OpenAI’s Whisper: Technological Breakthrough with Hidden Risks
Despite this, its application in the medical field is noteworthy.
Whisper by OpenAI: A Controversial Transcription Tool
Picture yourself at the doctor’s office, sharing detailed symptoms, only to find out your account has been misinterpreted due to a faulty transcription. This is the risk patients face at medical centers using Whisper, OpenAI’s transcription tool. An investigation by ABC News revealed that over a dozen developers, software engineers, and academic researchers have found Whisper to produce “hallucinations” – fabricated text involving non-existent medications, racial comments, and violent remarks.
Disturbing “Hallucinations”
The occurrences of these errors are numerous and alarming. A University of Michigan researcher found fabricated text in eight out of ten audio transcriptions of public meetings. In another study, computer scientists identified 187 “hallucinations” from over 13,000 audio recordings. A machine learning engineer noticed errors in half of the transcriptions from a corpus of over 100 hours, while a developer observed hallucinations in nearly all of the 26,000 transcriptions processed by Whisper.
Potentially Severe Consequences
The potential impact of these transcription errors is particularly severe in the medical field. A company named Nabla uses Whisper for its medical transcription tool, employed by over 30,000 clinicians and 40 health systems, having transcribed about seven million visits so far. However, the system for verifying transcription accuracy is flawed. According to Nabla’s CTO, Martin Raison, the tool deletes all audio recordings for “data security reasons”, and providers must quickly edit and approve transcriptions, although improvements to this system are considered.
AI: A Double-Edged Tool
Despite these issues, the latest version of Whisper was downloaded 4.2 million times from the open-source platform HuggingFace last month. The tool is also integrated into Oracle’s and Microsoft’s cloud computing platforms, as well as some versions of ChatGPT. These figures highlight the need for vigilance regarding AI use, whose outcomes can sometimes be misleading or even hazardous.