Hidden Dangers in OpenAI's Whisper Transcription Tool (Full Transcript)

Speaker 1: AI is changing the world around us, powering tools that make life easier, like Whisper, OpenAI's transcription model. But there's a hidden flaw in Whisper that could turn these conveniences into serious risks. OpenAI has been promoting Whisper as a robust transcription tool, claiming it delivers near-human-level accuracy. From transcribing interviews to generating text in consumer tech and creating video captions, Whisper has rapidly expanded across industries worldwide. But according to software engineers, developers, and researchers, Whisper has a critical flaw – hallucinations. Unlike typical transcription errors like misspellings, hallucinations create completely fabricated content. Experts say this can range from racial commentary and violent language to imaginary medical advice. For instance, researchers have seen Whisper fabricate shocking statements. In one transcription, a speaker says, he, the boy, was going to take the umbrella, and Whisper added, he took a big piece of a cross, he killed a number of people. Another transcription turned a description of people as two other girls and one lady into two other girls and one lady, which were black. And Whisper even created a non-existent drug, hyperactivated antibiotics. This isn't just limited to consumer tech. Whisper has found its way into hospitals and clinics. Over 30,000 clinicians across 40 health systems now use Whisper-based tools to transcribe patient consultations, aiming to save doctors' time. But OpenAI warns against using Whisper in high-stakes situations where accuracy is crucial. Imagine the potential risks – a doctor receiving a hallucinated transcription could lead to severe misinterpretations, even misdiagnoses. Alondra Nelson, a former White House science advisor, stated it best, nobody wants a misdiagnosis, there should be a higher bar. The deaf and hard-of-hearing also rely on Whisper's transcriptions for captions. Yet fabricated phrases could be mistaken as real content, putting them at an even greater risk since they might not have another way to verify the text. The spread of Whisper has also raised privacy concerns. California lawmaker Rebecca Bauer-Cahan declined to sign a consent form allowing her child's medical audio to be shared with tech companies, fearing sensitive data might end up in the hands of corporations like Microsoft, OpenAI's largest backer. Former OpenAI employees and industry experts have urged the company to address Whisper's hallucinations, and some even believe federal regulations are needed to oversee these AI models. Whisper is powerful, but in high-risk fields unaddressed hallucinations are simply too dangerous. OpenAI is aware of the hallucination issue, saying they're working on reducing errors. But with millions of transcriptions already out there, this is a problem we can't afford to ignore. For AI to truly help, it must be trustworthy.