OpenAI's transcription tool is making up things no one said
OpenAI's Whisper, an automatic speech recognition (ASR) system, is under fire from software engineers and academic researchers for its propensity to "hallucinate," or make up false information while transcribing. This is especially alarming in the transcription industry where precision is everything. According to the Associated Press, this AI tool has been observed inserting incorrect details into transcripts, from racial comments to non-existent medical procedures.
Researchers find hallucinations in majority of Whisper transcriptions
A University of Michigan researcher found eight out of 10 audio transcriptions from public meetings contained hallucinations. A machine learning engineer who analyzed over 100 hours of Whisper transcriptions, found inaccuracies in more than half of them. A developer also reported seeing hallucinations in nearly all the 26,000 transcriptions he generated using Whisper. These findings raise significant concerns about the reliability and accuracy of OpenAI's transcription tool.
OpenAI acknowledges issues
In light of these findings, an OpenAI spokesperson said the firm is "continually working to improve the accuracy of our models, including reducing hallucinations." The spokesperson also noted that their usage policies prohibit using Whisper "in certain high-stakes decision-making contexts," admitting the risks of inaccuracies in transcription. The company thanked researchers for sharing their findings, showing willingness to address these issues.
