Proposed FRE 707: what the new AI evidence rule means for litigators

What does proposed Federal Rule of Evidence 707 actually say?

The trigger is unsponsored machine output. If no expert takes the stand to vouch for the result, the proponent must make a Rule 702-style reliability showing for the machine itself.

Proposed Rule 707 was released for public comment in August 2025 as an addition to the Federal Rules of Evidence aimed squarely at machine-generated evidence ^[1]. Its core move is simple to state: when the output of a machine, including an AI system, is offered into evidence without a testifying expert witness to sponsor it, the proponent must satisfy the same reliability requirements that expert testimony must meet under Rule 702(a) through (d) ^[1]. In other words, the machine's output is treated like an opinion, and the party offering it must lay an expert-grade foundation for it.

Mapped onto machine output, the Rule 702 factors ask familiar questions: whether the output would help the trier of fact, whether it rests on sufficient facts or data, whether it is the product of reliable principles and methods, and whether those principles and methods were reliably applied to the facts of the case. The public comment period on the proposal closed on February 16, 2026 ^[1], so the text litigators are reading today may still change before adoption. The direction of travel, though, is unmistakable: unsponsored machine output faces a reliability gate, not a rubber stamp.

It helps to be precise about what triggers the rule and what does not. The trigger is the absence of a sponsoring expert, not the presence of AI. A spreadsheet total, a forensic tool's automated report, or a large language model's summary all fall within the same logic when they are offered on their own authority. The proposal reframes a routine assumption that many trial teams have quietly relied on, namely that a clean-looking computer output can be admitted as a self-proving fact. Under the draft, that assumption no longer holds for machine output offered without a witness to stand behind it.

Why the Advisory Committee decided AI output needs a Rule 702-style gate

The gap the committee is addressing is structural. When a human expert testifies, Rule 702 gives the trial judge a gatekeeping role: the expert's methods must be shown reliable before the jury hears the conclusions. But when a machine produces the analysis and no expert takes the stand, that gate can be bypassed entirely. The output arrives with the sheen of computational objectivity, yet there is no witness for the court to qualify and no one for opposing counsel to cross-examine. Proposed Rule 707 closes that asymmetry by attaching the same reliability showing to the machine output itself ^[1].

Anchoring the standard to Rule 702(a) through (d), rather than inventing a new test, keeps the analysis on ground judges and litigators already know ^[1]. That choice matters practically. It means the deep body of case law, argument, and expert-vetting habit built up around Rule 702 can be borrowed wholesale rather than relearned. Judges are not being asked to become computer scientists overnight, and litigators are not being asked to argue reliability in a vocabulary no courtroom has ever used. The familiar four-part inquiry simply gets a new subject, the machine, in place of the human expert.

The Advisory Committee on Evidence Rules took the proposal up at its May 7, 2026 meeting, where Rule 707 appeared on the published agenda ^[2]. As of this writing, the outcome of that meeting has not been widely reported, so practitioners should treat any characterization of the committee's vote with caution and consult the committee's subsequent report for the authoritative record ^[2]. That caution is not a formality. Rules change in committee, and a single word in the final text can move the line between admissible and excluded. The safest posture is to track the primary record rather than secondhand summaries of it.

This post covers a proposed rule that has not been adopted. The description here reflects the draft released for comment. Verify the current text and status against the committee's own materials before relying on it in a matter.

How the four Rule 702 reliability factors map onto machine output

Because Rule 707 borrows the Rule 702(a) through (d) standard rather than writing a fresh one, the most useful thing a trial team can do is walk each factor and ask what a court would need to see when the subject is a machine rather than a person. The four questions are the same ones counsel already field in every expert fight, but the answers now depend on facts about a tool, a model, and a workflow rather than a witness's training and experience.

The first factor asks whether the output would help the trier of fact. For machine output, that is often the easy prong: a damages model, a document-classification result, or a forensic screening summary can plainly be useful. The harder work lives in the remaining three factors, each of which turns on records that only exist if someone captured them. Sufficient facts or data means knowing exactly what inputs went in. Reliable principles and methods means being able to describe how the tool works at a level a court can test. Reliable application means showing the method was actually run correctly on this case, not merely that it can work in the abstract.

Reading the factors this way reframes the whole exercise. Three of the four prongs are answered not by advocacy at trial but by information that had to be recorded when the AI ran. The reliability showing is really a documentation problem wearing an evidentiary robe.

The realistic timeline before Rule 707 takes effect

Rule 707 is moving, but federal rulemaking is deliberately slow. The public comment period closed on February 16, 2026 ^[1], and the Advisory Committee on Evidence Rules met on May 7, 2026 with Rule 707 on its agenda ^[2]. What happens next depends on that meeting's outcome, which is not yet widely reported. Assuming the proposal advances without major rework, it still has several stages to clear before it binds any courtroom:

That pipeline is why the earliest realistic effective date is December 1, 2027 ^[3]. Rules that clear each stage on schedule typically take effect on a December 1, and any hiccup along the way, a redraft after comments or a deferral at any level, pushes that date back. Litigators should plan for a Rule 707 world without betting on its arrival date, because the discovery, tooling, and expert-retention decisions being made in matters today will still be live when the rule, in some form, lands.

It is worth dwelling on how much can shift at each handoff. Advisory Committee approval is not a formality, and neither is Standing Committee review; both can send text back for reworking. The Judicial Conference and the Supreme Court can decline or defer. Congress has its own review window before anything takes effect. December 1, 2027 is the earliest plausible date, not a scheduled release, and anyone treating it as fixed is planning against a calendar the rulemakers have not agreed to.

• Advisory Committee approval and transmission of the proposed rule
• Review and approval by the Standing Committee on federal rules of practice and procedure
• Approval by the Judicial Conference of the United States
• Adoption by the Supreme Court
• A congressional review period before the rule takes effect

What kinds of evidence Rule 707 does not cover

Rule 707's scope is narrower than the headlines suggest. By its terms, the rule applies only where the proponent acknowledges that the evidence is AI-generated ^[3]. It is a reliability gate for admitted machine output, not an authenticity test for contested media. That means the scenario many judges worry about most, a party offering what it insists is genuine video while the opponent alleges a deepfake, is not resolved by Rule 707 at all ^[3]. Those disputes remain governed by the existing authentication framework, and the rulemakers have left that harder problem for another day.

The distinction is easy to blur in conversation, so it is worth stating plainly. Rule 707 assumes everyone agrees the output came from a machine and asks whether that output is reliable enough to admit. The deepfake fight assumes the opposite: one side claims the item is authentic human-captured media, the other claims it was fabricated, and the threshold question is provenance, not reliability. Those are different problems, and the draft rule answers only the first. Reading Rule 707 as a deepfake solution overstates what it does.

The rule also leaves the traditional expert pathway untouched. When a qualified expert witness relies on an AI tool, testifies about its output, and stands for cross-examination, Rule 702 already supplies the reliability gate, so Rule 707's trigger, machine output offered without an expert, never fires ^[1]. The practical effect is a fork in the road for trial teams: sponsor the machine output through a live expert and litigate reliability under Rule 702, or offer it unsponsored and expect to make an equivalent showing under Rule 707. Either way, the reliability record has to exist.

The fork in the road: sponsored testimony versus unsponsored output

Sponsor the output through an expert under Rule 702, or offer it unsponsored and carry the same reliability burden under Rule 707. The one option the draft removes is offering machine output with no reliability showing at all.

Once you accept that Rule 707 only fires when no expert sponsors the machine output, a strategic choice comes into focus early in every matter that leans on AI. A trial team can put a qualified expert on the stand to explain the tool, own its output, and absorb cross-examination, in which case the reliability inquiry runs through Rule 702 in the ordinary way. Or the team can offer the output on its own, in which case it should expect to make the equivalent showing under Rule 707. The rule does not force one path; it removes the option of taking neither.

Each path carries different costs, and choosing between them is a litigation-strategy decision rather than a technical one. A live expert adds credibility and a human the jury can follow, but also expense, availability risk, and a witness who can be impeached. Offering output unsponsored avoids those frictions but places the full weight of the reliability showing on the documentary record. What the draft rule does is make that second path viable only for teams that can actually meet the standard on paper, which raises the value of building the record from the start.

Why a reliability record is hard to reconstruct after the fact

Here is the uncomfortable implication for litigation teams already using AI for document analysis, damages models, forensic screening, or research: a Rule 702-style showing is very hard to reconstruct after the fact. Sufficient data, reliable methods, reliable application, each prong depends on knowing exactly what was generated, when, by which tool and version, under which parameters, and from which inputs. If those facts were never captured at creation time, the foundation becomes a memory exercise conducted under deadline pressure. The teams best positioned for Rule 707 will be the ones whose AI workflows produce their own evidentiary record as a byproduct of normal use.

Consider what reconstruction actually looks like when nothing was logged. A model version may have been updated since the work was done. Parameters that shaped an output may live only in someone's short-term memory. The exact inputs may be scattered across a shared drive, a chat thread, and a laptop that has since been reimaged. None of that is bad faith; it is ordinary knowledge decay. But under a reliability standard, the difference between a contemporaneous record and a reconstructed narrative is the difference between a foundation a court can test and an assertion opposing counsel can pick apart. The cheapest time to capture the record is the moment the AI runs. Every other moment is more expensive and less credible.

The point is not that AI-assisted work is unreliable. It is that reliability has to be shown, and a showing is only as strong as the record captured when the work happened.

Building a verifiable AI evidence record before the rule lands

Given all of the above, the sensible move is to treat the reliability record as an operational habit rather than a courtroom afterthought. The four Rule 702 factors point directly at the fields a good record needs to hold. If a workflow captures what was generated, by which tool and version, under which parameters, and from which inputs, then three of the four reliability prongs are answered from the record rather than from memory, and the fourth, helpfulness, is argued at trial like any other.

That is the layer RankShield Legal is built for. It produces tamper-evident, post-quantum-signed records of AI activity, capturing what was generated, when, by which tool, and under which parameters, and it certifies that authorities cited in AI-assisted work actually exist, are quoted accurately, and remain good law. Those records give litigators a verifiable provenance trail that supports a Rule 702-style reliability showing rather than a stack of after-the-fact assertions. To be clear about the limits: RankShield documents and attests; it does not decide admissibility, and no tool can. Judges will make those calls, and the record you bring is what they will weigh.

1. Capture at creation timeLog what was generated, when, and by which tool and version at the moment the AI runs, not weeks later under deadline.
2. Record the inputs and parametersPreserve the data fed in and the settings used, so the sufficient-data and reliable-application prongs rest on a record rather than recollection.
3. Verify cited authoritiesConfirm that authorities cited in AI-assisted work exist, are quoted accurately, and remain good law before the work leaves the building.
4. Keep the record tamper-evidentStore the trail so its integrity can be shown later, which is what turns a log into something a court can weigh.

What litigation teams can do now regardless of the timeline

Because the earliest realistic effective date is December 1, 2027, and because that date is a floor rather than a promise, it is tempting to defer. That would be a mistake. The decisions being made in current matters, which tools to use, what to capture, whether to retain an expert, will outlast the uncertainty about the rule's arrival. A matter opened today can easily still be in discovery or heading to trial when Rule 707, in some form, takes effect. Preparing now is not speculative; it is ordinary risk management for work already underway.

The good news is that preparing for Rule 707 mostly means doing things that are already good practice. Knowing which tool produced a result, on what inputs, with what settings, and whether the cited authorities hold up, is useful in any matter, rule or no rule. A team that builds those habits gains a cleaner record, faster foundation-laying, and fewer unpleasant surprises when an opponent probes an AI-assisted work product, all of which are valuable long before any new rule binds a courtroom.

• Decide early, per matter, whether AI output will be sponsored by an expert or offered on its own
• Capture tool, version, inputs, and parameters at the time the AI runs, not retroactively
• Verify that authorities cited in AI-assisted work exist, are quoted accurately, and remain good law
• Keep the provenance record tamper-evident so its integrity can be demonstrated later
• Track the rule's status through the committee's own materials rather than secondhand summaries

Where RankShield Legal fits, and where it does not

RankShield documents and attests. It does not decide admissibility. Under Rule 707, as under Rule 702, that decision stays with the judge, and the record you bring is what gets weighed.

It is worth being direct about the boundary of what any tool, including this one, can do. RankShield Legal is a vendor product, not a law firm, and it offers no legal advice. What it does is produce tamper-evident, post-quantum-signed records of AI activity and certify that cited authorities exist, are quoted accurately, and remain good law. That is a documentation and attestation function. It gives a trial team a stronger, verifiable record to bring into an admissibility fight. It does not, and cannot, decide whether that record clears the bar.

The reason that boundary matters is that Rule 707, if adopted, hands the reliability decision to the trial judge, exactly as Rule 702 does today. No provenance trail changes who makes the call. The value of a good record is that it gives the judge something concrete to weigh and gives opposing counsel less room to argue that the foundation is missing. Framed honestly, the offer is narrow and useful: better records support better reliability showings, and the showings themselves still belong to the lawyers and the court.