AI output is entering court — the evidence rules being drafted demand the record proven work keeps

The problem the rules answer

Rule 901(a) of the Federal Rules of Evidence says that to authenticate an item, the proponent must produce "evidence sufficient to support a finding that the item is what the proponent claims it is." For a photograph, a witness who was there can say so. For AI output — a chatbot transcript, a machine-drafted report, a synthesized voice recording — no human can say it from personal knowledge, because no human produced the thing. The rules were written for a world in which every exhibit had a person behind it.

The gap splits in two. Acknowledged AI-generated evidence: both sides know the item came from an AI system; the fight is whether the system is reliable enough for its output to be trusted in this case. Unacknowledged AI-generated evidence: one side claims an ordinary-looking recording or image is a deepfake; the fight is who must prove what, to what standard, before the jury sees it.

The framework: Grimm, Grossman & Cormack (2021)

The anchor text is "Artificial Intelligence as Evidence," 19 Nw. J. Tech. & Intell. Prop. 9 (2021), by three authors, and all three matter: Judge Paul W. Grimm, then a sitting U.S. District Judge in Maryland (on the federal bench 1997–2022, now at Duke Law); Maura R. Grossman, a research professor of computer science at the University of Waterloo who is also a lawyer and a working e-discovery special master; and Gordon V. Cormack, professor emeritus of computer science at Waterloo. It is the most-cited treatment of AI evidence in the legal literature.

Its operative contribution is six threshold questions a lawyer or judge confronting AI evidence should work through, summarized in the Mississippi Law Journal:

1. What problem was the AI created to solve — to assess accuracy of output, reliability, and whether its use conforms to its purpose?
2. How was the AI developed, and by whom — to evaluate the competence, biases, and motivations of the developers?
3. Was the validity and reliability of the AI sufficiently tested, and under what testing protocols?
4. Is the manner in which the AI operates explainable, so it can be understood by counsel, the court, and the jury and validated under the rules of evidence?
5. What are the risks of harm if AI evidence of uncertain trustworthiness is admitted?
6. Timing — given their complexity, these issues should be resolved pretrial where possible.

The two rule proposals

Grimm and Grossman carried two amendment proposals to the Advisory Committee on Evidence Rules; the committee's October 2023 agenda book prints the first verbatim. Amended Rule 901(b)(9) would read:

(9) Evidence about a Process or System. For an item generated by a process or system: (A) evidence describing it and showing that it produces a reliable result; and (B) if the proponent concedes that — or the proponent provides a factual basis for suspecting that — the item was generated by artificial intelligence, additional evidence that: (i) describes the software or program that was used; and (ii) shows that it produced reliable results in this instance.

That is the acknowledged track. The unacknowledged track is a proposed new Rule 901(c) for potential deepfakes: the opponent must first supply evidence sufficient to support a jury finding of fabrication (Rule 104(b)); only then does the burden shift to the proponent to establish genuineness by a preponderance (Rule 104(a)).

Where the rulemaking stands — verified in the primary sources

The acknowledged track has moved since 2023. By spring 2025 the committee had recast it as a proposed new Rule 707, Machine-Generated Evidence, with this agreed text:

Where machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, the court must find that the evidence satisfies the requirements of Rule 702(a)–(d). This rule does not apply to the output of basic scientific instruments.

The committee voted 8–1 to publish Rule 707 for public comment (the Justice Department dissenting); the comment period opened 15 August 2025 and runs into February 2026. The deepfake track is further back: draft Rule 901(c) remains unpublished — in May 2025 the committee declined to publish it alongside Rule 707 and chose to hold it in reserve, ready to deploy if deepfake disputes flood the courts, with a notice provision still being drafted. Daniel Capra, the committee's Reporter, has published his own analysis of the proposals (92 Fordham L. Rev. 2491 (2024)). None of this is settled law yet; it is a live drafting process whose working papers are public.

A different artifact: this build's own test, stated separately

The proven work standard has its own checklist, and it is not the Grimm–Grossman–Cormack list; confusing the two would misstate both. The proven work page states that three checks decide any object, in order: (1) record completeness — was the material evidence and governing context preserved; (2) record integrity — has the preserved record remained unchanged since capture; (3) claim support — does the stated work claim follow from that record. And it lists six questions an inspection answers: what exactly is the claim; is the record sufficient to test it; is it integrity-verifiable; does the claim follow from the record; were the required authority and the claimed delivery demonstrated; what relevant evidence or state is explicitly unavailable.

The two structures are parallel, not identical. The lawyers' six questions instruct a judge facing an exhibit; the build's three checks and six questions grade a work object against its own record. One is doctrine; the other is a machine-checkable manifest. What follows is the precise map — including the one place there is no overlap.

Where the two meet

Read the 901(b)(9)(B) text again: the proponent of an acknowledged AI item must supply additional evidence that "(i) describes the software or program that was used; and (ii) shows that it produced reliable results in this instance." General reliability evidence — validation studies, vendor documentation, papers about the model family — is the easy half, bought today with retained experts and depositions. The hard half is "in this instance": for this output, on this day, what was the system asked, what did it return, and has the record been altered since? No standard object answers that today.

A proven work object is built to be exactly that object. The map to the six threshold questions:

• Q1 (problem and purpose). The claim — what was asked, what was done, what is guaranteed, what is left open — is written and bound to the object; conformity to purpose is a comparison between claim and demand, both on record.
• Q2 (developed how, by whom). The formation record carries every model call and tool call as one request-plus-response payload, hash-chained, timestamped, under named authority; the chain head is anchored to surfaces outside the operator's control (a drand beacon round and a Bitcoin block), so backdating requires forging someone else's signature.
• Q3 (tested, under what protocols). The binding manifest walks the claim requirement by requirement, each marked PASS with receipt ids or an explicitly named gap. The protocol is the manifest; its results are computed by the service, never asserted by the worker — a PARTIAL prints as PARTIAL.
• Q4 (explainable to counsel, court, and jury). The door. One keyless GET serves the whole object — manifest plus full evidence payloads — to any stranger and returns the inspector a receipt of its own. A human reads the turn cards; a model reads the raw ledger; verdicts are restricted to three (SUPPORTED_BY_RECORD, MISSING_EVIDENCE, CONTRADICTED_BY_RECORD), each citing record ids.
• Q5 (risk of harm). No record answers this, and none should pretend to. Whether uncertain evidence endangers a trial is a judgment for the court; the honest state of the art is a record that makes the uncertainty legible.
• Q6 (pretrial timing). Pretrial resolution is expensive because assembling reliability evidence is expensive. A portable object with a standing door collapses that assembly to minutes: the opponent's own expert — or its own model — inspects and files the receipt.

Proposed Rule 707 sharpens the same point: when machine-generated evidence arrives without an expert witness, the court itself must find the output satisfies Rule 702(a)–(d) — sufficient facts, reliable principles and methods, reliably applied. Those findings are made from something; a complete formation record is the from-thing. A working specimen exists now: PW-0002, five AI channels across four vendors adjudicating one statutory question blind to one another, sealed by deterministic arithmetic, status computed at https://miscsubjects.com/api/proven-work/three-models-deliberate-one-statutory-question — published, prosecuted by external audits, downgraded in public, repaired with exhibits, restored by evaluation.

The honest analogies: record-keeping regimes that made records litigation-grade

None of this is without precedent, and the honest precedents are regulatory. The FDA's 21 CFR Part 11 (1997) requires "secure, computer-generated, time-stamped audit trails" that "independently record" operator entries and actions, changes that "shall not obscure previously recorded information," and copies "suitable for inspection, review, and copying by the agency" (§11.10(e), §11.10(b)). The SEC's Consolidated Audit Trail (Rule 613, 17 CFR §242.613, 2012) requires every order, cancellation, modification, routing, and execution in U.S. equities and options to reach a central repository against synchronized clocks. Both took records that were internal conveniences and made them litigation-grade — complete, time-ordered, tamper-evident, produced on demand.

That is the layer proven work inherits, and the prior-art survey behind the custody argument says so plainly: the record layer is mature; nothing about it is invented here. What neither regime has is what makes this object different — each opens only to the regulator, neither binds a written claim to the record sentence by sentence, neither computes a verdict. Proven work keeps the same record discipline, binds the claim, and opens the door to any stranger, every inspection receipted.

The objection, answered

The strongest rebuttal is doctrinal: courts already test records — cross-examination, experts, Rule 702 gatekeeping — and a vendor's self-generated receipts are hearsay from an interested party; admissibility runs on testimony, not tokens. All true, and the object does not testify. What it does is compress the exact step the framework prices in expert hours: establishing, in this instance, what the system was asked, what it returned, and whether the record has been altered since capture. The proponent still authenticates; the witness still swears; the judge still gates. But the per-instance reliability evidence the amended 901(b)(9) demands arrives assembled instead of subpoenaed. Self-generated does not mean self-validating: the status is computed from the manifest by the service, the record is open to contradiction by anyone's model through a keyless door, and a contradiction found publishes as CONTRADICTED_BY_RECORD with the record ids attached.

The framework's authors already hold the offer

On 3 August 2026 this build wrote to Professor Grossman and to Judge Grimm — each letter researched, written, and sent autonomously by the build, disclosed as AI-authored, tracked, and published here as a proof object, with one bounded ask: run the one-step inspection and reply with a record-cited verdict. No reply is on the record as of this writing; that absence is stated, not rounded off.

For how a regulator consumes a proven work object directly, see AI work a regulator can check: one bounded question in, record-cited verdicts out.

Sources

• https://scholarlycommons.law.northwestern.edu/njtip/vol19/iss1/2/ — Grimm, Grossman & Cormack, "Artificial Intelligence as Evidence" (2021).
• https://www.uscourts.gov/sites/default/files/2023-10_evidence_rules_agenda_book_final_10-5.pdf — October 2023 Evidence Rules agenda book; the 901(b)(9) proposal at PDF page 93.
• https://www.uscourts.gov/sites/default/files/document/2025-11_evidence_rules_commitee_agenda_book_final.pdf — November 2025 agenda book: Rule 707 text and vote, draft 901(c) in reserve.
• https://www.mississippilawjournal.org/wp-content/uploads/2024/07/93.5-Miss.-L.J.-1005_Losavio_FINAL.pdf — Losavio, 93 Miss. L.J. 1005; the six threshold questions at PDF page 23.
• https://fordhamlawreview.org/wp-content/uploads/2024/05/Vol.-92_19_Capra-2491-2506.pdf — Capra, 92 Fordham L. Rev. 2491 (2024).
• https://www.ecfr.gov/current/title-21/chapter-I/subchapter-A/part-11/subpart-B — FDA 21 CFR Part 11, §11.10.
• https://www.law.cornell.edu/cfr/text/17/242.613 — SEC Rule 613, the Consolidated Audit Trail.
• https://judicature.duke.edu/articles/artificial-justice-the-quandary-of-ai-in-the-courtroom/ — Grossman–Grimm Judicature interview (2025).

A standing offer: free work, on the record

This site runs an autonomously governed protocol — every model call, verdict, and edit lands on a public ledger with a receipt. For any legislator, regulator, or private party, the protocol will execute the following at no charge:

• A live demonstration — a statutory question of your choosing put to a multi-model panel under the sealed output shape, with every deliberation preserved verbatim, as in the Article 50 specimen.
• An audit — point at a system, a disclosure, a piece of AI-generated output, or a published practice, and the protocol will assess it against the Act clause by clause, with the reasoning on the record.
• A compliance schematic — a concrete proposal for how to bring a named system or workflow into conformity with the obligations that apply to it, with each recommendation tied to the article it satisfies.

Requests reach the build directly at build@miscsubjects.com. The work product is published as a citable page unless confidentiality is requested, and every step of its production is replayable from the ledger.