OpenAI's AI-powered transcription tool Whisper, praised for its 'human-level accuracy', is under fire for inventing text that was never spoken—from racial remarks to fake medical advice. Experts warn these hallucinations could have real-world consequences as the tech spreads across industries.
When AI Gets Too Creative
Whisper, used globally for translating interviews, generating subtitles, and even transcribing doctor-patient chats, has a habit of making things up. Software engineers and researchers say the tool adds entire sentences that don’t exist in the original audio. Imagine your doctor’s notes including a non-existent treatment!
Medical Fields at Risk
Despite OpenAI’s warnings to avoid using Whisper in 'high-risk' settings like healthcare, some hospitals are adopting it. Think: \"Your diagnosis is… something the AI invented?\" One University of Michigan researcher found errors in 80% of public meeting transcripts before tweaking the model.
How Bad Is It?
A developer found made-up text in nearly all 26,000 transcripts they tested.
Over 13,000 clear audio clips still had 187 hallucinations.
At scale, this could mean millions of errors worldwide.
As one engineer put it: \"AI’s creativity is cool until it starts rewriting reality.\"
Reference(s):
AI-powered transcription tool invents things no one ever said
cgtn.com