OpenAI’s artificial intelligence model is delusional

In 2022, the OpenAI company unveiled a series of “Whisper” models; An artificial intelligence model developed for speech transcription. It has recently become clear that this artificial intelligence model is struggling with serious problems.

According to software engineers, developers, and academic researchers, productive AI tools in general tend to be delusional; But the strange thing about the Whisper tool is that it also has problems in transcribing speech. While Whisper is expected to transcribe exactly what is said; But this tool cannot accurately transcribe speech. The problem referred to is very serious and is not related to the common mistakes of these types of models in writing a similar word instead of the original word or misspelling.

According to users of this AI model, the service includes things like racial commentary and imaginary medical treatments in its transcriptions, which can be very dangerous. Many different hospitals and medical centers are now using this tool in medical fields, and correct transcription may lead to disaster.

One of these researchers from the University of Michigan says that the Whisper tool was hallucinating in 8 out of 10 transcription cases. Also, the machine learning engineer who studied more than 100 hours of Whisper’s transcriptions, says that more than half of these transcriptions contain errors and delusions. In addition, the developer also claims to have found hallucinations in all 2,600 hours of transcriptions of the tool.

OpenAI responded to the reports, with a company spokesperson saying that they are constantly working to improve the accuracy of the models and reduce the illusions. He also said their policies prohibit the use of Whisper “in certain high-stakes decision-making contexts.” Finally, the company commended the researchers for sharing their findings.