On Monday, researchers documented cases of artificial intelligence (AI) hallucinations in the Whisper tool of OpenAI.
Based on reports, the machine learning model is encountering frequent episodes of error despite its rapid usage in high-risk industries. The problem refers to instances where the large language model (LLM) detects non-existent patterns.
In connection with this, experts indicate that the audio tool produces text that features racial commentary, fictitious medical treatments, and inappropriate language.
On the other hand, OpenAI claimed that the model possesses near-human-level accuracy.
However, a researcher from the University of Michigan asserted that he discovered that eight out of every 10 speech transcriptions contain hallucinations.
Likewise, a machine learning engineer experienced problems in nearly half of over 100 hours of transcriptions.
Moreover, the LLM is supposedly used by more than 30,000 clinicians across 40 United States (US) health systems. Correspondingly, several researchers raised a warning against the adoption of Whisper following the found issues.
Additionally, AI transcription assistants are commonly expected to commit typographical mistakes. However, engineers stressed that they have never encountered as many hallucinations with other models powered by smart technology.
Meanwhile, Microsoft emphasized that the OpenAI tool Whisper is not intended for high-risk cases, including healthcare.
Orion Model Set to be Released by OpenAI
As indicated by reports, OpenAI will be releasing in December the Orion, its advanced AI model.
In line with this, the organization plans to release the product gradually to allow the development of customized features and products.
More importantly, Microsoft anticipates early access to Orion and integrates it into the Azure cloud platform.
Also, OpenAI is yet to announce whether the new AI tool will directly succeed GPT-4 or represent another type of model.
Furthermore, Chief Executive Officer Sam Altman targets unifying its LLMs into a version that provides a technology surpassing human-level intelligence.
