Could bad AI captions make a good speaker look less clear?

Before telling someone their recorded presentation was hard to follow, check whether the captions got their words right. A 2026 paper, ‘Lost in Transcription,’ reports a preregistered experiment with usable responses from 207 US-based participants. They watched short talk clips with accurate or error-prone automatic subtitles, with audio available. Faulty subtitles lowered ratings of both the speakers and the content. The speaker questions included whether they sounded clear and knowledgeable. The study used AI-generated voices to vary the accents while keeping the people on screen unchanged. It did not find an additional penalty for the accent groups tested once subtitle quality was controlled. Nor did it establish whether viewers blamed the person or the software. This is evidence about these controlled clips, not a measured effect on hiring decisions or every captioning tool. For a team sharing recorded talks, I would make caption checking part of publication rather than leave the speaker to repair their reputation afterward. If a passage seems muddled, compare it with the audio before writing feedback. Keep captions available; correct the words that misrepresent the person. That is a practice I would propose, not one tested by this paper. How does your team separate a confusing explanation from a transcription mistake?