The Caribbean Court of Justice bans generative AI from evidence, allows it for drafting, and puts the verification burden on the filer

What the CCJ is, and why that matters here

Most US practitioners have never had a matter in front of this court, so start with the institution. The Caribbean Court of Justice describes itself as a hybrid: "a municipal court of last resort and an international court vested with original, compulsory and exclusive jurisdiction in respect of the interpretation and application of the Revised Treaty of Chaguaramas." It was established on 14 February 2001 by the Agreement Establishing the Caribbean Court of Justice.

Those are two different jobs in one building. In its Original Jurisdiction the CCJ interprets the treaty that governs the CARICOM single market, which is the kind of work a trade tribunal does. In its Appellate Jurisdiction it sits as the final court of appeal for those Caribbean states that have adopted it in place of the Judicial Committee of the Privy Council. The practice direction is issued under both sets of rules, Part 18.1 of the Appellate Jurisdiction Rules 2024 and Part 32.1 of the Original Jurisdiction Rules 2024, so it reaches both streams of work.

Note the date it was issued. 14 February 2025 is the CCJ's own anniversary. That is a scheduling observation on our part, not something the document comments on.

One boundary worth stating plainly: this direction governs proceedings before the CCJ and nothing else. It does not reach the national courts of CARICOM member states, and nothing in the text purports to.

Who it binds, and one group it never names

The opening line lists the audience: the direction "provides guidance on the permissible use of Generative Artificial Intelligence ('GenAI') tools by attorneys, parties, witnesses, self-represented persons, and other court users." Then it fixes the scope: "This Practice Direction applies to all proceedings before the Caribbean Court of Justice".

Part III gives tailored guidance to exactly two of those groups. The first is attorneys and legal practitioners. The second is self-represented persons, who "may use GenAI tools to draft submissions or familiarise themselves with legal principles, but they remain responsible for verifying the accuracy and appropriateness of such outputs." Parties and witnesses are named in the preamble and are caught by the evidence prohibition in Part II, but they get no section of their own.

Court staff are absent. The word does not appear. Judges, registry personnel and judicial assistants are not addressed anywhere in the three pages, and "other court users" in context reads as people who come to the Court rather than people who work for it. If you were expecting the internal-use rules that some judiciaries publish alongside their filer rules, they are not in this instrument. Whether the CCJ has issued something separate for its own staff is not something we checked.

The prohibition sits on evidence, not on AI

This is the structural choice that makes the direction worth reading. The definitions section carves "Court Documents" to mean material filed or submitted, such as written submissions and reports, and then expressly excludes affidavits, witness statements, material reflecting the evidence or opinion of deponents or witnesses, and material tendered into evidence or used in cross-examination.

Everything in that excluded category is off limits. General Principle 1: "The use of GenAI is prohibited in the generation of the content of affidavits, witness statements, or any other material intended to reflect the evidence or opinion of a deponent or witness, or other material tendered into evidence or used in cross-examination. Such documents must reflect solely a person's knowledge and must not include artificial intelligence-generated content."

General Principle 2 closes the obvious workaround. GenAI is prohibited "for the purpose of altering, embellishing, strengthening, diluting or otherwise rephrasing the evidence of a witness or deponent." So you cannot run a witness statement through a model to tighten the prose. That is a broader bar than a naive reading of principle 1 would give you, because polishing is not generating, and principle 2 says polishing is out too.

On the other side of the line, General Principle 4 is unusually direct for a court rule: "The Court does not prohibit the use of GenAI tools for the preparation of Court Documents, provided that users comply with this Practice Direction and any relevant legislation, rules, and ethical obligations." Part III adds that GenAI "may be used to draft submissions, summarise legal arguments, or conduct basic research".

There is a narrow escape hatch for the evidence side. In exceptional circumstances a party may apply for permission to use GenAI in preparing an annexure or exhibit, and the application has to specify five things: a detailed description of the intended use, the programme and version, whether it is closed-source or open-source, whether it has privacy or confidentiality features, and the anticipated advantages. That third item is doing real work, and it connects to the confidentiality rule discussed below. As a drafting note, item (ii) in the published PDF reads "the GenAI programme, system or proposed for use", which appears to be missing a word in the original. We quote it as it stands.

Part IV: verification, disclosure, and a confidentiality rule narrower than it looks

Two obligations that US readers often conflate are separated here, and the difference is practical.

Verification is unconditional. Part III tells attorneys that outputs "must be thoroughly fact-checked, reviewed for accuracy, and adapted to comply with legal standards and requirements" and that "Citations or references generated by GenAI tools must be independently verified using reliable sources." Part IV paragraph 2 repeats it at the level of every court user: outputs "must be checked against authoritative legal texts, case law, or statutes." It then adds an evidentiary hook. Court users "may be required to demonstrate how such verifications were conducted." That is a rule with an audit trail attached, and it is the paragraph we would tell a litigator to plan around, because demonstrating a verification you performed six months ago is a records problem, not a legal one.

Disclosure works differently. Part IV paragraph 3: "The Court may require a user to disclose whether a GenAI tool was employed in the preparation of any document or evidence." That is a discretionary power the Court holds, not a standing certification duty on the filer. Nothing in the direction requires you to tag AI-assisted submissions on the way in. The follow-through is a readiness standard: users "should be prepared to identify specific portions of their submissions influenced by GenAI and explain the steps taken to ensure accuracy."

Read those together and the design is coherent. The CCJ did not want a disclosure box on every filing. It wanted the ability to ask, and the expectation that you can answer on the spot.

Then there is Part IV paragraph 1, the provision most likely to be misread in a firm memo. The text: "Court users must not input sensitive, confidential, or privileged information into open source GenAI tools to avoid unintentional disclosure of private and confidential information."

The prohibition is scoped to open source tools. It is not a general ban on putting privileged material into any model. The direction has already defined GenAI systems as ranging across closed-source and open-source products, so the distinction is deliberate rather than loose, and the exhibit-permission checklist asks about closed versus open source as a separate question from whether the tool has privacy features.

Our reading, and we mark this as inference rather than the Court's holding: the literal text does not prohibit inputting privileged material into a closed-source commercial product, which is the category most law firms actually license. What the direction offers instead is a recommendation. Secure platforms "which adhere to established ethical standards and legal requirements" are recommended for processing private and confidential information. Recommended, not required.

The practical consequence is that a firm relying on this paragraph alone for its policy would be relying on a narrow prohibition plus a soft recommendation. Professional conduct rules and client engagement terms will usually be the stricter constraint, and General Principle 6 points back at exactly that: attorneys are reminded of their duty to ensure filings are independently verified and comply with rules of professional conduct.

What non-compliance costs, and what it does not

Part V is short and worth quoting in full for the ceiling it sets. Non-compliance "may result in" two things: "Costs orders against the offending party" and "Rejection or diminished evidentiary weight of improperly prepared documents or submissions."

That is the list. There is no contempt provision, no referral to a disciplinary or bar authority, no strike-out power, no fine scale, and no mandatory reporting. We are not saying those consequences are unavailable to the CCJ under its general powers. We are saying this instrument does not specify them, and anyone drafting a client alert should not import the sanctions from the English or US fabricated-citation cases into a summary of this document.

The second limb is the interesting one. Diminished evidentiary weight is a remedy aimed at the material, not the lawyer. It fits the evidence-centred structure of the whole direction, and it means the risk of getting this wrong lands partly on the client's case rather than only on counsel's record.

Six Caribbean courts, one pattern

The reason we are covering an eighteen-month-old rule is that it is the anchor of a regional pattern that has attracted almost no attention outside the Caribbean. On our tracking, six Caribbean courts now have generative AI practice directions or equivalent judicial guidance: the CCJ, Jamaica, the Turks and Caicos Islands, The Bahamas, Belize, and Trinidad and Tobago.

Concentration like that in a set of small jurisdictions is not an accident of independent drafting, but we did not open the other five instruments for this piece and we are not going to characterise their contents, their dates, or their similarity to this one. What we will say is that the CCJ sits above several of those systems as a final appellate court, which gives its direction the widest formal reach of the six even though it is the court with the smallest docket.

For a US firm the question is narrow and answerable. If you have a CARICOM treaty matter, or a client whose final appeal runs to the CCJ rather than to the Privy Council, this is a live filing rule you are already subject to and have probably never read. If you do not, it is still the most clearly drafted evidence-versus-drafting split we have seen in a court AI rule, and it is three pages long.

What we did not verify

What we opened: the primary source, Practice Direction No. 1 of 2025, in full, as a three-page PDF from ccj.org, extracted twice by two different methods and cross-checked. Every quotation above was character-matched against that extraction. We also opened the CCJ's own overview page for the description of the institution.

On the date: the scout flagged 2025-02-14 as coming from press coverage rather than the instrument. We checked. The document's own closing line reads "on 14th February 2025" above the signature of Adrian Saunders, President, so in this instance the reported date and the instrument date agree. We are using the instrument's date. The direction takes effect "from the date on which it is issued", so issue and effect are the same day.

What we did NOT open: the CCJ (Appellate Jurisdiction) Rules 2024 and the CCJ (Original Jurisdiction) Rules 2024, so we have not independently confirmed what Part 18.1 and Part 32.1 authorise the President to do beyond the fact that this direction cites them. We did not open the AI practice directions of Jamaica, the Turks and Caicos Islands, The Bahamas, Belize, or Trinidad and Tobago, and nothing above describes their contents. We did not confirm which CARICOM member states have adopted the CCJ's Appellate Jurisdiction, so we have named no country as being under it. We did not check whether the CCJ has issued any separate internal AI policy for judges or registry staff, and we found no revision to this direction, but we did not exhaustively search for one.

What we refuse to claim: that the direction binds any court other than the CCJ; that it imposes a standing duty to disclose AI use on filing, which it does not, because the power to require disclosure sits with the Court; that any sanction beyond costs orders and rejection or diminished evidentiary weight is specified in this text; and that the confidentiality paragraph prohibits inputting privileged material into closed-source tools. That last point is our reading of the words "open source" in Part IV paragraph 1 and it is offered as analysis, not as the Court's position.

Related briefings

• Trinidad and Tobago judiciary AI practice direction
• Qatar QICDRC AI court practice direction
• Ireland Court of Appeal CA16 GenAI practice direction
• Australia Federal Court GPN on AI and GenAI disclosure
• AI Regulation News hub