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

A couple of months in the past, my physician confirmed off an AI transcription instrument he used to document and summarize his affected person conferences. In my case, the abstract was tremendous, however researchers cited by ABC Information have discovered that’s not at all times the case with OpenAI’s Whisper, which powers a instrument many hospitals use — typically it simply makes issues up solely.

Whisper is utilized by an organization called Nabla for a medical transcription instrument that it estimates has transcribed 7 million medical conversations, in accordance with ABC Information. Greater than 30,000 clinicians and 40 well being programs use it, the outlet writes. Nabla is reportedly conscious that Whisper can hallucinate, and is “addressing the issue.”

A gaggle of researchers from Cornell College, the College of Washington, and others found in a study that Whisper hallucinated in about 1 p.c of transcriptions, making up complete sentences with typically violent sentiments or nonsensical phrases throughout silences in recordings. The researchers, who gathered audio samples from TalkBank’s AphasiaBank as a part of the examine, word silence is especially frequent when somebody with a language dysfunction referred to as aphasia is talking.

One of many researchers, Allison Koenecke of Cornel College, posted examples just like the one beneath in a thread about the study.

The researchers discovered that hallucinations additionally included invented medical circumstances or phrases you would possibly count on from a YouTube video, reminiscent of “Thanks for watching!” (OpenAI reportedly used to transcribe over a million hours of YouTube movies to coach GPT-4.)

The examine was presented in June on the Affiliation for Computing Equipment FAccT convention in Brazil. It’s not clear if it has been peer-reviewed.

OpenAI spokesperson Taya Christianson emailed a press release to The Verge:

We take this concern significantly and are frequently working to enhance, together with decreasing hallucinations. For Whisper use on our API platform, our utilization insurance policies prohibit use in sure high-stakes decision-making contexts, and our mannequin card for open-source use contains suggestions towards use in high-risk domains. We thank researchers for sharing their findings.