Who is responsible when an AI transcription or note-taking tool adds words nobody said, and how do you prove what was really said?

What should I check, step by step?

1. Find out whether your transcription tool keeps the original audio, for how long, and whether you can export it. If it deletes audio by default, decide whether you can keep your own copy within your privacy and consent rules.

2. When a transcript is made, hash the audio file and the transcript (SHA-256) and record the tool, model and version, so you can show later which audio produced which text.
3. Have a named person review a transcript before it becomes a record: a signed clinical note, filed minutes, a quoted statement. Record who approved it and when.
4. Review hardest where hallucinations cluster: long pauses and silence, background noise or music, speech impairments, accents and lower-resource languages. Whisper's model card says hallucinations may be worse in lower-resource languages.
5. Watch for the telltale additions: sentences at the end of a recording, names, medications, numbers or relationships nobody mentioned, and video-style lines such as thanking viewers or asking people to subscribe.
6. When someone disputes a transcript, go back to the audio, not the text. If the audio is gone, say so; don't treat the transcript as proof of what was said.
7. Keep corrections visible: store the original machine transcript, the corrected version and who made the change, rather than overwriting the text.
8. Tell people when they're being recorded and transcribed, and get consent where the law requires it. Whisper's model card also cautions against transcribing people recorded without their consent.
9. Check your contract with the vendor for who bears errors, how long audio and transcripts are kept, and whether you can export both if you leave.

How often does Whisper hallucinate words that weren't in the audio?

Careless Whisper (Koenecke and colleagues, ACM FAccT 2024) tested OpenAI's Whisper on thousands of short audio clips. Most transcriptions were highly accurate, but roughly 1% contained entire hallucinated phrases or sentences that didn't exist in any form in the audio. In one example, Whisper correctly transcribed a simple sentence and then added five more containing words like terror, knife and killed.

• The authors sorted the harmful hallucinations into perpetuating violence, inaccurate associations (made-up names, locations, relationships and health statuses) and false authority (lines like a YouTuber's sign-off, or links to websites). They found 38% of hallucinations included explicit harms, and that hallucinations happened disproportionately for speakers with longer non-vocal pauses, a common symptom of aphasia.
• Whisper's own model card explains the cause: because the models are trained on large-scale noisy data, they combine predicting the next word with transcribing the audio, so predictions may include text not actually spoken. It recommends against use in high-risk domains.

Why does deleting the original audio make AI transcription errors hard to catch?

The AP reported in October 2024 that Nabla's Whisper-based medical tool was used by over 30,000 clinicians and 40 health systems, and that its chief technology officer said the tool erased the original audio for data safety reasons. Nabla said clinicians edit and approve the notes. Without the audio, though, nobody can later check a disputed sentence against what was actually said.

• Deleting audio can be a reasonable privacy choice. If you make it, the review before approval is the only check you get, so record that it happened: who reviewed, when, and a hash of the exact text they approved.

Who is responsible for errors in an AI-generated transcript or medical note?

No court has set a general rule for transcription hallucinations. But records are usually judged as the record of whoever signs or adopts them, so the clinician who signs a note, the company that files minutes and the reporter who quotes an interview own what the transcript says.

• That's why the useful evidence isn't the transcript alone. It's the link between three things: the audio, the machine transcript and the approved version, each with a hash and a time, plus who approved it. With that, a dispute can be settled by listening to the audio; without it, it comes down to whose memory is better.