The Expert RecordThe standard for expert witnesses using AI

The Standard · Clause 4 of 8

Clause 4: Checkability before adoption

If you cannot name what would check it, do not use it.

55 matters in the record engaged itVersion 1.0
The clause, verbatim

Checkability before adoption. Before adopting any output, the expert identifies the specific thing against which it would be checked. Where no such thing can be identified, the output is not adopted.

A test applied at the moment of adoption

Adoption (the moment a piece of output is taken into the work as the expert's own) is the point after which an assertion stands on the expert's authority, whatever produced it. This clause asks one question just before that point: what, specifically, would this be checked against? "I would verify it" is not an answer. A page of the production, a standard obtained from the issuing body in the revision operative on the relevant date, a recomputation in the spreadsheet, a transcript page and line: those are answers.

If a specific, obtainable thing comes to mind, the tool is working as an instrument, and the check is a mechanical step that can be performed, described and logged. If the honest answer is that the output seemed right, matched the expert's understanding, or agreed with what the model said earlier, it is entering on the tool's authority and stays out. The test does not require knowing how a model works, only what would settle the question.

Green, yellow and red work

The question sorts tasks into the book's three categories. Green work is checkable against material already in hand: optical character recognition where the page images are kept, deduplication by hash, timestamp normalization. Yellow work needs something obtained from outside, as when a model names the governing standard and the expert then buys it and reads the section. Red work returns no answer at all, because the output would be the expert's judgment; that belongs to Clause 2. The pre-use checklist puts this question before every use.

Readings the clause does not support

The clause concerns adoption, not reading. It does not require exhaustive verification of every response before it is even considered. Nor is it satisfied by asking a second model, or the same one again: a cross-check is useful only if its own basis can be examined, which is the point the entry on Matter of Weber makes about a damages expert's Copilot cross-check. A check against memory does not qualify either. It is the companion of Clause 3: this clause names the check before adoption, and independent authority requires that the check actually be made against the source.

Where the record applies it

In State of Washington v. Puloka, the defense offered a bystander video enhanced with AI software. The frames the model added by prediction could not be traced to anything in the original file, and after a Frye hearing (an inquiry into general acceptance in the relevant scientific community) the trial court excluded the enhanced clip. The case was tried on the original footage. The order is unpublished, and the account rests on press and practitioner reporting.

The Cour d'appel de Rouen, RG 25/01677 (Copilot analysis) reversed a summary order on the strength of dye tracing, site reports and professional technical opinions, and said expressly that a Copilot analysis of a contractor's letter was not among the evidence it relied on. The only thing that analysis could be checked against was the letter, which the court already had.

Concord Music Group v. Anthropic (Chen declaration) shows a check that failed. The underlying article was real and had been found by a person, but when counsel used Claude.ai to format the reference it returned a wrong title and authors, and a manual citation check did not catch the error. The magistrate judge struck the paragraph and imposed no sanctions.

The rules it sits beside

The Standard does not restate any court rule and does not depend on one. In United States federal courts, Federal Rule of Evidence 702 asks whether a method can be tested; output adopted without knowing what would check it cannot be. Other systems ask the same questions in their own terms, and the matters below include tribunals in several of them.

Matters in the record that engaged Clause 4

55 of the 103 matters in AI in the Courtroom engaged this clause. Each is listed with the sentence from its entry that says how. The mapping describes conduct, not blame, and each entry sets out the facts it rests on.

Expert witnesses and AI

Fabricated citations and facts in filings

AI-generated and altered evidence

Judges, courts and AI

AI inside the courtroom

  • Williams v. State (Ind. Ct. App.) AI-assisted transcript “That is Clause 4 of the Expert Record Standard, checkability before adoption, and Clause 3, independent authority: the words attributed to a speaker should be confirmed against the audio, not accepted on the system's rendering.”
  • State v. Horcasitas (Pelkey AI victim impact statement) “Checkability, Clause 4, is the one most directly engaged.”
  • Arizona Supreme Court AI news reporters “That is checkability, Clause 4, in its simplest form: each video can be checked against a human-authored text before and after it is published.”
  • DoNotPay "robot lawyer" and FTC order “That is the situation non-delegation, Clause 2, addresses: the argument would have been composed by the system in real time, and neither the defendant nor anyone else could have checked it before it was spoken, contrary to Clause 4.”

Rules, orders and guidance

Clause 3: Independent authority · All eight clauses · Clause 5: Reproducible analysis

Top