So Who Must Prove the AI Inference Is Reliable?

Under the revised proposal discussed by the Advisory Committee on Evidence Rules in May 2026, the party offering an expert-like AI inference must establish its reliability. The ordinary way to do that is to provide an expert who can explain how the system produced the evidence, although the draft leaves a narrow route for exceptional proof such as convincing validation studies. The answer is therefore direct but conditional: the proponent carries the proposed reliability showing, yet the proposal is not an enacted rule and does not apply to every computer output.

The practical value of that answer lies in knowing when it applies and what it demands. A party cannot reach the conclusion merely by saying that a computer produced it. But an opponent cannot avoid the ordinary distinction between measurement and inference by calling every digital item “AI evidence.” The analysis begins with the system’s function, continues through four reliability gates and the evidence supporting them, and then asks a separate question about whether the digital item itself is authentic.

The Proponent’s Burden Begins With Classification

The first decision is not who has the better expert. It is whether the offered output is an expert-like inference at all. A machine that records a sensor reading, preserves metadata, reports coordinates, or performs straightforward arithmetic may raise accuracy and authenticity questions without necessarily supplying an expert opinion. A system that predicts, classifies, compares, reconstructs, or generates a conclusion is doing more than preserving an observation. Its output depends on a method that connects data to meaning.

That functional distinction matters because the proposed Rule 707 is triggered by a counterfactual: would a human witness offering the same conclusion be subject to Rule 702? If the answer is yes, the proponent cannot avoid the reliability inquiry by presenting the output directly, labeling it a business record, or calling the person who operated the program as a lay witness. If the system merely performed a measurement or retrieval, the proposed expert-like framework may not be the right starting point. The classification must follow the operation in the case, not the product name.

Once the output is classified as an expert-like inference, the proponent’s proposed showing has four parts. First, the evidence must help the factfinder with an issue in the case; technical novelty or persuasive force is not enough. Second, the output must rest on sufficient facts or data, including inputs that fit the population and purpose involved. Third, the principles and methods must be reliable. Fourth, those principles and methods must have been reliably applied to the case-specific facts.

The four gates are connected but not interchangeable. A result may address the right issue and still rest on inadequate data. A method may be well validated in general and still be used outside the conditions for which it was tested. A case may contain suitable inputs while the proponent has not shown that the system handled them correctly. The proposed architecture asks the offering party to connect the entire path from purpose to input, from input to method, and from method to conclusion.

The proponent’s burden also explains why the revised draft ordinarily requires an expert. An operator can often say when software was run, what was entered, and what output appeared. An expert can address why the system is suitable for the question, how it was trained or validated, what its error information means, what its limitations are, and how it was applied here. The sources do not support treating every expert as a guarantee of correctness. They support treating expert explanation as the ordinary means of making an expert-like inference intelligible and contestable.