Are AI Transcripts Admissible in Court?

Vendors tend to answer "is it accurate?" when asked "is it admissible?" — they are not the same question, and the distinction has practical consequences.

Accuracy is empirical: does the text match what was said? Certification is procedural: has a person authorised by the jurisdiction attested that it does, under a rule that specifies who may attest and how? Admissibility is evidentiary: will this court accept this document for this purpose?

An AI transcript can satisfy the first and fail the second outright, because certification is a question of who signed it, not how good it is. A transcript with a 99.9% accuracy rate and no certification is not an official transcript in a jurisdiction that requires one.

This varies more than most practitioners outside their home jurisdiction expect, and the variation is not a formality.

California requires transcripts used as the official record to be certified by a licensed Certified Shorthand Reporter, who must list their CSR number on both the cover page and the certificate. Since 1 January 2023 delivery must be electronic unless a party entitled to the transcript requests paper. Formatting is prescribed down to line numbering and spacing.

Pennsylvania is distinctive in requiring each contributor to certify their own portion of a transcript, rather than one signature covering the whole document. North Carolina publishes an approved transcriptionist list through the state court system, so the question of who may prepare a transcript has a concrete published answer.

Florida is the outlier worth knowing about: it has no statewide court reporter licence at all. Certification requirements are set circuit by circuit under Rule 2.535, which means the rule that applies depends on which circuit your matter sits in. A vendor claiming to provide "Florida certified transcripts" is describing something that does not exist as a statewide category.

State certification rules do not govern federal proceedings, and a practitioner who has confirmed the rule for their state court has answered only half the question if the matter is federal.

Under 28 U.S.C. § 753, each federal district court appoints official court reporters, and the reporter certifies the transcript of proceedings. Individual districts publish their own transcript policies, including format requirements and redaction procedures for personal identifiers. Two courts sitting in the same city — a state trial court and a federal district court — can impose materially different requirements on the same transcript.

The practical consequence is that "certified in this state" is not a portable answer. Confirm the rule for the specific court and the specific purpose, because the cost of discovering a transcript is unusable is paid at the point you need it.

There is a second category that the certification discussion tends to obscure entirely: a transcript of a recording offered as evidence, rather than a transcript of a court proceeding.

Where a party seeks to introduce a recorded phone call, a body-camera recording, or a covert recording, the transcript functions as an aid to the finder of fact rather than as the official record of a hearing. The governing questions become authentication of the underlying recording and whether the transcript accurately reflects it — not whether a licensed reporter produced it. Courts commonly handle disputes by having the parties submit competing transcripts or stipulate to an agreed one.

This is where a carefully reviewed transcript does substantive work, because accuracy is the actual issue in dispute and no certification requirement stands in the way. It is also where an unreviewed automated transcript is most dangerous: opposing counsel will read it as closely as you did, and a mistranscribed passage becomes a credibility problem rather than a typo.

Framing this as a limitation understates how much litigation work involves transcripts that were never intended for filing.

Speaker attribution is the failure mode specific to legal transcription, and it is worse than a garbled word. A transcript that attributes testimony to the wrong speaker is not a degraded transcript; it is an actively misleading document. Automated transcription is at its weakest precisely where depositions are most contested — overlapping speech, rapid exchanges, interruptions.

The same applies to names, figures, dates, and citations. These appear once, are never repeated, and a model faced with an unfamiliar proper noun produces a plausible substitute rather than flagging doubt. A human reviewer working against the audio is what converts a probably-right transcript into a checked one.

So the honest position is narrower than "AI transcripts are fine" and broader than "AI transcripts are useless": they are reliable working documents when reviewed, and they are not official records unless certified by someone the jurisdiction authorises.