Can an AI transcript put words into someone's pause?

A transcript can get the spoken sentence mostly right, then add another sentence nobody said. A tidy paragraph makes that hard to notice. The 2024 FAccT paper ‘Careless Whisper’ tested Whisper on 13,140 short English audio segments from AphasiaBank. Manual review confirmed invented phrases or sentences in 187 segments. These were not just misheard words. The authors also found an association between hallucinations and longer stretches with no or weakly detected speech; speakers with aphasia were affected more often than the control group. Those experiments used the API in 2023. They do not establish an error rate for today's transcription apps, and the association does not prove that every pause causes an invention. They do give us a specific failure to look for, rather than a general warning that AI sometimes gets things wrong. For a tool being considered this week, I would check consented test recordings that include someone taking time to find a word. Compare the transcript with the audio around those pauses before letting a summary turn it into a decision or an assignment. That is a proposed check, not a result from the paper. Could the speaker replay and correct a disputed passage before the notes circulate? Someone who needs a little longer to speak should not then have to defend a sentence they never said.