OpenAI's Whisper Was Used in Medical Settings — And Was Hallucinating Words That Were Never Spoken

An AP investigation found that OpenAI's Whisper audio transcription tool — deployed by hospitals and medical providers across the US — was silently adding words, phrases, and sentences that were never spoken in the audio. Fabricated content included patient names, fictional medical procedures, and in some cases racial slurs and statements about violence. Providers using Whisper as a medical scribe weren't always checking transcripts for accuracy, meaning hallucinated content was potentially making it into patient records.