Tech giant OpenAI has been lauded for its artificial intelligence-powered transcription tool, Whisper, which promises near human-level robustness and accuracy.
However, a significant flaw has emerged: Whisper frequently generates fabricated text, including racial commentary, violent rhetoric, and even imagined medical treatments. More than a dozen software engineers, developers, and academic researchers have reported these issues, referring to them as hallucinations.
These fabrications are particularly problematic as Whisper is widely used across various industries to translate and transcribe interviews, generate text for consumer technologies, and create video subtitles. Experts are especially concerned about the tool's use in medical centers for transcribing patient consultations, despite OpenAI's warnings against deploying Whisper in high-risk domains.
The full scope of Whisper's inaccuracies is challenging to determine, but evidence suggests a pervasive problem. A University of Michigan researcher found hallucinations in eight out of every ten audio transcriptions he reviewed. Similarly, a machine learning engineer discovered errors in about half of over 100 hours of transcriptions, while another developer identified issues in nearly every one of 26,000 transcripts created with Whisper.
Even in well-recorded, short audio samples, Whisper's reliability is questionable. A recent study by computer scientists uncovered 187 hallucinations in over 13,000 clear audio snippets, indicating that tens of thousands of faulty transcriptions could result from millions of recordings.
As Whisper continues to be integrated into various sectors, the prevalence of these errors underscores the need for improved oversight and enhancements to ensure accuracy and reliability.
Reference(s):
AI-powered transcription tool invents things no one ever said
cgtn.com
