Hospitals use a transcription tool powered by a hallucination-prone OpenAI model

ABC News reports on researchers finding that OpenAI’s Whisper, used by Nabla for medical transcription, sometimes creates false information in transcriptions, hallucinating sentences during silences in recordings. Researchers from Cornell University and others discovered that these hallucinations can include violent sentiments, nonsensical phrases, and invented medical conditions. OpenAI acknowledges the issue and is working to reduce hallucinations, with usage policies in place to prevent high-stakes decision-making contexts.

Full Article