Kenne Michael 
The organization acknowledges on May 31, that addressing hallucinations is a crucial stage in creating AI systems compatible with human values.
The release of ChatGPT-4 in March signified a significant advancement in the field of generative AI chatbots. However, guaranteeing factual accuracy has long been challenging for such systems, as they occasionally produce erroneous data, also known as “hallucinations.” Through a post on their official website, OpenAI announced the reduction of these hallucinations.
The term “AI hallucinations” refers to situations in which AI systems generate outputs that are factually inaccurate, misleading, or unsubstantiated by real-world data. These hallucinations can manifest in various ways, including the generation of misleading information, the fabrication of nonexistent events or people, and the provision of inaccurate information on particular subjects.
OpenAI researched to evaluate the effectiveness of two categories of feedback to address this issue: “outcome supervision” and “process supervision.” In outcome supervision, feedback is provided based on the outcome, whereas in process supervision, feedback is provided for each step in a chain of reasoning. OpenAI evaluated these models utilizing mathematical problems, producing multiple solutions, and selecting the solution with the highest ranking based on each model’s feedback.
The research team discovered, through in-depth analysis, that process supervision improved performance by encouraging the model to adhere to a human-approved reasoning process. In contrast, outcome supervision was more challenging to examine consistently.
OpenAI recognizes that the implications of process supervision extend beyond mathematics and that further research is required to comprehend its effects in various domains. According to the organization, if the observed outcomes hold in broader contexts, process supervision may offer a more desirable combination of performance and alignment than outcome supervision. OpenAI has made available the entire dataset of process supervision in support of research endeavors, inviting exploration and study in this area.
While OpenAI did not cite specific incidents that prompted its investigation into hallucinations, two recent incidents illustrate the issue in real-world situations.
In one instance, Lawyer Steven Schwartz, in Mata v. Avianca Airlines case, acknowledged using ChatGPT as a research resource. The fact that the information provided by ChatGPT has been wholly fabricated highlights the issue at hand.
The OpenAI ChatGPT is not the only AI system to experience hallucinations. During a March demonstration of its chatbot technology, Microsoft’s Bing AI chatbot analyzed earnings reports and produced inaccurate figures for companies such as Gap and Lululemon.