Whisper Speech-to-Text AI Reportedly Found to Create Violent Hallucinations

February 12, 2024

Researchers at Cornell reportedly found that OpenAI's Whisper, a speech-to-text system, can hallucinate violent language and fabricated details, especially with long pauses in speech, such as from those with speech impairments. Analyzing 13,000 clips, they determined 1% contained harmful hallucinations. These errors pose risks in hiring, legal trials, and medical documentation. The study suggests improving model training to reduce these hallucinations for diverse speaking patterns. And to stay informed about AI incidents, visit JOIN US.

Matched TAIM controls

Suggested mapping from embedding similarity (not a formal assessment). Browse all TAIM controls

Alleged deployer
Whisper users, Whisper, Organizations integrating Whisper into customer service systems, OpenAI, Companies using Whisper
Alleged developer
OpenAI
Alleged harmed parties
Users whose speech is misinterpreted by Whisper, Professionals relying on accurate transcriptions, People with disabilities, Individuals with speech impairments, General public

AI governance case studies

For forensic AI governance failure analysis (TAIMScore™ case studies), browse Human Signal’s Failure Files™.

Source

Data from the AI Incident Database (AIID). Cite this incident: https://incidentdatabase.ai/cite/732

Data source

Incident data is from the AI Incident Database (AIID).

When citing the database as a whole, please use:

McGregor, S. (2021) Preventing Repeated Real World AI Failures by Cataloging Incidents: The AI Incident Database. In Proceedings of the Thirty-Third Annual Conference on Innovative Applications of Artificial Intelligence (IAAI-21). Virtual Conference.

Pre-print on arXiv · Database snapshots & citation guide

We use weekly snapshots of the AIID for stable reference. For the official suggested citation of a specific incident, use the “Cite this incident” link on each incident page.