AI transcription tools have gotten remarkably good at one thing: transcribing clean audio of a single speaker in a quiet room. In that specific scenario, the best tools genuinely perform well. The problem is that legal proceedings are almost never that scenario.

Depositions involve multiple speakers, overlapping speech, technical terminology, accents, poor room acoustics, and remote connections with compression artifacts. Add in the legal weight of every word on the page, and the gap between "pretty accurate" and "accurate enough to stand as an official record" becomes enormous.

What follows are five scenarios drawn from reported attorney experiences and publicly documented cases where AI transcription produced errors that had real consequences. Details have been generalized where necessary to protect attorney-client privilege, but the types of failures are documented and consistent across the industry.

23%
Average error rate for AI transcription in real legal audio conditions
5x
Higher error rate for speakers with non-standard accents vs. clean audio
$4,200+
Average cost to retake a deposition when the original transcript is unusable

The 5 Failures

01
Employment Litigation

A single pronoun changed who was accused of what

During a deposition in an employment discrimination case, a witness described a conversation between two supervisors. The AI tool misattributed one speaker's words to the other throughout a roughly four-minute exchange, the result of overlapping voices and a momentary audio dropout.

What the AI produced: "He told her that she should have thought about that before she filed."
What was actually said: "She told him that he should have thought about that before he filed."

The reversal flipped the accusation entirely. The plaintiff's attorney caught the error only after submitting the transcript to opposing counsel, requiring a written correction, a follow-up session to re-read the passage on the record, and an explanation to the court about the discrepancy.

Cost: Half a day of attorney time, a formal errata sheet, and an embarrassing explanation to the judge. The case settled, but opposing counsel cited the transcript confusion in negotiations.
02
Medical Malpractice

A drug name became a completely different drug

A medical expert witness in a malpractice case was asked to describe the prescribed dosage of a specific anticoagulant. The AI tool had no problem transcribing the conversation around the medication but consistently rendered the drug name phonetically, producing a word that referred to a different class of drug entirely.

What the AI produced: "The patient was administered warfarin at a dose of..."
What was actually said: "The patient was administered apixaban at a dose of..."

Warfarin and apixaban are both anticoagulants, but they have different mechanisms, different standard dosages, and different contraindications. The substitution changed the expert's entire opinion as it appeared on the page.

Cost: The plaintiff's team discovered the error during trial prep, three months after the deposition. Correcting the official transcript required a motion, a hearing, and a reconvened session with the expert present. Total delay: six weeks.
03
Personal Injury

An objection that never made it into the record

A defense attorney objected to a line of questioning as leading and argumentative. The objection was raised while the witness was finishing a sentence, resulting in roughly two seconds of overlapping speech. The AI produced a transcript in which the objection appeared as fragmented noise characters, then simply disappeared from the record.

What the AI produced: "[inaudible] ...so the loading dock was always kept clear."
What was actually said: "Objection, counsel is leading the witness. [Witness continuing] ...so the loading dock was always kept clear."

A transcript without the objection is a transcript that suggests the testimony was given without challenge. The absence of the formal objection affected the evidentiary record and raised questions about preservation for appeal.

Cost: The attorney had to produce the audio recording to demonstrate the error, reopen the deposition record, and file a motion to correct. The opposing party disputed the correction, adding two additional weeks to the discovery timeline.
04
Immigration Court Proceeding

A witness whose testimony was half gibberish on the page

A witness in a removal proceeding testified in accented English. The AI tool produced a transcript that was accurate for roughly 60% of the testimony and increasingly unreliable for any sentence where the witness's regional accent was pronounced or where vocabulary was specific to their country of origin.

What the AI produced (sample): "I was working the [unintelligible] and the border [unintelligible] came at [unintelligible]."
What was actually said: "I was working the harvest and the border patrol came at four in the morning."

The gaps were not evenly distributed across unimportant details. They clustered around the most factually significant portions of the witness's account, precisely because those sentences involved proper nouns, unfamiliar place names, and dialect-specific vocabulary the AI had not been trained to recognize.

Cost: The proceeding was continued to allow for a human interpreter and certified reporter to be present. The original AI-produced transcript was excluded from the record entirely. The witness's account was effectively delayed by four months.
05
Commercial Litigation

A remote deposition where key testimony simply dropped out

A deposition was taken over a video conferencing platform. The witness was in a different time zone and experienced intermittent connection issues throughout the session. An AI transcription tool was used to produce the official record. Audio compression artifacts and brief freezes caused the tool to drop portions of testimony, particularly during the most detailed part of the witness's account of a contract negotiation.

What the AI produced: "We agreed that the payment schedule would [audio dropout] and then the addendum [audio dropout] which both parties understood."
What was actually said: "We agreed that the payment schedule would begin sixty days after delivery, and then the addendum clarified that the warranty period ran concurrently, which both parties understood."

The missing language contained the specific terms that were central to the dispute. The case turned on exactly when payment obligations began and whether the warranty ran concurrently or consecutively. None of that survived the transcript.

Cost: The opposing party refused to stipulate to a correction. A second deposition was required. Scheduling the witness again took seven weeks and cost both sides roughly $12,000 in attorney time and travel expenses.

What Every One of These Has in Common

None of these failures involved obscure edge cases or unusual circumstances. Overlapping speech happens in nearly every deposition. Medical terminology is standard in malpractice cases. Witnesses with accents are the norm in a diverse country. Remote depositions have been routine since 2020. Connection issues are an accepted part of video proceedings.

These are not the failure modes of a technology that is almost ready. They are the failure modes of a technology that is fundamentally not designed for what legal transcription requires. AI transcription optimizes for intelligibility on average. Legal transcription requires accuracy on every single word, in every condition, with no exceptions.

"The problem is not that AI transcription makes mistakes. Every tool makes mistakes. The problem is that in a legal proceeding, a mistake in the record is not a mistake you can just go back and fix informally."

The Hidden Cost Most Attorneys Do Not Calculate

When attorneys evaluate AI transcription tools, they tend to compare the cost of the tool against the cost of a certified reporter. That comparison usually makes AI look attractive. What the comparison misses is the cost of a single significant error.

Correcting a material error in a deposition transcript can involve any combination of the following:

The cost of hiring a certified court reporter for a full-day deposition is typically between $400 and $1,200 depending on location, complexity, and turnaround time. The cost of a deposition retake starts at several thousand dollars before you count attorney time. The math is not difficult.

Failure Type AI Transcription Risk Certified Court Reporter
Speaker attribution errors Common with overlapping voices Resolved in real time
Technical terminology Phonetic guesses, often wrong Researched before the proceeding
Objections and interruptions Frequently dropped or garbled Captured and attributed accurately
Non-standard accents Error rates 5x higher than standard Human judgment, real-time clarification
Remote audio quality issues Dropouts render testimony unusable Can pause, request repeat, flag issues
Legal admissibility of transcript Disputed without certification Certified, court-admissible record

A Note on Where AI Transcription Is Actually Useful

It is worth being clear: this is not an argument that AI transcription has no place in legal practice. For internal purposes, first-pass review of audio, and non-evidentiary note-taking, automated tools can save real time. Many firms use them exactly this way.

The problem arises when firms use AI transcription as a substitute for a certified reporter in any proceeding where the transcript will carry legal weight. That is the specific use case where the failure modes above apply, and it is the use case that carries real risk.

The next article in this series examines the full side-by-side comparison across every dimension that matters: speed, cost, accuracy, real-time capability, and legal admissibility. Read it here: AI vs. Court Reporters: A Full Comparison.

Court Reporters vs AI — A Six-Part Series

Need a Certified Court Reporter?

Search our directory of verified professionals. Available for depositions, hearings, trials, and remote proceedings nationwide.

Find a Court Reporter