If you labor really even close to legal, then you’ve probably muted this talk at one aspect during a very last 12 months. An office worker proposes recording deposition audio through an AI transcription app, because it’s cheap and nearly instantaneous. Someone else pushes back. And no one really knows where the line is meant to lie.
With that in mind, here is a candid analysis of exactly where the line falls – in particular with what AI transcription excels, how and why it quietly breaks down often times – and chuckle all you like – the legal field has had its audio surrender to software last.
Transcript =/= notes
Legal transcripts, however, live under a separate set of rules. At trial, you can read back a deposition transcript to the witness: used for impeachment. There is an exhibit entered containing a transcription of the 911 call. A transcript of an arbitration can becomes a basis for appeal. These are not documents that the one audience who stands to gain from an error in them, opposing counsel, will be loath to parse.
We actually think the real question is not, ‘is AI transcription good? You train it end of October 2023, and the question is always: «is this good enough to last an attack? And that is a far, much higher reality.
The place solely AI transcription fails with regards to legal audio
That accuracy number, which AI tools love to brag about – in the 85–95% range somewhere (perhaps even higher!) – is based on clean audio. A speaker or two, decent mics – no one interrupting anyone. Legal audio is pretty much the ideal stress test for everything automatic speech recognition (ASR) struggles to cope with.
Take crosstalk. Depositions, by their nature of being adversarial. Over on the attorneys politely interrupt their answers, over on witnesses halt answering questions while other people stop in mid-sentence to cut them off. AI will either stitch those overlapping voices into a single, mushy speaker or simply omit the overlap altogether – but “Objection, form” has to continue appearing in the record even if it lands right under someone’s sentence.
Next are negations, which play their own low-level havoc. Barely a thing, barely aloud- One little unstressed syllable divide “I did sign it” and “I didn’t sign it.” AI training, built from a sea of data up to October 2023 just plucks the more likely word – and never tells you it guessed when audio gets difficult. That one guess, that is silent can change what the answer means when you write it on a testimony.
Combined with that is just simply bad audio. Big, empty speakerphones, echoey courtrooms with the odd emotional witness breaking down at just the wrong time: one droopy intonation ruining a questioning; heavy accents. You are trained on the parts that we, as humans working from context – what was asked or said by whom a minute ago – write in and mark anything spotty with an (inaudible). AI itself doesn’t really have an (inaudible) setting. It always comes up with something whether it heard correctly or not.
AI transcription tools are practically designed to do the opposite of what legal work requires. Deliberately, they’re trained to smooth speech into open text – trimming hesitations, collapsing repetitions and tidying up grammar. Okay, I mean meeting notes sure that’s a feature. For testimony – it’s an automated, unseen redaction of the record. Nothing gets logged anywhere. You would never ever even look at what you washed away.
And then there is the question of privilege
Someone honestly should answer some questions before a firm uploads an attorney-client recording to the free transcription app. And where does that audio actually go? Sitting on whose servers? Is it retained afterward? Are you going to train the model using it?
As with most consumer AI tools, those answers exist somewhere in a terms-of-service agreement that no one has ever read, and they aren’t questions likely to be answered by any firm bound as an attorney relationship privileged. Legal transcription differs from this model – non-disclosure agreements, limited access to files, secure transport and storage of your data as well as a real named individual bearing the responsibility for that recording. There is a true answer to the question – whom had access to your audio if ever a client asks.
“But AI is so much cheaper”
Upfront, yes, dramatically so. To the auto-transcriber, a two-hour deposition costs a few bucks and appears in minutes.
Then the real cost starts. Someone needs to listen back and correct speaker labels, reinsert any skipped objections or testimony; format the transcript for a trial preparation file. To properly correct output of AI you must listen to everything, since the mistakes sound like completely ordinary sentences. A few hours of cleanup at paralegal billing rates cancels out the savings several times over – and anything that escapes review just sits in the file, waiting for opposing counsel to find it first.
On the other hand, it costs more on the invoice but definitely less in real life – a transcript can come done and dusted, formatted up and ready to reference.
This is where AI really earns its crust
In truth, AI transcription is not without merit in a law office. Far from it, really. Excellent for transforming your own dictation into a searchable rough draft, skimming through two hours of audio to find the ten minutes worth saving, or just internal notes that will never ever leave your desk.
The bottom line
For common audio, AI has made transcription faster and cheaper than ever before (which I consider a legitimate win – no arguments here). However, legal transcripts work a different kind of magic. They did not need to be just mostly correct – they needed to have defensibility in every word where it mattered, confidentiality by necessity and presentation formal enough that the result would stand as a record. That’s still human work. It’s what we’ve done since 1999 at American Transcription Services -experienced, U.S.-based legal transcriptionists, or true verbatim available on request and iron-clad confidentiality covered with every file. If you have audio bound for a case file, get in touch, we’ll quote and take it from there.
