The remarkable skills of OpenAI’s transcription tool, Whisper, are being praised for reaching almost “human level robustness and accuracy.” But specialists have found a serious problem: Whisper frequently makes up text; this is known as “hallucination.” According to interviews with researchers, developers, and software engineers, these errors might be anything from small typos to completely made-up words that contain offensive language or racial comments.
Whisper’s extensive use across a number of industries presents significant issues, especially in the healthcare industry where it is being used more and more to transcribe patient sessions. Many healthcare facilities are implementing Whisper in spite of OpenAI’s warnings against its use in “high-risk domains,” which could endanger patient care. Experts worry that these lies may result in incorrect medical advice or major misdiagnosis.
The magnitude of Whisper’s errors seems significant. During a study on public meetings, a researcher from the University of Michigan reported that in 80% of the audio transcriptions he examined, he found hallucinations. In a similar vein, a machine learning engineer discovered that about half of the more than 100 hours of Whisper transcriptions contained fake material. Another developer used the technique to create 26,000 transcripts, and in almost all of them, he found hallucinations.
Even with high-quality audio recordings, these problems still exist. Computer scientists found 187 hallucinations in over 13,000 undistorted audio samples, indicating that there may be tens of thousands of inaccurate transcriptions in millions of recordings. This raises serious concerns regarding the tool’s dependability, particularly in situations where precise information is essential.
These errors have particularly serious repercussions in medical settings. Former White House Office of Science and Technology Policy head Alondra Nelson pushed for a higher level of accuracy in AI systems used in healthcare, highlighting the serious repercussions that could result from incorrect diagnosis.
The dangers are further increased by Whisper’s use in producing closed captions for the Deaf and hard of hearing people. Deaf person Christian Vogler noted that this group is especially at risk since they are unable to quickly identify errors concealed in transcriptions.
In conclusion, the ubiquity of hallucinations poses serious hazards, particularly in high-stakes settings like healthcare, even though OpenAI’s Whisper tool exhibits considerable promise. This demonstrates how urgently strict regulation and advancement of AI technology are required to guarantee their safe and efficient use.
SOURCE :
AP NEWS
