Federal Rule of Evidence 707 and the Future of Machine-Generated Evidence in U.S. Courts

Written By Brian Walters

Introduction: A measured rule for a fast-moving problem

Courts are already seeing parties offer outputs from algorithms—facial recognition matches, similarity scores, automated classifications, and even deepfake videos—as if they “speak for themselves.” The proposed Federal Rule of Evidence 707 (Machine-Generated Evidence), released for public comment in May 2025, is an attempt to keep evidentiary doctrine aligned with reality. In short: when parties seek to admit machine-generated evidence without a human expert, they would need to satisfy the same reliability requirements that govern experts under Rule 702. The Advisory Committee on Evidence Rules voted 8–1 to publish a draft—an exploratory step that explicitly invites robust public comment.

Below, I synthesize the most relevant commentary (May–October 2025), explain how Rule 707 fits with Rules 702 and 902(13), unpack what counts as “machine-generated evidence,” and translate the proposal into practical action items for litigators, judges, and technologists.

What the Committee is proposing—and why now

The core move

The draft would require courts to apply Rule 702’s reliability gatekeeping to machine-generated outputs when they are offered without expert testimony, while exempting “basic scientific instruments.” The logic is straightforward: if the evidence would need a Rule 702 foundation when presented by a human expert, it should not bypass that scrutiny just because a machine produced the output.

Judge Jesse M. Furman’s framing

U.S. District Judge Jesse M. Furman, who chairs the Advisory Committee, has emphasized that publishing the draft is not a foregone conclusion but a way to get ahead of a rapidly changing landscape and learn from the public record. He has repeatedly framed the project as exploratory and iterative. Reuters+2Bloomberg Law News+2

Why Rule 702 alone may not be enough

The Committee’s materials reflect a recurring problem: many litigants try to offer automated outputs directly—or through witnesses who cannot meaningfully explain the model—leaving courts without a clear doctrinal hook for reliability screening. The 2023 amendments to Rule 702 sharpened courts’ focus on sufficient facts/data, reliable principles and methods, and reliable application, but they presuppose a testifying expert. Rule 707 addresses the gap when no expert is offered. United States Courts

The lone “no” vote: DOJ’s opposition and concerns

The Department of Justice cast the sole vote against publishing Rule 707 for comment. Reporting identifies Elizabeth Shapiro as the dissenter. The DOJ’s core position: Rule 702 already does the job when experts rely on algorithmic tools, and a new rule risks unnecessary complexity, cost, and potential overreach—particularly around disclosure of proprietary systems. In their view, the better answer is to use existing doctrine rather than create a new, technology-specific rule. Reuters+2Bloomberg Law News+2

Practical takeaway: whether or not Rule 707 is adopted, parties relying on machine outputs should plan as if a full Rule 702 showing will be required, including validation data, error rates, and a reliable application to the facts.

What counts as “machine-generated evidence”?

The draft and its commentary contemplate several common categories (illustrations—not a closed list): United States Courts

  • Predictive or inferential outputs from AI/ML systems (e.g., risk scores, market-impact inferences, damages estimates).

  • Forensic image/video analyses, such as facial recognition, pattern matching, or enhancement pipelines.

  • Synthetic media (deepfakes)—images, audio, or video purporting to depict reality but algorithmically fabricated.

  • Automated software analyses that classify, rank, or infer facts without expert oversight.

By contrast, “basic scientific instruments”—thermometers, standard lab devices—remain outside 707’s scope. (The Committee judged them less likely to present the opacity, dataset, and validation risks of modern ML systems.) United States Courts

Several commentators argue 707’s trigger—limited to uses “without an expert”—is too narrow, because the same reliability risks exist even when a human witness shows up to “sponsor” a black box. Expect that point to feature prominently in public comments. druganddevicelawblog.com

How Rule 707 fits with Rules 702 and 902(13)

  • Rule 702 (as amended in 2023) already requires courts to find, by a preponderance, that an expert’s opinion rests on sufficient data, employs reliable principles and methods, and reliably applies those methods. Rule 707 ports that reliability test to machine-generated evidence offered without an expert. United States Courts

  • Rule 902(13) provides self-authentication for certain electronic data via certification. That satisfies authenticity, not reliability. Even self-authenticated machine outputs may face Rule 707 (reliability) scrutiny at admission. United States Courts

  • The Committee also considered a new 901(c) concept for deepfake challenges: when an opponent makes a plausible showing of AI fabrication, the proponent may have to prove authenticity by a higher burden than the usual prima facie standard. That deepfake-specific measure remains under discussion but underscores the Committee’s dual-track approach: authenticity and reliability are different gates. National Law Review+1

Practical takeaway: Do not conflate authentication (Rules 901/902) with reliability (Rule 702 / proposed 707). A document can be what it purports to be (authentic) and still be unreliable.

The commentary landscape (May–October 2025)

Courts and rulemakers

  • Advisory Committee agenda book (May 2025): Includes the draft text and Reporter’s memorandum; it describes the rationale, scope, and instrument exemption. United States Courts

  • Standing Committee (June 2025): Took up the question of publishing for public comment; reporting confirms approval to seek public feedback. United States Courts+1

  • Judge Furman’s public framing: exploratory posture; need to “get ahead” of rapid technological change; explicit request for robust, critical comment. Reuters+1

Bar, firm, and practitioner analysis

  • Womble Bond Dickinson: Useful overview of scope, the 8–1 vote, and the “no-expert” trigger; flags cost and implementation concerns. Womble Bond Dickinson

  • National Law Review (multiple pieces): Highlights the Rule 707/Rule 702 alignment, the potential two-step deepfake framework under 901(c), and practical uncertainties about gatekeeping. National Law Review+2National Law Review+2

  • Drug & Device Law Blog: Critiques the “without an expert” limitation as under-inclusive; urges a broader rule for algorithmic outputs. druganddevicelawblog.com

  • Villanova Law Review (blog): Explains mechanics and concerns about definitional breadth; good for academic framing. Villanova Law Review

Practical takeaway: The consensus welcomes a uniform federal standard but stresses implementation details—what must be disclosed, how courts will assess reliability, and how to protect trade secrets while enabling adversarial testing.

State-level movement: not Rule 707 clones, but signals are clear

No state has adopted a direct analog to FRE 707. But two developments matter:

  1. Deepfake criminal/civil laws:

    • New Jersey enacted a deepfake law (April 2025) creating criminal and civil penalties for deceptive AI-generated media; the Governor’s release and independent reporting confirm penalties and intent. NJ.gov+1

  2. Judicial administration rules for AI:

    • California Judicial Council adopted Rule of Court 10.430 (July 2025), requiring every state court to either ban generative AI or adopt a policy by set deadlines; the rule addresses confidentiality, verification, and disclosure. Reuters+1

    • New York Unified Court System issued an interim AI policy (October 2025) restricting use to approved tools, prohibiting feeding confidential material into public systems, and mandating training and disclosure/verification practices. Reuters+1

Practical takeaway: Even without Rule 707-style evidence codes, states are hardening the institutional perimeter—deepfake bans and court-system AI policies are becoming common. Expect state evidence committees to revisit authentication and reliability for deepfake evidence and AI in litigation as federal deliberations advance.

The litigation challenges Rule 707 is trying to solve

  1. Reliability & bias: Models may be opaque (no access to weights/source), trained on biased data, or susceptible to drift and adversarial manipulation. 707 would require a showing akin to Daubert’s valid principles/methods and reliable application, backed by validation, error rates, and—where feasible—peer review or benchmarking. United States Courts

  2. Reproducibility & testability: Opponents need a defensible way to probe the model (e.g., hold-out tests, audit logs, version history). The Committee materials and commentary anticipate protective orders and supervised testing to balance discovery with IP protection. United States Courts

  3. Authenticity vs. reliability (deepfakes): Traditional Rule 901 was built for real items; synthetic items can evade the usual provenance proofs. The contemplated 901(c) approach would front-load the authenticity fight where there’s a plausible deepfake claim, separate from the reliability fight. National Law Review+1

  4. Hearsay fit: Fully automated outputs often do not involve a human declarant; they may be non-hearsay—but still must meet reliability to be admitted as substantive evidence. Rule 707 supplies the gatekeeping even when Rule 801 does not. United States Courts

What litigators should do now (checklist)

Treat every machine-generated output like expert material under Rule 702. Even if you intend to proceed “without an expert,” build the record as if a Daubert hearing will happen.

  1. Model provenance file

    • Training data: sources, representativeness, data hygiene, known limitations.

    • Validation: cross-validation method, benchmarks, error rates, sensitivity/specificity.

    • Versioning: model iterations, change logs, and when the case-relevant run occurred.

    • Controls: guardrails against drift, adversarial perturbations, and overfitting. United States Courts

  2. Explainability & testability

    • Provide confidence scores, feature attributions, or interpretable summaries where available.

    • Be prepared to re-run the model on agreed datasets; propose neutral-expert or court-supervised testing under a protective order to mitigate trade-secret exposure. United States Courts

  3. Disclosure strategy

    • Decide early whether to wrap a qualified expert around the machine to reduce 707 risk.

    • If proceeding without an expert, ensure you can still make a preponderance showing on each 702 factor with documentation and third-party validations. United States Courts

  4. Authentication plan (deepfakes)

    • For audio/video, plan chain-of-custody, device metadata, hashing, and forensic reports.

    • If you anticipate a deepfake challenge, prepare to clear a heightened 901(c)-style showing if that amendment (or a local analog) is adopted. National Law Review+1

  5. Protective orders & confidentiality

    • Negotiate tiered access to code/weights/data or use trusted-neutral review to allow adversarial testing without IP surrender. Courts have long experience crafting such orders; expect more of them. United States Courts

  6. Rule 403 foresight

    • Even if you clear 707/702, anticipate Rule 403 arguments (unfair prejudice, jury confusion). Keep use cases modest and validated, and consider limiting instructions or demonstrative-only use where appropriate. National Law Review

Guidance for judges (gatekeeping without technophobia)

  • Structure the record: Ask for validation studies, error rates, and how the model generalizes to the case facts. Require an audit trail (inputs → outputs). United States Courts

  • Phase reliability: Consider pretrial 104(a) hearings or appointing a neutral expert in complex cases.

  • Separate authenticity from reliability: If a plausible deepfake challenge is made, take up authentication first; only then analyze 707/702 reliability. National Law Review

  • Balance transparency and secrecy: Encourage protective orders or special master processes to enable meaningful adversarial testing while protecting proprietary materials. United States Courts

  • Use Rule 403 judiciously**:** When an output risks overawing the jury (e.g., glossy but weakly validated heatmaps), weigh probative value vs. confusion/prejudice. National Law Review

How technologists can future-proof tools for “AI in litigation”

Vendors and in-house data teams who expect their tools to surface in court should:

  • Build auditability in from the start: logging, reproducibility, documented data lineage, version control, and error tracking.

  • Include explainability features (feature importance, counterfactuals, confidence intervals).

  • Prepare a “litigation disclosure mode”: a redacted, court-friendly way to let opponents (or a neutral) re-run tests without compromising IP.

  • Maintain a living validation dossier that maps cleanly to Rule 702 factors.

These design choices reduce litigation friction and can make the difference between admissible machine-generated evidence and a successful exclusion under Federal Rule of Evidence 707. United States Courts

Where states are heading: deepfakes and court-system AI policies

As noted, we are not yet seeing state evidence rules mirror 707, but there is rapid movement on deepfake evidence and AI governance in courts:

  • New Jersey: criminal/civil penalties for deceptive deepfake evidence (April 2025). Useful signal for authenticity and source-validation expectations. NJ.gov+1

  • California: statewide court-system AI rule requiring adoption of policies (or bans) by set deadlines; disclosure and verification of AI-generated work feature prominently. Reuters+1

  • New York: interim AI policy for courts (October 2025)—approved tools only, training, confidentiality protections. Reuters+1

Practical takeaway: expect state evidence committees and judicial councils to continue experimenting at the edges—especially with deepfake evidence—even as federal Rule 707 proceeds through the comment process.

What happens next: timeline and expectations

  • The Standing Committee approved soliciting public comment (June 2025). Expect a public comment window extending into early 2026, followed by revisions. If the process continues, the Supreme Court would transmit any final rule to Congress by May preceding the effective date, making December 1 of a future year a typical target (commentators are floating 2027 as plausible). Reuters

How to prepare now

  • Litigators: Build 707-ready records. Even if you plan to use an expert, assemble validation and explainability materials so the opinion is resilient under Rule 702 and any Rule 707 overlay.

  • In-house & vendors: Create standard validation packets and neutral testing protocols aligned with AI in litigation best practices.

  • Judges & court admins: Develop bench cards, identify trusted neutrals, and consider local standing orders for machine-generated evidence and deepfake evidence authentication.

Final analysis: a prudent, workable path—if we do the work

The proposed Federal Rule of Evidence 707 is not a tech panic; it is a modest doctrinal nudge that restores parity between human expert testimony and machine-generated evidence. It will not cure every problem (e.g., it does not itself compel source-code disclosure), and the DOJ’s concerns about cost and redundancy warrant serious attention. But the alternative—allowing algorithmic outputs to bypass reliability review simply because no expert is offered—invites inconsistency and, ultimately, error. Reuters

For now, the smartest course is to internalize 707’s logic today: treat algorithmic outputs as expert-grade evidence that must be validated, explainable, and testable. That approach serves litigants, courts, and—most importantly—the truth-seeking function of trials in an era when AI in litigation and deepfake evidence are no longer hypotheticals but routine realities. United States Courts

Sources & further reading

Brian Walters
Previous

The $125,000 Business Personal Property Tax Exemption: What Texas Small Businesses Need to Do in 2026
Next

TRAIGA Deep Dive: Texas’s “Prohibited Practices” AI Law, the Biometric Trapdoor, and the Enforcement Math