Software program engineers, builders, and educational researchers have critical issues about transcriptions from OpenAI’s Whisper, in keeping with a report within the Related Press.
Whereas there’s been no scarcity of dialogue round generative AI’s tendency to hallucinate — principally, to make stuff up — it’s a bit stunning that this is a matter in transcription, the place you’d anticipate the transcript intently observe the audio being transcribed.
As a substitute researchers instructed the AP that Whisper has launched every part from racial commentary to imagined medical remedies into transcripts. And that might be notably disastrous as Whisper is adopted in hospitals and different medical contexts.
A College of Michigan researcher learning public conferences discovered hallucinations in eight out of each 10 audio transcriptions. A machine studying engineer studied greater than 100 hours of Whisper transcriptions and located hallucinations in additional than half of them. And a developer reported discovering hallucinations in almost all of the 26,000 transcriptions he created with Whisper.
An OpenAI spokesperson mentioned the corporate is “regularly working to enhance the accuracy of our fashions, together with decreasing hallucinations” and famous that its utilization insurance policies prohibit utilizing Whisper “in sure high-stakes decision-making contexts.”
“We thank researchers for sharing their findings,” they mentioned.