Whisper is a popular transcription tool powered by artificial intelligence, but it has a major flaw. It makes things up that were never said.
Curated from apnews.com · 26 October 2024 →
Whisper is OpenAI's speech-to-text model, and by late 2024 a product built on it was in use at around 40 health systems and had been through some seven million medical visits. The AP spoke to engineers and researchers who had studied its output and reported that none had seen a transcription tool invent as much: whole sentences that nobody said, appearing most often in the silences, and including invented medications and racial commentary. One study of over 13,000 clean audio clips found hallucinations in 312 of them. The compounding problem in the clinical case is that the vendor deletes the original audio for privacy reasons, so once the note is written there is nothing left to check it against.