AI Evidence Trail: The 4 Gates a Machine-Made Document Must Pass

Is there a rule for AI-generated evidence yet?

Not a final one. The US Advisory Committee on Evidence Rules spent three years on the question and, in May 2025, identified two separate problems: machine output that may be unreliable in ways buried in the program and difficult to detect, and audiovisual evidence that is fake. For the first it drafted Rule 707; the committee voted 8 to 1 to publish it, with the Department of Justice voting against.

The draft is one sentence: when machine-generated evidence is offered without an expert witness and would be subject to the expert-evidence rule if a witness gave it, the court may admit it only if it meets that rule's reliability requirements. It excludes the output of simple scientific instruments.

The committee said publication was not a presumption that the rule would be enacted. The Standing Committee's report to the Judicial Conference for September 2026 lists Rule 707, and a possible new Rule 901(c) on deepfakes, only as matters of continued consideration. Neither appears among the amendments now on course to take effect.

What are the four gates?

Every evidence system asks some version of the same four questions. The US Federal Rules of Evidence state them unusually compactly, which is why they are the standard reference in discussions of AI material.

Authenticity: is it what you say it is?

Rule 901(a) asks the party offering an item for evidence sufficient to support a finding that the item is what the proponent claims it is. For an ordinary document, a witness who made or received it is usually enough.

AI changes the claim you must be able to support. A photograph is authentic if it shows what the camera recorded; an AI-enhanced photograph is not the same exhibit, and describing it as the original is where a genuine document becomes a misleading one. The same applies to a chat log summarised by an AI tool, a transcript produced by speech-to-text, or a timeline an assistant built from your emails. Each can be useful. None is the underlying record.

Reliability: can a machine's answer be trusted?

This is the gate draft Rule 707 was written for. The committee's reasoning is worth knowing even outside the US: the usual safeguard against unreliable statements is cross-examination, and a machine cannot be cross-examined. So the court looks instead at the inputs, the method and whether it is known to work, the questions Rule 702 asks of a human expert.

For a self-represented party the practical version is simple. If an AI tool calculated, classified or predicted something that matters to your case, such as the interest owed, whether two contracts match, or what a blurred number says, be ready to show what went in, what tool was used, and a way to check the result by hand. If you cannot, present the underlying documents and your own calculation instead.

Does any of this apply in a Hong Kong small claims case?

The gates are softer there, not absent. Section 23(2) of the Small Claims Tribunal Ordinance says the rules of evidence shall not apply to proceedings in the Tribunal, which may receive any evidence which it considers relevant. An AI-assisted summary will usually be received. The question is what it is worth.

The Tribunal also has tools of its own. Under section 16 it must inquire into any matter which it may consider relevant, whether or not a party raised it, and it may summon any witness and require the production of any document, record, book of account or other thing. An Adjudicator who doubts a screenshot can ask for the phone.

Hong Kong's judges have been told to look. The Judiciary's guidelines on generative AI, issued in July 2024, say that where it appears a litigant in person has used generative AI to prepare submissions or documents, the judge should make enquiry with the litigant, and ask what checks for accuracy have been undertaken. On 13 April 2026 the Judiciary said it was drawing up separate guidelines for legal practitioners and other court users, for consultation with the profession later in the year.

What is an evidence trail?

It is the record that lets someone else follow a document from where it started to the copy in your bundle. Building one takes minutes when the document is created and can be impossible months later.

What if you think the other side's evidence is AI-made?

Say so early and specifically, and ask for the original rather than an expert.

• Point to the inconsistency. A date that does not match the metadata, a voice that says something out of character, a document format the sender never used.
• Ask for the source. In Hong Kong's Small Claims Tribunal, ask the Adjudicator to require production of the original file or device under section 16(2).
• Ask who made it. A witness who must explain how a document was created, on oath, is the best test there is. The Judiciary warns that a witness who is found not to have told the truth may face criminal prosecution.
• Do not overclaim. Calling a genuine document a deepfake damages your credibility with the Adjudicator more than the document did.

Pressure-test your evidence before the hearing

The four gates are questions, and questions can be rehearsed. Put each exhibit through them: what fact it proves, who vouches for it, how it was made, and whether it could mislead. Trial AI's mock court runs an adversarial examination of your own bundle, so the document that cannot answer those questions shows up before the Adjudicator finds it.

That rehearsal is preparation, not legal advice, and it does not predict how a court will rule on admissibility. Where the rules of evidence do apply, and in any case above small claims, take advice on the specific exhibit.