The machine was right. I should say that at the top, because it is the part that makes the rest uncomfortable.
Some years ago, in a country that posts its speed limits in kilometers, an envelope arrived with my license plate on it, a date, a time accurate to the second, and a number I had no memory of reaching. I looked at the photograph for four seconds — long enough to confirm the car was mine — and paid.
Here is the confession. I never checked the camera. It did not occur to me to ask when that device had last been calibrated, or who signed the certificate, or what would have happened if I had written back and said: prove it. I audited the photograph. I never audited the machine.
Small failure, large shape. In most legal systems, "prove it" lands back on the person making it: the device starts out presumed correct, the human starts out arguing uphill. That allocation is the whole fight — and on May 7, 2026, the committee that writes America's rules of evidence sat down to have it about a much larger class of machines. They debated for two hours, then walked away until fall.
What the committee actually did on May 7
I came expecting to write a different article: that proposed Federal Rule of Evidence 707 — the first federal rule written specifically for evidence a machine produced — had cleared its last serious hurdle, and that the earliest it could bind anyone was December 1, 2027. Both halves are now wrong.
The report the evidence committee filed in May says the Advisory Committee on Evidence Rules "does not recommend action on the proposed Rule 707 at this time". It has instead revised the rule and plans "further study on it and another issue relating to artificial intelligence — the problems posed by 'deepfakes' — at its next meeting." Bloomberg Law put it less diplomatically: the committee "can't agree on whether to proceed", "never mind the best approach."
Then the sentence that kills the calendar. The committee agreed the revised rule "would require re-publication" were it to go forward — and then "decided not to propose that the revised Rule 707 be released for a new period of public comment at this time." A rewrite big enough to need a fresh national comment round, with none scheduled.
Not cold feet, exactly. The committee wanted the draft vetted at its fall meeting "by technology experts and others in the field of AI and law," input that would help it "determine whether an amendment is necessary". After a published rule, two hearings and a full comment file, the live question is still whether there should be a rule at all. That meeting is October 15, 2026, in Boston, "open to the public for observation but not participation," and it sits on the judiciary's own calendar.
Federal rulemaking runs through a minimum of seven stages of formal comment and review. If a rule survives them, the Supreme Court transmits it to Congress by May 1, and absent congressional action the amendments "take effect as a matter of law on December 1."
Start to finish, the judiciary says, the process usually takes two to three years. Rule 707 was not rejected; it was sent back to a stage it had already cleared. I will not run the arithmetic forward and hand you a new date: no document states one, and a rule that needs republishing before it re-enters the pipeline has no knowable end. The clock did not tick down. It reset.
The whole rule, and the hole it was written to close
Here is the rule as published for comment in August 2025, in full:
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.
Rule 702 is the expert-witness rule: before a witness may tell a jury what something means, the side offering that opinion must show that it will help, that it rests on sufficient facts or data, and that it comes from reliable methods reliably applied. A person vouches, and a person can be cross-examined.
Now the hole — and the committee's own explanation is the clearest I have read anywhere. Where a testifying expert relies on AI, the reliability requirements bite. But if AI is offered without one — "for example through a witness who applied the program but knows little or nothing about its reliability" — then "Rule 702 is not obviously applicable". "Yet it cannot be that a proponent can evade the reliability requirements of Rule 702 by offering AI output directly or through a lay witness."
That inverts the framing I arrived with. I thought the gap was a flaw in Rule 707. The gap is today's law; Rule 707 was the patch.
Andrea Roth, the Berkeley evidence scholar who has been in the room for this rule from the start, made it concrete: under the status quo a judge could admit "an analytical conclusion by Cellebrite technology through a lay witness without any Rule 702 scrutiny". Not the raw data — the conclusion. Her summary, per the minutes: untenable.
The committee was equally blunt about why a machine is not simply a quiet witness. "A human expert can be cross-examined… But it may be more difficult to attack the weight of AI output." So its draft note tells judges to consider instructing juries that "evidence should not be assumed to be reliable simply because it was produced by a machine" — a warning label, not a rule of procedure.
The version that emerged in May is a different animal. The rewrite even changed the rule's name, to "Evidence Produced by Artificial Intelligence and Presented at Trial Without an Expert," and made the burden explicit: "the proponent must establish" that the evidence will help the trier of fact, rests on sufficient facts or data, and is the product of reliable principles and methods reliably applied.
It also adds a notice provision and "focuses on 'artificial intelligence' rather than 'machine-generated' information" — the loudest single demand in the comment file.
The evidence is already in the building
It is tempting to file this under things that might matter later. The courtroom record says otherwise.
In 2024 a Washington state trial judge in State v. Puloka refused to let a jury watch AI-enhanced video of a shooting: the court held that the enhancement tools, which use machine-learning algorithms, "have not been peer-reviewed by the forensic video analysis community, are not reproducible by that community, and are not accepted generally in that community."
The older machines are already deep in the system. In material the Federal Judicial Center publishes for federal judges, probabilistic genotyping systems — software that untangles DNA mixtures too messy for a human analyst to read — are "the most widely adopted methods for interpreting complex DNA mixtures." One of them, TrueAllele, had been used in more than 850 criminal cases when The Markup examined it in 2021, and the government crime labs running it "don't get access to the program's source code."
The ACLU has tracked one woman as the fourteenth person in the United States wrongfully arrested after police relied on flawed facial recognition; "at least thirteen other people are publicly known" to have been arrested the same way before her. Publicly known is a floor, not a count.
And here is how such material reaches a jury today. In its comment on Rule 707, the ACLU described the ordinary practice: prosecutors often seek to admit evidence through police officers testifying that they used an investigative tool as they were trained to do, "even though the officers have no knowledge of how that tool works."
The best argument against the rule comes from the prosecutors
The other side's case is stronger than I expected, and it arrives from two directions.
The Justice Department does not want Rule 707 at all. Speaking for the Department in May, Ms. Shapiro said the DOJ "would rather have no Rule 707 at all" and that the subject remains "an anticipatory problem that has yet to materialize in court". That is a characterization of the future, not a finding about the present — but it comes from the federal system's largest litigant, and when the committee voted to publish the draft at all, the single vote against was the Justice Department's, 8 to 1.
The second objection comes from the opposite corner of the bar, and is the more serious one. The American Association for Justice — the plaintiffs' trial bar — called the rule "over-inclusive," warned it "will result in satellite litigation over the use of routinely admitted evidence," and asked the committee to pause for redrafting. Its list of what would be swept in lands hard: surveillance video, geolocation data, and "specific machines, tools, technology or datasets that are a core component of a job or profession." The published rule said machine-generated, not AI-generated.
The comment file registers that unease precisely: two public hearings, 59 written comments by the deadline, and of those only 3 expressed unqualified support, 27 supported the rule subject to revisions, and 27 opposed it. The committee had also said, on release, that it was "not treating release for public comment as a presumption that the rule should be enacted".
So that is settled, then. Except the same minutes carry the rebuttal from a few chairs away: Ms. Owens said such a rule is not premature, because federal defenders are already seeing this kind of evidence coming into cases.
And many commentators, the committee recorded — including the ACLU and the New York City Bar Association — said the problem is occurring today, especially in criminal cases. Same room, same afternoon, two accounts of the same present tense.
Australia runs the same problem from the opposite end
Now the comparison that reframed the whole thing for me — and it starts, appropriately, with my speeding ticket.
Australia's uniform evidence law does the precise opposite of what Rule 707 proposes. As the Australian Law Reform Commission describes it, section 146 "creates a rebuttable presumption that … if the device or process is one that, if properly used, ordinarily produces a particular outcome, then in producing the document or thing on this occasion, the device or process has produced that outcome." In plain English: the machine is presumed to have gotten it right, and whoever doubts it carries the burden — which, in a criminal case, usually means the person with the least money and the fewest experts.
The Commission's worked example tells you which era the provision was built for: under section 146, it would not be necessary to call evidence to prove that a photocopier "normally produced complete copies of documents and that it was working properly when it was used to photocopy the relevant document." Sensible for a photocopier; considerably more interesting for a model whose output nobody can reproduce. And this is live law — the Evidence Act 1995 is still on the books as a Commonwealth statute in force, in a compilation registered in 2025.
There is a second inversion behind the first. American evidence law spent three decades building a gatekeeping test for expert reliability; Australian courts went the other way. After the Honeysett decision, Professor Roberts wrote on Melbourne Law School's High Court blog that the provision governing expert opinion "has been interpreted in a way that precludes consideration of the issue of reliability," and that the courts have steered well clear of Daubert. That is his reading, and he says so: "It is difficult to avoid the conclusion that the High Court was intent on avoiding discussion of reliability as a condition of the admissibility of expert evidence."
So you would expect Australia to be miles behind on AI in the courtroom. It is ahead — on a different track. There is still no blanket rule that admits or excludes AI outputs; standard principles apply "with additional layers of scrutiny." What moved was everything around the statute. The Victorian Law Reform Commission's report was tabled in Parliament on February 3, 2026 with 30 recommendations, including that judicial guidelines "prohibit the use of AI tools for judicial decision-making" — the first inquiry by an Australian law reform body into AI in courts and tribunals, which found more than a third of surveyed lawyers, experts and self-represented litigants in Victoria had already used AI there. Then in April, while Washington was still redrafting, the Federal Court of Australia shipped binding practice guidance, released April 16, 2026: where AI has been used to "summarise, analyse or generate material that informs evidence or opinion (or its admissibility)," those uses "must be transparently disclosed at the start of the document."
Two systems, two speeds, opposite defaults. Australia moved fast on who must tell the court an AI was involved, and has not touched the burden question. The United States is trying to move the burden and cannot get the proposal out of committee. Only one of them has a statute that presumes, on the page, that the device got it right — and if you have ever paid a camera ticket without asking a question, you know how that feels.
The courtroom this is heading toward
Now let me get imaginative, because the trend line is not hard to read.
Picture an ordinary robbery trial five years from now. No confession, no eyewitness who saw a face. The state's case is four machines: a face-match narrated by the detective who typed in the query; a route reconstruction built by a model from license-plate reader pings and cell-tower records; a risk flag from an analytics package the department licenses; and an interrogation transcript translated by a model, no human interpreter in the room. Nobody in that courtroom built any of it. For each item the witness is the person who pressed the button, and each can testify, truthfully, that they pressed it exactly as trained.
Then the defense asks the only question that matters: how did the route model weight the ambiguous pings, the ones putting the car in two places at once? Nobody in the building knows — which is precisely the case the committee's draft note anticipated, where an AI process "can sometimes develop in such a way that nobody is able to explain how the system has reached a result, because the machine has developed the ability to program itself." Under the rewritten rule, if the process cannot be explained, "the court should in most cases find that the proponent has not established more likely than not that the methodology is reliable." Under the rules we actually have, it is a question put to a witness who cannot answer it — and then the jury retires.
Push one step further: if courts eventually demand a reliability showing, vendors will sell one — court-ready report modes, validation summaries written by the company selling the tool. The photocopier presumption would not be repealed so much as re-created commercially.
The people who have read the whole file
The most interesting critique in the comment file came from neither camp you would expect.
A six-organization coalition — the ACM's US Technology Policy Committee, Asian Americans Advancing Justice | AAJC, the Center for Democracy & Technology, EPIC, Fight for the Future and UnidosUS — argued that borrowing Rule 702 for machines is the wrong instrument entirely. The criteria for assessing an AI system's reliability differ from those for assessing an expert witness's testimony, they wrote, "making this an improper fit." AI output is more reliable when the system was trained on unbiased data of high quality, and nothing in Rule 702 assesses the training data — the rule looks at the knowledge and experience of a human expert, because that is what it was built to look at.
The New York City Bar Association landed nearby from a more institutional angle: it supports adoption of rules governing machine-generated evidence, "however, we strongly recommend revisions to make more explicit the issues targeted by the evidence rule," and argued that such evidence "should principally be introduced through expert testimony."
The deepest version of the argument is nine years old. In Machine Testimony, published in the Yale Law Journal in 2017, Roth wrote that courts "shoehorn" machines "into existing rules by treating them as 'hearsay,' as 'real evidence,' or as 'methods' underlying human expert opinions" — attempts that are intellectually incoherent and "fail to fully empower juries to assess machine credibility."
Notice the shape of that disagreement: the Justice Department and the plaintiffs' trial bar both want the brakes pulled, for opposite reasons, while the ACLU, the federal defenders and a coalition of technologists all want a rule — and several of them think this one is built on the wrong chassis.
What to do with this, if you ever end up in a courtroom
You are far likelier to meet this as a juror than as a defendant, and that is where the leverage sits.
If you are called for jury duty, listen for who is actually vouching. When a number, a match or an animation appears, ask yourself: did this witness build the tool, run it, or merely receive its output? The committee's draft note warns that machine output should not be assumed reliable simply because a machine produced it. Nobody may say that out loud in your trial. You are allowed to think it.
If you are ever a party, ask one question early and loudly. Is any evidence against me the output of software? If so: who sponsors it, what version, and can the other side produce a human who can explain how it reached that result?
If your work generates records, assume they are evidence. Claims software, HR screening, fraud scoring, security analytics. Keep the version numbers and validation reports, and know who at the vendor could testify.
Do not confuse this with the deepfake problem. Both travel to the same October meeting, but the committee itself calls deepfakes "another issue": one asks whether a recording is real, the other whether a machine's conclusion is reliable.
If you want a say, watch for republication — and for October 15. Fifty-nine comments shaped what happened in May, and the revised rule needs a new comment round before it can go forward. The Boston meeting is open for observation.
The default is the whole argument
My speeding ticket taught me this years too late: someone always has to prove the machine right, or prove it wrong, and a rule of evidence is just a decision about which of those jobs goes to whom. Australia assigned it to the objector, in a rule built around a photocopier. Rule 707 would assign it to whoever wants the jury to believe the output. Neither default is neutral, and pretending there is a neutral option is how defaults get set by accident.
On May 7 the United States declined, for now, to choose — and declining is itself a choice, because the existing default keeps running in the meantime. The fall conference may well produce a better rule, narrower and aimed at AI conclusions rather than every surveillance camera in America. But I would rather a delay be honest than invisible. So here is the sentence I would want on the record: as of today, in a federal courtroom, if nobody sponsors the machine, nobody has to prove it. Not the government, not the vendor, not the officer who typed in the query. The jury simply gets the answer.
My vote? Fix the chassis, not the wording. If Rule 702 was never built to interrogate training data, bolting machines onto a rule about human expertise will keep producing sentences judges cannot apply. Write the reliability test machines actually need — provenance, version, validation, error rate, and who can explain it — and write it while the technology is still arriving, not after a decade of verdicts has settled the question in the vendors' favor.
Somebody always has to prove the machine right, or prove it wrong; a rule of evidence is just a decision about which one of you it is. The HAIA Foundation spends its time on that sentence — and if you would rather learn how it works before the summons arrives, subscribe.