The Party Offering the AI Inference Must Defend It
The subject of this series is proposed Federal Rule of Evidence 707, a rulemaking effort aimed at AI-produced evidence offered in court without a human expert who can explain and defend the system. The central question is not whether courts should “trust AI.” It is narrower and more practical: when a machine has produced an expert-like conclusion, which party must establish that the conclusion is reliable, and what would that showing require? The answer developed by the latest draft is that the party offering the inference ordinarily carries the burden, usually through an expert foundation—but that answer belongs to a proposal still under study, not to an enacted rule.
That qualification matters. The Advisory Committee on Evidence Rules’ May 17, 2026 report says the Committee did not recommend action on the earlier version of Rule 707. Instead, it revised the draft, favored a version that ordinarily requires an expert while allowing exceptional alternative proof, added a notice requirement, and planned further study with technology experts and lawyers at a Fall 2026 meeting. The report records a divided public-comment process: three comments expressed unqualified support, 27 supported the rule subject to revisions, and 27 opposed it. Those numbers describe a live rulemaking debate, not a settled legal standard.
Yet the problem the draft is trying to organize is already easy to see. Imagine a party offering a facial-recognition result, a predictive assessment, an AI reconstruction, or a generative-AI conclusion as evidence of what happened. A technician may be able to say which button they pressed and when the output was printed. That testimony may establish how the exhibit entered the courtroom. It does not necessarily explain whether the underlying data were suitable, whether the method was validated for this use, whether the output can be audited, or whether the system applied its method reliably to the facts of this case. The proposed rule asks the court not to confuse operation with reliability.
The Rule Is About Expert-Like Inferences, Not AI in General
The legal problem is not that a computer was involved. Modern evidence often passes through machines. A spreadsheet may perform arithmetic. A GPS receiver may report coordinates. A sensor may record a reading. Metadata may preserve a time or location. These outputs can be wrong, but their ordinary reliability questions often concern calibration, handling, or mechanical accuracy rather than a hidden act of expert reasoning.
The proposed Rule 707 is directed at a different kind of output: one in which a system interprets data and draws a conclusion through statistical, model-based, probabilistic, predictive, classificatory, or generative processes. The New York City Bar’s public comment calls this “machine-generated inferential evidence.” Its examples include probabilistic genotyping, facial or voice recognition, AI-based image or audio reconstruction, predictive analytics, and conclusions produced by generative-AI systems. The distinction is important because the system is no longer merely preserving an observation. It is supplying an interpretation that may influence the factfinder’s view of a disputed issue.
That is also why the proposed rule is tied to a counterfactual question: would this output be subject to Rule 702 if a human witness testified to the same conclusion? Rule 702 is the Federal Rule that governs expert testimony. If a human expert said, in substance, “this data means X,” the court would ordinarily ask whether the opinion would help the factfinder, rest on sufficient facts or data, use reliable principles and methods, and apply those principles reliably. The draft Rule 707 seeks to prevent a party from avoiding those reliability questions merely by presenting the conclusion directly or through a lay witness.
This is a rule about substance rather than packaging. Calling an analytical result a “computer output,” placing it in a business-record format, or having a person who knows little about the system print it does not change what the system did. The report’s proposed committee note gives the example of a technician who enters a question and produces an answer without expertise about the program’s validity. The technician may authenticate the act of operating the program. They may not be able to establish that the program’s conclusion deserves the weight or admissibility associated with an expert analysis.
The proposal is also deliberately narrower than a claim that every use of software needs an expert. The committee’s minutes use a phone-extraction tool as an illustration. If the tool simply mirrors information from a device, that use may resemble retrieval. If the information is then processed to identify a person’s key contacts or infer relationships, the system is performing analysis. The same product can therefore fall on different sides of the reliability question depending on the operation performed in the case.
The First Divide Is Between Measurement and Inference
This functional distinction supplies the opening test for the whole subject: what did the system do here? “Machine-generated” describes where an output came from, but not whether it measured, calculated, interpreted, predicted, or invented a conclusion. The City Bar argues that a rule using only the broad source-based label would sweep in routine outputs that the proposed reliability framework is not meant to target. It recommends language that points directly to inferential outputs and explains the difference in the committee note.
The distinction is not a promise that measurement is always reliable or that inference is always unreliable. A sensor can be miscalibrated, a coordinate can be assigned to the wrong device, and a deterministic process can be applied to incomplete inputs. Nor does a statistical or generative method fail simply because it is probabilistic. The point is that the type of reasoning changes what the proponent must be prepared to explain. A measurement can often be checked against calibration and operating conditions. An inference may require a deeper account of training data, assumptions, validation, error, and the connection between a general model and the particular case.
The difference also protects against a subtler form of evasion. A party may present a machine’s conclusion as though it were a neutral fact when the machine has actually made a judgment analogous to one an expert would make. Facial recognition does not merely preserve an image; it proposes an identity match. Predictive analytics does not merely store records; it estimates a future or a probability. An AI reconstruction does not merely reproduce an input; it may fill gaps or generate content based on learned patterns. Those are reasons to ask how the conclusion was produced before asking how persuasive it seems.
The reverse mistake is possible too. If a court treats every digital output as an expert inference, it may impose unnecessary scrutiny on ordinary measurement and obscure the special concern presented by an analytical conclusion. The committee’s revised draft moved away from the earlier “machine-generated evidence” formulation and focused on artificial intelligence, but the source materials acknowledge that the boundary remains difficult. Hybrid systems may combine extraction, measurement, and inference. Future tools may not fit a familiar label. The reliable classification is therefore functional and case-specific, not a simple list of approved or disfavored products.
At this stage, the opening answer to the title can be stated carefully. If the offered item contains an expert-like AI inference that would be treated as expert testimony if a human supplied it, the proposed framework places the reliability showing on the proponent. That does not mean the opponent must prove unreliability before the court can ask questions. It means the party seeking the benefit of the conclusion must be prepared to establish the conditions that make the conclusion admissible under the proposed framework.
The Burden Question Begins With the Offering Party
The draft’s allocation becomes clearer when reliability and authenticity are kept apart. Reliability asks whether the system’s principles, inputs, and application support the conclusion. Authenticity asks whether the item is what the proponent says it is—whether a recording, document, image, or output is genuine and connected to the claimed source. A certificate that authenticates computer output does not, by itself, establish that an AI-generated inference is methodologically sound. Conversely, a genuine output can still be an unreliable analysis.
The report’s separate discussion of possible Rule 901(c) treatment for generative-AI fabrication makes the distinction concrete. A deepfake challenge concerns whether an item offered as genuine was fabricated. The proposed Rule 707 concerns the reliability of an AI-produced conclusion. The two problems may appear in the same exhibit, but they do not ask the same question or automatically place the burden in the same place. This series will therefore treat them as connected but separate tracks rather than use “AI evidence” as a single category.
The practical reasoning will proceed from that boundary. First, the system’s function must be identified: measurement or inference, retrieval or analysis. Then the proposed reliability gates can be examined: whether the output will help the factfinder, rests on sufficient facts or data, comes from reliable principles and methods, and reflects a reliable application of those principles to the case. Only then does the question of foundation become intelligible: ordinarily an expert would explain the system, although the draft leaves room for exceptional proof such as convincing validation studies. The quality of that proof depends on whether the opponent receives enough notice and information to test it.
The point is not to make machines impossible to use in court. It is to prevent a persuasive conclusion from receiving an evidentiary shortcut simply because no human witness is standing behind it. The question that comes next is therefore concrete: when does a computer merely record or measure, and when has it performed the kind of interpretation that makes the proposed Rule 707 relevant?