What makes a legal transcript 'official'?
In federal court, only the transcript a reporter certifies is the official record. 28 U.S.C. § 753(b) makes no transcript official unless it's made from records certified by the reporter, and a certified transcript is deemed prima facie a correct statement of the testimony. An AI draft, however clean, isn't that record.
The format isn't optional either. The Judicial Conference has prescribed the federal transcript format since 1944 so every party is treated equally nationwide, and no court, judge, or transcriber may deviate from it. PART 18.17 requires the reporter or transcriber to authenticate the transcript with a certification on the last page – "I (we) certify that the foregoing is a true and correct transcript" – which is exactly what a filing relies on.
So separate two jobs. One is the certified record for the court; the other is every working copy around it – the draft you search, annotate, and quote in a memo. AI is for the second job. Multi-speaker transcription of a deposition or hearing gets you a fast, speaker-labeled draft to work from, not a substitute for the certified transcript.
Who can certify legal transcription?
Certification comes from a credentialed reporter or transcriber, not from the software. The National Court Reporters Association issues the stenographic RPR (Registered Professional Reporter), its foundational certification, alongside the realtime CRR. For digital audio, the American Association of Electronic Reporters and Transcribers issues the CER and CET: the electronic reporter who captures the proceeding, and the transcriber who produces the transcript from that recording.
These credentials still set a hard accuracy floor. NCRA's RPR skills tests require 95% accuracy on each leg to pass. AAERT's CET practical exam requires 95% accuracy for exams on or after March 26, 2026, lowered from 98%, though AAERT says 95% still matches the highest base-level requirement in the industry. Relaxed or not, that's the bar a certified human clears – and the reason an unedited machine draft doesn't file itself.
Where does AI legal transcription fit?
Use AI for volume and speed, not for the certified record. Manual verbatim transcription runs up to six hours of work per hour of audio (Bell et al. 2018, via Fleiss et al. 2024) – unworkable across a matter's worth of recordings. An AI first pass turns each hour into minutes of processing plus targeted cleanup. Because legal audio arrives in bursts, a deposition here, hours of discovery there, a pay-once model where credits never expire fits matter-based volume better than a monthly seat.
The safe uses are internal: reviewing discovery audio, drafting a deposition summary, pulling quotes for a brief, or indexing hours of recordings to find the one exchange that matters. A clean recording still decides most of your accuracy, so the interview-recording fundamentals apply here too. In each case, a human reads the draft against the audio before anything load-bearing leaves the building.
Where AI needs you most is exactly where legal transcripts are unforgiving: proper nouns, case citations, numbers said quickly, overlapping speakers, and the gap between "can" and "can't." Bracket anything unclear as [inaudible] with its timestamp rather than guessing – a flagged gap is defensible, a confident mistake isn't. Once you've checked it, export a clean copy to DOCX for the matter file.
What about consent and confidentiality?
Recording law and the duty of confidentiality govern this before a word is transcribed. Federal law requires one-party consent, while about 11 states require all-party consent – California, Florida, Illinois, and Washington among them – so where a recording crosses state lines, get everyone's yes. We can't give legal advice; check the rule in your jurisdiction before you record.
Once you hold the audio, the Illinois Rule of Professional Conduct 1.6 frame applies (a verbatim adoption of ABA Model Rule 1.6): a lawyer must not reveal information relating to a client's representation without informed consent, and Rule 1.6(e) – the ABA's Rule 1.6(c) – requires reasonable efforts to prevent inadvertent or unauthorized disclosure of, or access to, that information. That duty reaches every recording and transcript in the matter, wherever they sit.
Practically, that means minding where the audio and drafts live. Use a tool that doesn't train AI on your files and lets you delete them after processing. Pepys never trains on your audio or transcripts, and you can auto-delete files once they're transcribed – which matters when the recording is privileged or sensitive.
