On May 31, OpenAI announced an effort to enhance ChatGPT's mathematical problem-solving abilities, with the aim of reducing instances of artificial intelligence (AI) hallucinations. OpenAI emphasizes that alleviating hallucinations is a critical step in developing consistent AI.
In March, the launch of the latest version of ChatGPT ChatGPT-4 pushed AI further into the mainstream. Generative AI chatbots, however, have long struggled with factual accuracy, occasionally generating false information, often referred to as "hallucinations." A post on the OpenAI website announced efforts to reduce these AI illusions. AI hallucinations are situations where artificial intelligence systems produce output that is actually incorrect, misleading, or unsupported by real-world data. These hallucinations can manifest in various for ms, such as generating false information, making up events or people that don't exist, or providing inaccurate details about certain topics. OpenAI conducted research to examine the effectiveness of two types of feedback: "outcome monitoring" and "process monitoring".Outcome supervision involves feedback based on final results, while process supervision provides input for each step in the thought chain. OpenAI evaluates these models using mathematical problems, generates multiple solutions and chooses the highest-ranked solution based on each feedback model.
After a thorough analysis, the research team found that process supervision yields superior performance because it encourages the model to adhere to human-approved processes. In contrast, outcome oversight proved more difficult to review on an ongoing basis. OpenAI recognize s that the impact of process supervision goes beyond mathematics and requires further investigation to understand its impact in different domains. It expresses the possibility that process monitoring can provide a favorable combination of performance and consistency compared to outcome monitoring if the observed res ults hold in a broader context. To facilitate research, The company has publicly released a complete process supervision data set, and sincerely invites exploration and research in related fields. While OpenAI didn't't provide clear examples that prompted its investigation of hallucinations, two recent incidents demonstrate real-life problems.
In a recent incident, Steven Schwartz, an attorney in the Mata v. Avianca airline case, admitted to relying on chatbots as a research resource. However, it turned out that the information provided by ChatGPT was completely fabricated, highlighting the problem at hand.
OpenAI's ChatGPT isn't the only example of an AI system encountering hallucinations. In a demonstration of chatbot technology in March, Microsoft's Bing AI chatbot checked earnings reports and generated inaccurate numbers for companies like Gap and Lululemon.
