By: Jeffrey Ramsey
I recently had the privilege of writing the record for the intramural moot court competition at the University of Tennessee Winston College of Law. I drew inspiration from two sources: the controversial HALT Drunk Driving Act and a cruel essay question on an Evidence exam administered by Associate Dean Briana Rosenbaum.[1] From what I can remember, her exam question asked whether the Confrontation Clause is implicated when artificial intelligence (“AI”) is used to test substances believed to be drugs and the AI system is trained on signed statements by human drug analysts from previous cases. In other words, the AI system was trained on statements that were clearly testimonial, at least when used against the defendants for whom the reports were made.[2]
Now, I thought I nailed this answer. At the time, it seemed pretty clear: in that scenario, artificial intelligence is acting as the “mouthpiece” for clearly testimonial statements by humans. Accordingly, Smith v. Arizona controls, and the statements implicate the Confrontation Clause.[3] Case closed. But when I spoke with Dean Rosenbaum about this essay question while drafting the moot court problem, she pointed out a flaw in my analysis.
Dean Rosenbaum explained that Smith v. Arizona, while instructive, is not dispositive of this issue. In Smith, a surrogate drug analyst relied on statements made by the actual analyst to form an independent expert opinion about whether the substance the actual analyst identified as cocaine was, in fact, cocaine.[4] The State argued that the actual analyst’s statements merely served as a basis for the surrogate analyst’s testimony and that the State, through the surrogate, did not offer the actual analyst’s statements for their truth.[5] The Court disagreed, concluding that the surrogate’s testimony carried weight only if the underlying statements were true.[6]
In short, the question in Smith was whether, consistent with the Confrontation Clause, analyst A can test drugs allegedly possessed by defendant B and analyst C can use analyst A’s report at trial to draw a conclusion against defendant B. But Dean Rosenbaum’s question poses a very different issue: whether analyst C (an artificial intelligence model) can rely on analyst A’s report against defendant B, and thousands of other reports made against different defendants, to draw conclusions against defendant D. In other words, under Smith, the government cannot, consistent with the Confrontation Clause, introduce AI-generated statements against a criminal defendant when the underlying model was trained on testimonial statements about that defendant. But Smith does not clearly answer whether artificial intelligence can learn statistical relationships from testimonial statements concerning defendants A through Y and apply those relationships to draw conclusions about defendant Z.
This distinction sent me looking for an answer. After my conversation with Dean Rosenbaum, I researched how courts have treated Confrontation Clause issues relating to artificial intelligence. And, to my dismay, I found remarkably little. In fact, I decided not to include the issue in the moot court problem because there was insufficient case law to construct a fair problem. Instead, after a conversation with another evidence professor, I learned that the legal community cares more about how evidentiary rules apply to artificial intelligence than about any potential Confrontation Clause issues. In 2025, the Advisory Committee on Evidence Rules proposed a new Rule 707 addressing machine-generated evidence that would be subject to Rule 702 if presented by a human expert.[7] The proposal would require courts to assess whether such evidence is helpful, based on sufficient facts or data, produced through reliable principles and methods, and reliably applied to the facts of the case.[8] The Committee continues to revise its proposal.[9]
That debate is important, but it largely concerns the reliability of evidence. And that is where I think the legal community has overlooked something: reliability does not resolve the Confrontation Clause problem. Since Crawford v. Washington, the constitutional inquiry has focused on whether the government has introduced testimonial statements without giving the defendant an opportunity for cross-examination.[10] Yet lower courts confronting machine-generated evidence have generally stopped a step early. They have concluded that such evidence is not a “statement” because a machine cannot be a “declarant.”[11] United States v. Washington, for example, decided that raw machine-generated data were not hearsay because “only a person may be a declarant,” and therefore the data did not implicate the Confrontation Clause.[12]
This is where Professor Andrea Roth enters the conversation. Roth has been a trailblazer in identifying the problems that machine-generated evidence creates for traditional evidence and the confrontation doctrine. Roth argued that machines can create their own “black box” dangers—errors arising from programming, inputs, operation, or the machine-learning process—and proposed safeguards that would give juries foundational information and context necessary to assess the accuracy of machine-generated evidence.[13] In a later article, Roth went further, arguing that treating nonhuman conveyances as categorically outside the Confrontation Clause is difficult to square with the Clause’s “text, history, logic, and principle.”[14] She proposed understanding confrontation as a meaningful opportunity to scrutinize the government’s proof, in whatever form, and to attack that evidence before the factfinder.[15] In her view, the right to confront a witness is not limited to cross-examination of human witnesses.[16]
Roth’s work gets us much closer to the problem I have in mind. If the government wants to use a machine as a source of proof, the defendant should have meaningful access to whatever information is necessary to determine how that machine produced its conclusion.[17] But Roth’s proposal leaves one important question unanswered. What happens when the information the machine learned from consists of statements made by human witnesses?
Artificial intelligence does not “know” anything. Machine-learning systems are trained on data, and they adjust their internal parameters to identify patterns and relationships to apply to new data.[18] In supervised learning, for example, a model learns from training data containing observed inputs and known outputs, allowing it to develop rules for predicting outputs when presented with new inputs.[19] That matters because, once trained, the model can use those learned relationships to predict an output about new data, even when no human has ever made that particular conclusion. But the machine may have learned how to reach that conclusion from statements that humans did make.
Suppose a machine is trained on thousands of drug analysts’ reports. The machine may not reproduce any particular analyst’s statement. Instead, it may learn relationships between characteristics in those reports and conclusions about whether a substance is a particular drug. The final output may therefore be genuinely machine-generated. But the machine’s conclusion may nevertheless depend, at least in part, on human assertions from the training data.
And that brings me to my proposal: Courts should not end the Confrontation Clause inquiry merely because the ultimate output came from a machine. At a minimum, courts should consider whether the machine’s output materially depends upon testimonial statements made by human declarants. If the underlying statements were not testimonial, there may be no Confrontation Clause problem. If they were testimonial, however, giving the defendant access to the machine’s training data may not be enough to vindicate his confrontation rights.
Roth is right that a defendant needs meaningful access to information necessary to scrutinize machine-generated proof. But imagine telling a defendant, “Here is everything the artificial intelligence system was trained on. You can examine the reports and the data, but you cannot cross-examine the people who made the testimonial statements contained within that training set.” This seems to miss Crawford’s point. The Confrontation Clause does not merely guarantee access to information. It guarantees a particular method of testing testimonial evidence: confrontation.[20]
Using evidence generated by an AI system that was trained on human statements does not necessarily violate the Confrontation Clause in every case. And not every statement in AI training datasets is testimonial. But when machine-generated evidence materially depends upon statements made by humans, courts should determine whether the underlying statements are testimonial rather than categorically disregarding the Confrontation Clause because the final output came from a machine.
[1] Briana Rosenbaum, Univ. of Tenn. Winston Coll. of L., https://winston.utk.edu/directory/briana-rosenbaum/ (on file with the Tennessee Law Review) (last visited Oct. 23, 2026).
[2] See Melendez-Diaz v. Massachusetts, 557 U.S. 305, 310–11 (2009) (concluding that forensic analysts’ certificates identifying a substance were testimonial); Bullcoming v. New Mexico, 564 U.S. 647, 657–60 (2011) (concluding that a testimonial forensic report could not be introduced through a surrogate analyst without an opportunity to confront the original analyst).
[3] Smith v. Arizona, 602 U.S. 779, 783, 797–801 (2024) (explaining that when a testifying expert conveys an absent forensic analyst’s statements in support of the expert’s opinion, and those statements provide that support only if true, the statements are admitted for their truth and implicate the Confrontation Clause).
[4] Id. at 783, 796–97.
[5] Id. at 793.
[6] Id. at 783.
[7] Advisory Comm. on Evidence Rules, Report to the Standing Committee 2–3 (Dec. 1, 2025).
[8] Id. at 2–3; Fed. R. Evid. 702.
[9] See Advisory Comm. on Evidence Rules, May 2026 Agenda Book 116–96.
[10] Crawford v. Washington, 541 U.S. 36, 51 (2004).
[11] See, e.g., United States v. Washington, 498 F.3d 225, 230–31 (4th Cir. 2007); United States v. Lamons, 532 F.3d 1251, 1263–64 (11th Cir. 2008); United States v. Moon, 512 F.3d 359, 362 (7th Cir. 2008); United States v. Miller, 982 F.3d 412, 437 (6th Cir. 2020).
[12] Washington, 498 F.3d at 231.
[13] Andrea Roth, Machine Testimony, 126 Yale L.J. 1972, 1977–78, 1989–2000, 2022–51 (2017).
[14] Andrea Roth, What Machines Can Teach Us About “Confrontation”, 60 Duq. L. Rev. 210, 211 (2022).
[15] Id. at 211–12.
[16] Id. at 211.
[17] See id. at 210–12; Roth, supra note 12, at 2027–30.
[18] See Solveig Badillo et al., An Introduction to Machine Learning, 14 Clinical Pharmacology & Therapeutics 871, 871 (2020); Harry Surden, Machine Learning and Law, 89 Wash. L. Rev. 87, 89–91 (2014).
[19] See Badillo, supra note 17, at 875.
[20] See Crawford v. Washington, 541 U.S. 36, 61 (2004) (“[R]eliability [must] be assessed in a particular manner: by testing in the crucible of cross-examination.”); id. (“Where testimonial statements are involved, we do not think the Framers meant to leave the Sixth Amendment’s protection to the vagaries of the rules of evidence . . . .”)
