On May 31, OpenAI announced its efforts to improve the mathematical problem-solving capabilities of ChatGPT, with the goal of reducing hallucinations from artificial intelligence (AI). OpenAI highlighted that mitigating hallucinations is a crucial step towards developing aligned AI.
In March, the introduction of the latest version of ChatGPT, ChatGPT-4, further pushed AI into the mainstream. However, generative AI chatbots have long struggled with factual accuracy, occasionally generating false information, commonly referred to as “hallucinations.” Efforts to reduce these AI hallucinations are announced via a post on the OpenAI website.
AI hallucinations refer to cases where artificial intelligence systems generate results that are incorrect, misleading, or not supported by real-world data. These hallucinations can manifest in a variety of ways, such as generating false information, making up events or people that don’t exist, or providing inaccurate details about certain topics.
Open AI carried out an investigation to examine the effectiveness of two types of feedback: “monitoring of results” and “monitoring of processes”. Outcome monitoring is based on the end result, while process monitoring provides information about each step in a chain of thought. OpenAI evaluated these models using mathematical problems, generating multiple solutions and selecting the best rated solution based on each feedback model.
After extensive analysis, the research team found that process monitoring offered superior performance by encouraging the model to follow a human-approved process. In contrast, performance monitoring was more difficult to analyze in a coherent way.
OpenAI recognized that the implications of process monitoring go beyond mathematics and that more research is needed to understand its effects in different fields. He expressed the possibility that if observed results hold in broader contexts, process monitoring could offer a favorable mix of performance and alignment compared to outcome monitoring. To facilitate research, the company made the full process monitoring data set public, inviting exploration and study in this area.
Although OpenAI did not provide the explicit cases that prompted its research on hallucinations, two recent events exemplified the problem in real-life scenarios.
In a recent incident, Mata v. Avianca Airlines attorney Steven Schwartz admitted to relying on the chatbot as an investigative resource. However, the information provided by ChatGPT turned out to be completely false, which highlights the problem.
OpenAI’s ChatGPT is not the only example of AI systems suffering from hallucinations. during a demonstration of its chatbot technology in March, Microsoft’s Bing AI chatbot sifted through earnings reports and produced inaccurate numbers for companies like the Gap and Lululemon.
Clarification: The information and/or opinions expressed in this article do not necessarily represent the views or editorial line of Cointelegraph. The information presented here should not be taken as financial advice or investment recommendation. All investment and commercial movement involve risks and it is the responsibility of each person to do their due research before making an investment decision.
Investments in crypto assets are not regulated. They may not be suitable for retail investors and the entire amount invested may be lost. The services or products offered are not directed or accessible to investors in Spain.