLS sampling of AI court orders)
• Judge James Graham, S.D. Ohio, Standing Order (21 Oct 2025): PDF
• Judge Wendy Berger, M.D. Fla., Standing Order on AI (2 Apr 2026): PDF
• Shawnee County District Court Rule 3.125 (14 Jun 2024): shawneecourt.org
• Montana Fourth Judicial District Rules of Practice, Rule 3.G: PDF
• Connecticut Practice Book section 4-9 (23 Jun 2026): jud.ct.gov
• New York, 22 NYCRR Part 161 (effective 1 Jun 2026 per the tracker), read from the archive: nycourts.gov (archive, 2 Aug 2026)
• Supreme Court of Florida, In re Amendments to Rule 2.515, SC2026-0673 (28 May 2026), read from the archive: PDF (archive, 21 Jul 2026)
• Oklahoma Court of Criminal Appeals, 2026 OK CR 7, Rule 1.17 (18 Feb 2026), read from the archive: okcca.net (archive, 7 Jun 2026)
• Louisiana Act No. 250 of 2025 (effective date per the tracker), read from the archive: legis.la.gov (archive, 12 Aug 2026)
• Conselho Nacional de Justica, Resolution 615/2025 and Colombia’s Acuerdo PCSJA24-12243, Italy’s Law 132/2025, Spain’s Instruccion 2/2026, Israel’s Procedure 01-26 and the other judicial-use instruments: linked from their Regulation Tracker records
• SafeLegalAI data: Regulation Tracker · documents.json · map.json · taxonomy · related reports: AI in legal practice: what 130 country records require, US federal court AI standing orders, The sanctions ledger

Appendix C: changes to this report

None since first publication. This report replaces the July 2026 draft comparison of twelve jurisdictions at the same address. Planned for the next edition: read the Florida 17th Circuit order and the parallel circuit orders; add the Civil Justice Council’s final report; code the individual United States judge-level orders that the tracker has read into the appendix.