Transcription Flaws From OpenAI's Whisper Raise Concerns

OpenAI's transcription tool, Whisper, lauded for its near "human level robustness and accuracy," appears to have a significant flaw. Experts reveal that Whisper occasionally fabricates text—known as hallucinations—which can include racial comments or imagined medical treatments. Concerns arise as it gains traction in diverse fields, including medical centers transcribing patient consultations, despite OpenAI's caution against its use in "high-risk domains."

While the scope of Whisper's shortcomings is challenging to assess, a University of Michigan researcher who was studying public meetings said he found hallucinations in 80% of the audio transcriptions he reviewed; a developer said he identified hallucinations in almost every transcript he created—and he created 26,000 of them. Advocates call for government intervention and OpenAI's resolution of this issue.

The tool's integration into prominent platforms like Microsoft's cloud services has led to widespread global usage. "This seems solvable if the company is willing to prioritize it," said William Saunders, a former OpenAI employee. "It's problematic if you put this out there and people are overconfident about what it can do and integrate it into all these other systems." (This story was generated by Newser's AI chatbot. Source: the AP)