OpenAI's Whisper caught hallucinating in up to 80% of transcripts

If you rely on AI to transcribe meetings, you probably expect the occasional missed word or misspelled name. But what happens when the tool just starts fabricating whole sentences out of thin air, complete with violent rhetoric or fake medical treatments?

That is the exact problem hitting OpenAI's open-source speech recognition tool, Whisper. According to an AP investigation, researchers and software engineers are finding that the model hallucinates at an alarming rate.

At the University of Michigan, a researcher studying public meetings found hallucinations in eight out of every ten Whisper transcripts he inspected. Another engineer reviewed 100 hours of audio and spotted fabrications in roughly half of them. One developer even claimed that almost all of the 26,000 transcripts he generated contained hallucinated text.

Making things up in high-stakes settings

Researchers told the AP that they had never seen another transcription tool hallucinate this aggressively. The glitch doesn't just hit noisy or long recordings either; computer scientists found made-up text even in short, crystal-clear audio clips.

More than a dozen engineers and researchers confirmed that Whisper fabricates entire phrases and sentences, ranging from racial commentary to completely invented medical treatments.

That would be concerning enough on its own, but Whisper is already everywhere. It is integrated into versions of ChatGPT, used across call centers, hosted on cloud platforms from Microsoft and Oracle, and was downloaded over 4.2 million times from Hugging Face last month alone.

Crucially, medical clinics are increasingly using Whisper-based tools to transcribe doctor-patient consultations. As Alondra Nelson from the Institute for Advanced Study warned, nobody wants a misdiagnosis because an AI hallucinated symptoms or medications.

It is also a silent hazard for the deaf community. Christian Vogler, director of Gallaudet University's Technology Access Program, pointed out that deaf and hard-of-hearing users have no realistic way to catch hallucinations hidden deep inside otherwise normal text.

OpenAI claims Whisper approaches human-level accuracy and advises against using it in high-risk decision-making contexts. A company spokesperson stated they are actively studying ways to reduce hallucinations and will incorporate feedback into future updates.

Why it matters

Transcription was supposed to be the safe, boring part of AI. While everyone knows large language models can lie, people generally assumed speech-to-text tools just recorded what was actually said. Whisper's issues completely shatter that assumption.

For regular users, if you are transcribing a casual podcast or a routine interview, Whisper remains useful as long as a human reviews the output. But right now, we think it is reckless to rely on it in zero-review pipelines, especially in legal, academic, or healthcare workflows.

For the broader industry, this exposes a massive blind spot. Companies are rushing to bolt open-source models onto critical infrastructure to cut costs. But when the foundation model invents quotes out of silence, downstream tools inherit contaminated data.

OpenAI's playbook of open-sourcing powerful tech while slapping a disclaimer against high-risk use feels increasingly inadequate. When a tool is downloaded millions of times and pushed into clinics, fine print does not prevent medical errors. Unless transcription models find a way to stay strictly grounded in audio, ubiquity is turning into a massive liability.