The technology company OpenAI has launched Whisper, an AI-based transcription tool that stands out for its "robustness and near-human precision." Despite this praise, numerous engineers, developers, and researchers have discovered a significant issue with Whisper: it often invents fragments of text or even complete sentences.
According to collected testimonies, some transcriptions generated by Whisper include hallucinations that can be quite inaccurate and, in some cases, even disturbing. A machine learning engineer at OpenAI commented that the failures persist even in well-recorded audio tests, while a third developer claimed to have identified errors in almost all of the 26,000 transcriptions he performed with Whisper.
Despite the problems, Whisper has become extremely popular and is integrated into various platforms, from call centers to voice assistants. However, some researchers have identified notable errors in its operation. For example, in a study conducted by professors from Cornell University and the University of Virginia, racist comments invented by Whisper were found in some transcriptions.
Despite these warnings, many hospitals and medical centers are using tools like Whisper to transcribe medical consultations, with the idea that medical staff spend less time on administrative tasks. Although some companies claim to be aware of Whisper's hallucinations and are trying to address the issue, concerns about the accuracy and privacy of AI-generated transcriptions persist.
Due to the confidential nature of medical interactions, lawmakers and experts have expressed concern about the use of tools like Whisper in high-risk environments. Some researchers have found hallucinations in the majority of the reviewed transcriptions, which could have serious consequences, especially in the healthcare field.
The prevalence of hallucinations in tools like Whisper has led to calls for OpenAI to address these issues and improve the accuracy of its AI models. Despite the implications that transcription errors may have, some assert that, with the proper adjustments, these problems could be resolved.
