According to a report on October 28th by Fast Technology, OpenAI's AI speech transcription tool Whisper has been exposed for serious flaws, as it has been found to invent large segments and even entire sentences of false information out of nothing, causing widespread concern.

Particularly noteworthy is that some medical institutions have publicly admitted to using Whisper to record consultations between doctors and patients, which quickly sparked a furor on the internet.

Currently, numerous medical institutions, including the Mankato Clinic in Minnesota, USA, and Children's Hospital Los Angeles, have over 30,000 clinicians and 40 health systems using tools developed by French AI healthcare company Nabla based on Whisper.

However, it is worth noting that OpenAI has previously issued a clear warning that Whisper is not suitable for use in "high-risk areas."

Upon preliminary analysis of over 100 hours of Whisper transcription data, a machine learning engineer was shocked to find that approximately half of the content contained "hallucinations," meaning the transcribed content was significantly different from the original speech.

The specific cause of this serious flaw in Whisper remains unknown. However, some software developers speculate that these fictional contents are often related to pauses in speech, background noise, or music playback.

In response to this issue, OpenAI has stated that the company is actively researching ways to effectively reduce the generation of fictional content and expressed gratitude for the researchers' findings. Simultaneously, OpenAI has promised to incorporate corresponding feedback mechanisms into subsequent model updates to further improve transcription accuracy and reliability.