OpenAI's Whisper, a widely used transcription tool known for its high accuracy, has raised concerns among experts and industry users due to a recurring issue with creating fabricated text known as "hallucinations."
This error, where the tool invents random sentences or phrases, has been found in various fields, including medicine, where incorrect transcriptions could lead to severe consequences.
According to AP News, tech and academia experts report that Whisper's hallucinations can sometimes generate inaccurate transcriptions that include inappropriate or violent language, and even imaginary medical advice.
These errors are especially problematic as Whisper is commonly employed in transcription and translation across industries, including hospitals, customer service and video subtitle creation.
While OpenAI has warned users not to use Whisper in "high-risk" environments, this guidance hasn't stopped medical facilities from adopting the tool to assist with transcribing doctor-patient consultations.
Read also: N
Concerns Grow Over Whisper's Accuracy in Sensitive Fields
In one study, researchers found hallucinations in about 1% of Whisper transcriptions, particularly when background noise, silences, or pauses were present. Errors included incorrect descriptions of people's actions or invented conversations.
This flaw has significant implications for the deaf and hard of hearing community, who rely on transcription for communication but may miss fabricated text "hidden" within accurate information.
Some healthcare providers have integrated Whisper-based tools into their systems to reduce note-taking time. For instance, Nabla, a company that has adapted Whisper for medical use, estimates over seven million consultations have been transcribed using the tool.
Nabla encourages clinicians to review transcripts before relying on them, though some experts worry about the absence of original audio to confirm transcription accuracy.
Calls are increasing for stricter AI regulations to ensure responsible use, especially as AI tools like Whisper become integral to sensitive fields. OpenAI has pledged to refine Whisper's capabilities and welcomes feedback to minimize errors in future updates.
Read also: No
