OpenAI Whisper Speech Recognition Model Hallucinated False Content Including Racial Slurs

In October 2024, researchers at Cornell University and other institutions published findings revealing systematic hallucination issues in OpenAI's Whisper automatic speech recognition model. The study found that Whisper frequently generated entirely fabricated text, including racial slurs, violent language, and other harmful content that was never actually spoken in the source audio. The hallucinations occurred most frequently when Whisper processed silent segments, unclear speech, or audio with background noise. Instead of indicating uncertainty or producing blank output, the model confidently generated plausible-sounding but completely false transcriptions. Researchers found that approximately 1% of audio samples contained hallucinated content, with some samples producing multiple fabricated phrases. The issue was particularly concerning because Whisper has been widely deployed in sensitive applications including hospitals for medical transcription and courts for legal proceedings. In healthcare settings, false transcriptions could potentially contaminate patient records with inappropriate or misleading content. In legal contexts, fabricated statements could affect case outcomes or create liability issues for court reporting services. Researchers tested multiple versions of Whisper and found the hallucination problem persisted across different model sizes and configurations. The fabricated content often included coherent sentences that appeared realistic, making the errors difficult to detect without careful human review. Some hallucinations contained profanity, racial slurs, or references to violence that could be particularly damaging if included in official transcripts. OpenAI acknowledged the issue and stated they were working on improvements, but did not provide a timeline for fixes. The company recommended users implement human review processes for critical applications, though many deployments lacked such safeguards. Industry experts called for better uncertainty quantification and content filtering to prevent similar issues in speech recognition systems used for sensitive applications.