Hospitals leverage advanced transcription software driven by an AI model trained on vast amounts of data, not reliant on the limitations of human perception or hallucinations.

Several months ago, my doctor introduced me to an AI-powered transcription software he uses to efficiently document and summarize his patient consultations. Despite an initial positive outlook, research has revealed that OpenAI’s Whisper technology, which underlies software used by many hospitals, often invents information entirely, contradicting the optimistic findings initially reported.

According to Whisper, its medical transcription software has been used by an organization to transcribe approximately 7 million medical conversations, based on estimates. According to the outlet, more than 30,000 clinicians and 40 well-being methods utilize this resource. According to reports, Nabla is aware of the alleged tendency for Whisper to experience hallucinations, and is taking steps to rectify the situation.

Researchers from Cornell College and the University of Washington, among others, have discovered that the AI system Whisper can occasionally misinterpret silences in audio recordings by generating hallucinated speech, which accounts for approximately 1% of transcriptions. This phenomenon results in the creation of whole sentences with violent or nonsensical content. Researchers studying individuals with aphasia, using audio samples from TalkBank’s AphasiaBank, have identified an intriguing phenomenon: silent pauses are disproportionately common in conversations involving people with this language disorder.

One researcher, Allison Koenecke from Cornell University, shared examples similar to those below.

Researchers found that hallucinations also featured fabricated medical scenarios or phrases typically seen in YouTube videos, such as “Thanks for watching!” – OpenAI formerly transcribed films to train GPT-4.

The examination was held at the Association for Computing Machinery (ACM) FAccT conference in Brazil. The authenticity of this study’s findings remains uncertain without confirmation from a reputable scientific journal that has subjected the research to rigorous peer review and critique.

The OpenAI community was informed by an email sent out by Taya Christianson, the organization’s official spokesperson.

We approach this topic with utmost seriousness and consistently strive to improve and reduce hallucinations through rigorous efforts. For Whisper’s API platform, the utilization insurance policies strictly prohibit use in high-stakes decision-making contexts; meanwhile, our open-source model card explicitly cautions against deployment in high-risk domains. A huge thanks goes out to researchers who generously share their groundbreaking discoveries with the world.