“AI hallucinations” occur when AI systems generate false information that does not exist in the original audio or text input. Instead of just mishearing or misunderstanding, the AI makes up details that seem believable. This happens because generative AI tries to extend beyond its training data and sometimes produces wholly inaccurate or harmful results.<\/p>\n
<\/p>\n
In medical transcription, hallucinations might involve invented medical facts, wrong drug names, fake clinical cases, or added commentary that misrepresents what a healthcare provider said. For example, Koenecke et al. (2024) found that OpenAI\u2019s Whisper\u2014an AI transcription model used by companies like Nabla\u2014sometimes created fabricated violent incidents, racial remarks, and medical terms not originally spoken.<\/p>\n
<\/p>\n
A recent study by researchers from Cornell University and the University of Washington reported that Whisper hallucinated in about 1.4% of transcriptions. While this seems low, clinical work relies heavily on accurate patient records. Nearly 40% of these hallucinations had the potential to cause harm by distorting speaker intent or adding false clinical information. Errors in documentation can lead to misdiagnosis, incorrect treatments, or missed symptoms, all of which threaten patient safety.<\/p>\n
<\/p>\n
<\/p>\n
<\/p>\n
<\/p>\n
Nabla uses OpenAI\u2019s Whisper in its AI-powered medical transcription tool. Their system serves over 30,000 clinicians in more than 70 medical organizations across the United States and has processed about 7 million medical visits.<\/p>\n
<\/p>\n
The company compiled a proprietary dataset of roughly 7,000 hours of real medical audio, incorporating feedback from nearly 10,000 physicians to better tailor Whisper to healthcare-specific language. Despite these efforts, hallucinations still occur, and Nabla is actively working to reduce their frequency.<\/p>\n
<\/p>\n
A major concern is Nabla\u2019s policy of deleting original audio recordings after transcription. While this protects patient privacy and reduces storage needs, it prevents providers from verifying transcription accuracy or correcting errors. Since accurate documentation is key to diagnosis and treatment, not having the original audio reduces transparency and accountability.<\/p>\n
<\/p>\n
Nabla reports a 99.3% word accuracy rate, but this figure does not fully reflect the severity of hallucinations. Hallucinations often involve entire false sentences rather than simple word mistakes, so basic word error metrics can hide critical problems.<\/p>\n
AI hallucinations in medical transcription raise several risks for medical administrators and owners. The main concern is patient safety. Incorrect entries in Electronic Health Records (EHRs) can result in:<\/p>\n
<\/p>\n
<\/p>\n
<\/p>\n
Healthcare providers depend heavily on accurate records. Relying on AI transcripts without careful review poses dangers. These risks grow in complex cases or with vulnerable patients, such as those with aphasia or speech impairments. Research shows Whisper hallucinations happen more often during silences or pauses typical in such patients, risking miscommunication.<\/p>\n
<\/p>\n
Legal and regulatory problems may arise as well. Misdocumentation related to hallucinations could lead to malpractice claims or penalties for failing to meet care standards. Organizations like Microsoft advise legal review before using AI transcription in healthcare, making risk management essential for medical offices.<\/p>\n
<\/p>\n