Proposed FRE 707 would extend Daubert's reliability standard to any AI-generated evidence offered without a sponsoring human expert. Here is how plaintiff attorneys using AI tools in depositions should prepare — including a methodology disclosure exhibit you can attach to a motion in limine.
See how AI medical-record review links every fact to the exact Bates page that proves it — click any citation and jump straight to the record.
See the 60-second demo →Why this matters now. The Advisory Committee on Evidence Rules published a draft of new Federal Rule of Evidence 707 for public comment in August 2025. Comment closed February 16, 2026. If the rule is approved by the Judicial Conference, the Supreme Court, and Congress on the standard schedule, it takes effect December 1, 2026. Attorneys who use AI tools in depositions and at trial should be ready to defend admissibility today, not after the first adverse ruling.
Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993), supplied four non-exhaustive factors for evaluating the reliability of scientific testimony: testing, peer review, known or potential error rate, and general acceptance. Proposed FRE 707 takes these factors and applies them to evidence produced by an AI system that is offered without a sponsoring human expert who has personally adopted the conclusions. The honest analysis follows.
Has the methodology been tested? For Courtroom AI's contradiction-flagging pipeline, the answer is partial-yes: we run an internal regression suite of approximately 858 fixture-based test cases across all 22 production tools, the contradiction-detection module included. The tests cover transcript ingestion, segmentation, embedding similarity above a tunable threshold, and end-to-end alert generation against canned medical-record corpora. We have not published these test results in a peer-reviewed forum and we do not claim a published gold-standard accuracy figure. Sophisticated opposing counsel will press on this; the honest response is that the underlying language model and embedding model have both been benchmarked publicly by their providers (Anthropic and Voyage AI respectively), and the Tool's role is to surface candidate contradictions to the attorney rather than to issue a verdict.
The components of the system have varying degrees of peer review. Voyage AI's embedding models have been benchmarked publicly on the Massive Text Embedding Benchmark (MTEB) and on legal-domain BEIR-LegalBench tasks; the published technical reports describe retrieval accuracy at standard k values. Anthropic publishes model cards, system-card safety evaluations, and accuracy benchmarks on tasks such as MMLU, HumanEval, and various reasoning benchmarks. The CourtListener and PubMed APIs serve content that has itself been peer-reviewed in the original sources. The orchestration of these components into a contradiction-detection system is, however, MedLegal AI's own composition and has not yet been subjected to independent peer review. We are upfront about this gap.
This is the factor where attorneys must be most candid with the court. We do not have a peer-reviewed published error rate for end-to-end Courtroom AI contradiction detection. What we can responsibly cite, broken into the system's components:
voyage-law-2 model card reports retrieval recall@10 in the high 70s on legal benchmarks; voyage-3 reports comparable general-domain numbers. These are recall, not precision; an attorney evaluating an alert sees the cited source.The honest framing for a court: the Tool's component error rates are publicly disclosed by their providers; the composed system's error rate has not been peer-reviewed; and the architecture is therefore designed so that the attorney verifies every cited source before acting on any alert. Reliability of the assistive output, in other words, is bounded above by the reliability of the attorney who reviews it.
General acceptance of AI assistive tools in litigation practice is genuinely early-stage. Westlaw and Lexis offer AI-augmented research products. CoCounsel (Casetext, now Thomson Reuters) and Harvey are widely deployed at AmLaw firms. EvenUp, Supio, and Filevine offer AI-augmented case-development tools. ABA Formal Opinion 512 (July 2024) addresses lawyers' duties when using generative AI. State bar opinions in California, Florida, New York, New Jersey, and the District of Columbia have addressed AI use in the practice of law in the past 18 months. Live deposition AI is a smaller niche; general acceptance for that specific use case is best described as emerging rather than settled. Counsel should expect that opposing counsel will challenge novelty, and should be prepared with the methodology disclosure below.
This section summarizes the formal Methodology Disclosure document that MedLegal AI customers may attach as an exhibit to a motion in limine response, a Daubert opposition, or an expert disclosure. The full document runs roughly 1,000 words at filing length.
| Component | Provider |
|---|---|
| Speech-to-text | Commercial ASR (configurable per deployment) |
| Reasoning / contradiction detection | Anthropic Claude (Sonnet and Haiku tiers) |
| Semantic embeddings | Voyage AI voyage-3 and voyage-law-2 |
| Medical literature | NCBI PubMed E-utilities |
| Case law | CourtListener (Free Law Project) |
| Physician payments | CMS Open Payments |
| FDA adverse events | openFDA |
Every alert is rendered as a suggestion to the licensed attorney user. The Tool has no autonomous authority to ask questions on the record, address the witness, or appear in any capacity at the deposition or trial. The attorney remains the sole decision-maker. This is the same architecture as an attorney consulting Westlaw or a deposition binder during examination — tools whose use has never required Daubert qualification because the tool is not the witness.
Where the Tool returns an alert, it returns a confidence band drawn from three sources: embedding cosine-similarity between the questioned statement and the candidate source passage, the language model's self-reported confidence, and the freshness and authoritativeness of the cited source. Confidence bands are presented as High / Medium / Low, not as numerical probabilities, because the underlying model-reported confidence has not been independently calibrated against ground truth in published peer-reviewed studies.
The disclosure document explicitly states: ASR word-error rate is not zero; embedding retrieval is not exhaustive; large language models can produce plausible but incorrect statements; the Tool does not validate factual truth (it identifies divergence between statements and sources); the composed contradiction-detection accuracy is not peer-reviewed; and the Tool relies on the integrity of its inputs. Every cited source is presented to the attorney with a deep-link URL or document anchor so the attorney can verify the source before relying on it.
For each session the Tool stores artifacts that allow the chain from microphone to alert to be reconstructed:
The Tool's transcript is not the official record. The certified court reporter's transcript remains the official record under FRCP 30. The Tool's transcript is contemporaneous attorney work-product used as a research aid and is so labeled in any production. Where opposing counsel demands the Tool's transcript or audio in discovery, counsel should evaluate work-product, attorney-client, and trade-secret protections before producing.
Proposed FRE 707 targets a specific abuse: offering AI output as expert testimony with no qualified human standing behind it. That is a real risk, and the rule is sensible. It is also not how Courtroom AI is used.
Courtroom AI surfaces; the attorney and any retained expert testify. We never offer the Tool's output as the expert opinion in a case. The closest analogy is the one judges already understand: nobody objects when an attorney uses Westlaw to find a case and then cites that case at the podium. The case, not Westlaw, is what is offered. Westlaw is the research aid. The same logic applies here. The medical record, the prior deposition, the published article, the public payment record — those are what is offered. The Tool is the research aid that helped the attorney find them in the few seconds available between question and answer.
This distinction matters in three concrete ways:
If a competing AI product does offer its tool output as substantive evidence — for example, an "AI damages calculator" report tendered as the damages opinion without a sponsoring economist — that product faces a real FRE 707 problem. Courtroom AI does not.
Federal opinions touching AI in litigation in 2025 and 2026 are accumulating. Several patterns are visible: courts have sanctioned attorneys who filed AI-fabricated case citations (a Mata v. Avianca lineage that started in 2023 and continues); courts have addressed admissibility of AI-generated images and deepfake authenticity questions; and courts have begun requiring disclosure of AI use in filings under standing orders. None of these opinions, as of this writing, has finally resolved the question of admissibility for an AI tool used as a research aid by a licensed attorney during deposition or trial preparation. We do not pretend otherwise.
On the bar-regulation side: ABA Formal Opinion 512 (July 2024) addresses competence, confidentiality, supervision, and reasonable fees in lawyers' use of generative AI tools. State bar formal opinions from California, Florida, New York, New Jersey, the District of Columbia, and others have followed in the past 18 months and broadly track Opinion 512. None of these bar opinions prohibits use of AI as a research aid; they require that the attorney remain competent and accountable, which is precisely the architecture described in section 2 above.
Counsel should monitor the docket of the Advisory Committee on Evidence Rules for the final FRE 707 text and the implementing committee notes, which may answer specific questions about the rule's reach.
If opposing counsel files a motion in limine or Daubert challenge to your use of Courtroom AI, the playbook is straightforward.
The Tool is bounded in seven specific ways, and these bounds are by design:
If you are facing or expecting a Daubert challenge in the near term, our 5-second Daubert workup primer walks through the public-records lookup that Courtroom AI automates. Our April 2026 comparison matrix places the Tool against EvenUp, Supio, Filevine, Harvey, and CoCounsel on 13 features, including how each handles methodology disclosure.
See Courtroom AIvoyage-3 and voyage-law-2; available at docs.voyageai.com.This white paper is informational and is not legal advice. Counsel relying on it should evaluate the specific evidentiary standards of their jurisdiction, including state-court variations such as Frye in New York and certain California state-court contexts, and any local standing orders requiring AI disclosure. Consult outside counsel before filing.