Skip to content
LPBN-Icon-v1
EvidenceProf Blog
Editor: Colin Miller

Evidence No One Stands Behind

<em>Why FRE Article VII Has Nothing to Grab When No Witness Testifies</em>
By Colin Miller

Guest post by Jon Polenberg

Professor Miller was kind enough to invite me to share a few thoughts here, and I want to use the space to make an argument about Proposed Federal Rule of Evidence 707 that I develop at length in a forthcoming article, Judgment Without a Subject: Artificial Intelligence and Legal Attribution. See https://papers.ssrn.com/abstract=7248440. In the interest of disclosure: I submitted a public comment opposing the proposed rule, and I take part in a Sedona Conference Working Group 1 brainstorming group on AI admissibility. The views here are my own, and do not represent Becker, its clients, or The Sedona Conference.

Let me start with the rule, and then with the problem it could not solve.

What Rule 707 Tried to Do

The Advisory Committee on Evidence Rules saw a real difficulty. Litigants increasingly offer machine-generated output for its truth, often with no expert on the stand. That output takes familiar forms, among them an algorithm’s match, a model’s score, and a system’s risk assessment. If a human expert offered the same conclusion, Rule 702 would govern. The proponent would have to show a qualified witness applying reliable methods to sufficient facts and data. But when the conclusion comes straight from a machine, Rule 702 has nothing to grab because there is no witness.

One point about scope matters throughout. Rule 707 reaches only acknowledged AI. Acknowledged AI means output a party openly offers as a machine’s product and asks the court to credit for its truth. The rule does not reach deepfakes, which present the reverse problem and belong to a different rule. I return to that distinction at the end.

The published draft answered that gap by extending Rule 702’s reliability standard to the machine itself:

Rule 707. Machine-Generated Evidence. When machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702(a)–(d). This rule does not apply to the output of simple scientific instruments.

The Standing Committee approved that text for publication in June 2025, and it went out for public comment that August. The comments divided. Many thought “machine-generated” swept too broadly; some thought courts did not need the rule yet; others wanted an expert foundation required in every case. The Committee held public hearings in January 2026, the comment period closed that February, and at its May 7, 2026 meeting, the Committee declined to recommend the published text, turned to considering a revised draft, and set the subject for further study alongside deepfakes.

So the published version is not the rule. Rule 707 is now an information item rather than text moving through the Rules Enabling Act. Its trajectory is worth studying anyway because what it ran into is structural, not editorial. No redrafting on the reliability model reaches it.

The Move the Rule Made

Here is the move. Rule 702 rests on a person qualified as an expert who applies reliable principles to the facts and then submits to cross-examination. The reliability inquiry tests that witness’s reasoning. Rule 707 kept the reliability apparatus and left the witness behind. It asked whether reliable methods produced the output and whether the system applied them reliably. It could not ask the one question Rule 702 is built around. Whose judgment is this?

That substitution is the whole problem, and it is not new. Alan Turing made the same trade in 1950. He found the question “can machines think?” unanswerable, so he replaced it with a question any observer could score. Can the machine keep an interrogator from telling it apart from a person? Rule 707 switched the question in a similar way to Turing. It replaces “whose conclusion does this output state?” with “was the process that produced it sound?” Each move defends itself on its own terms. Each quietly declines the question its answer will be treated as settling.

Consequence Is Not Judgment

To see why the substitution matters, separate two things the law does. It assigns consequences and it credits judgments. These are different operations.

Assigning a consequence places a burden on someone because of an output. The burden takes ordinary forms, such as damages, exclusion, or invalidity. Law does this constantly without asking whose judgment produced the output. Product liability asks for a defect and a seller. Title VII asks for an employer and an adverse action. Neither asks who decided anything. Tort and employment law absorbed autonomous systems easily because neither ever had to name the subject behind the machine.

Crediting a judgment is different. It means accepting an output as the conclusion of a subject. The court treats the instrument’s reading as a finding, the program’s match as a determination, and the model’s score as an opinion. To credit is to say: this is someone’s call, and I will give it legal effect as such.

Evidence is where the second operation becomes unavoidable. When a party offers machine output for its truth, the factfinder is not pricing an event. She must decide whether to believe a conclusion. The output has stepped into judgment’s place. Rule 707 conceded as much when it imported Rule 702’s standard. Yet no one stands there. That is the difficulty I call judgment without a subject.

The carve-out for “simple scientific instruments” was reaching for this line without naming it. It also answers the natural objection that courts credit thermometers and scales every day with no subject behind them. A thermometer measures; it does not decide whether this patient is febrile under a governing standard. A spreadsheet computes what its operator directs. The operator supplies the rule, picks the inputs, and applies the result.

What makes a modern AI output different is that the system supplies the application itself. Kant’s word for that step is judgment: deciding whether a particular case falls under a rule. No one worries that a scale has no subject because a measurement is not judgment.

A party offers a risk score, a match, or a diagnosis as judgment, the very thing a person would otherwise supply. That is where the subject goes missing. So the claim is not that all machine output needs an author. It is that output offered as a conclusion, in the place a judgment would fill, needs one.

Hearsay: No Declarant to Test

Walk it through the existing rules and the gap surfaces three times. Start with hearsay. The rule polices a declarant’s out-of-court statement offered for its truth, and the Rules define a declarant as the person who made the statement. The whole apparatus runs to that person. Cross-examination tests her perception, memory, narration, and sincerity. Machine output falls outside it, and courts have decided that raw machine data and machine-generated markers are not “statements” by a “declarant” because a machine is not a person.

Hearsay guards against a human witness’s failings. She misperceives the event, forgets it, narrates it loosely, or lies about it. A model does none of that, which is why the machine-source cases come out right. The rule finds nothing to police, and the output clears a screen built for someone else.

As Andrea Roth has shown in her work on machine testimony, the machine has its own sources of error that the declarant model never anticipated. Miscalibration, flawed training data, and coding mistakes each corrupt an output. So, the output does an assertion’s work. It stands where a witness’s conclusion would stand, and the rule that would test a witness never engages. Hearsay does not admit the machine’s conclusion; it fails to stop it.

Authentication: Vouching for the Machine

If hearsay does not reach the output, what does? Authentication does. No one built it to carry that weight. Once the output is not a statement, the proponent needs only to show the item is what she claims. For a system, Rule 901(b)(9) asks whether it produces an accurate result; Rules 902(13)–(14) let certified electronic records self-authenticate.

A lay witness who knows the process can make that showing. She need not have made the determination, understood it, or agreed with it. She testifies that the system ran and that systems of this kind return accurate results. She is vouching for a machine. Ask her whether the applicant really was unqualified, or whether the sample really matched, and she has nothing of her own to say. She ran the system but decided nothing. Her testimony still carries the output to the jury, and the jury credits a conclusion the human in the room will not claim.

So, the proponent satisfies a rule about provenance and receives the benefit of a rule about credibility. That is the swap again, dressed as foundation.

Confrontation: The Borrowed Judgment with No Source

In criminal cases the gap takes on constitutional weight because the Sixth Amendment guarantees confrontation with the witnesses against the accused. A witness is a subject who can take an oath and answer questions. That guarantee assumes someone occupies the witness’s place.

The forensic-evidence line assumes that subject exists. Melendez-Diaz treated the analyst who prepared a certificate as a witness the defendant may confront. Bullcoming held that a surrogate who neither performed nor certified the test will not do.

Williams v. Illinois splintered the Court 4-1-4 over a machine-generated DNA profile. Smith v. Arizona cleared away the plurality’s “not for its truth” rationale in 2024. When an expert relays an absent analyst’s out-of-court statements for their truth, the Clause applies, because that expert’s opinion borrows whatever authority it has from an absent person. The Court left for remand whether those particular statements were testimonial. The problem lies in whose determination the factfinder credits, not in who happens to testify.

Now run the machine case through Smith. The technician on the stand borrows the conclusion the same way an expert borrows an absent analyst’s. The Clause finds no absent analyst to reach. The determination came from a source it cannot classify as a witness. Smith identified the borrowed judgment and named its source. The machine case supplies the borrowing without the source. It is the Bullcoming surrogate problem at its limit: not the wrong analyst, but no analyst whose judgment the output expresses.

Confrontation’s machinery presupposes that its mechanisms land on someone. A witness who lies faces perjury. One who equivocates faces the jury’s read on her demeanor. An expert who overstates faces impeachment. Each works because it lands on a subject. None lands on a machine. This is not a confrontation violation in the ordinary sense; it is something stranger. If the output is not a witness’s statement, the Clause has nothing to run against. The accused may then face proof she cannot confront at all.

Why the Rule’s Stall Changes Nothing

Here is the part that matters for where we are now. Rule 707 did not create this gap, and pulling the published text back did not close it. The three doctrines still face the same machine and register that the output has taken judgment’s place, but no one stands there to bear it.

The Committee could have written a rule that required a person to adopt the output. A qualified expert would read the machine’s conclusion, agree with it, and say so on the stand. Then Rule 702 would govern honestly, and the other side would have someone to cross-examine. Rule 703 already shows the law knows how to make an expert own the material she relies on. The published draft took the other path. It tested the process and let the conclusion in without anyone standing behind it. That choice was deliberate, and it is the difficulty in miniature. The Committee offered reliability as a substitute for a subject.

The Cost of Leaving the Subject Unnamed

Why press this if the published rule is not moving forward as written? Because the gap does not wait for a rule. Litigants are already offering these outputs, and courts are already crediting them. The machine-source cases mark the leading edge. Roth’s work documents the wider trend.

When a court credits machine output as judgment but names no subject, it does not leave the subject’s place empty. The status is not conferred; it accretes. A court credits a “determination” or an “assessment,” and the output becomes someone’s conclusion by default before anyone decides it should be. Call it electronic personhood by accretion. Scholars have debated a deliberate, bounded status and mostly rejected it. This is something else, an unlegislated artifact that crediting alone produces.

And the doctrines treat the same output inconsistently. Copyright gives it no author. Tort attaches it to a developer, deployer, or user. Evidence credits it as reliable. That’s what happens when law asks its consequence questions in every register while no one asks the prior question. Whose judgment is this?

The deepest cost is structural. Cross-examination presupposes a witness. Appellate review works on stated grounds, and stated grounds come from a reasoner. Liability presupposes an actor. Strip the subject from judgment and these mechanisms do not adapt cleanly; they strain toward an object that is not there. A determination that law credits as judgment but attributes to no one stands beyond the kind of review that presupposes a reasoner.

What Rule 707’s Successor Has to Confront

The Committee has not abandoned the field but divided it instead. It continues to study a revised Rule 707. It is also weighing a new Rule 901(c) aimed at deepfakes and revisiting Rule 703 after Smith v. Arizona. The division of labor is worth stating precisely because the deepfake work is sometimes treated as though it answered the Rule 707 problem. It does not.

Rule 707 governs acknowledged AI, which a party openly offers as the product of a machine and asks the factfinder to credit for its truth. Deepfakes are unacknowledged. Someone passes fabricated material off as authentic and conceals that a machine made it. One is disclosed judgment with no subject behind it; the other is hidden forgery.

A deepfake rule polices authenticity. It says nothing about who, if anyone, stands behind an output the parties agree a machine produced. So even a fully successful Rule 901(c) would leave the missing subject exactly where Rule 707 found it.

So I am not arguing that law must confer some new status on machines, or that any particular reform is the answer. My claim is narrower and prior. Before the Committee decides how to admit machine-generated evidence under Rule 707, Rule 901, or anywhere else, it has to decide what the admission is doing. If the output is a conclusion, then the rule needs a subject who adopts it and can answer for it. Otherwise the rule has to say plainly that it admits a conclusion no one will stand behind. It then has to accept what follows for cross-examination, confrontation, and review.

Reliability is a genuine value, and I do not doubt that machine evidence can be reliable. But reliability answers a question about process. Suppose a model never hallucinated, ran the same way every time, and answered correctly on every question. It would still not answer the question a factfinder actually faces when she credits an output for its truth: whose judgment, if anyone’s, is this? Getting the answer right is not the same as owning it. Rule 707 as published could not ask that question. Its successor should.

My thanks to Professor Miller for the space. The longer argument appears in Judgment Without a Subject: Artificial Intelligence and Legal Attribution. That article sets out the personhood theory behind this post and a July 2026 incident that tests the same gap in criminal law.