OpenAI has announced the development of a new model for language processing that aims to eliminate the problem of hallucinations in AI-generated text. These hallucinations occur when the AI creates responses that are entirely fabricated, leading to incorrect or even dangerous information. Two models have been developed by OpenAI to combat this issue, called outcome supervision and process supervision. The latter trains reward models to provide feedback at every step to create a chain of thought, mimicking human thinking. The tests on a math dataset showed that the process supervision approach led to significantly better performance. Although still in the research phase and only tested on math, OpenAI believes that this step-by-step verification can potentially resolve hallucinations on AI-generated text, ultimately producing better and safer language models for public use.
OpenAI is an artificial intelligence research laboratory consisting of the for-profit OpenAI LP and the non-profit OpenAI Inc. The organization is dedicated to researching and developing AI in a way that is safe and beneficial to humanity. The laboratory was founded in 2015 in San Francisco, California.
This week, a lawyer reported submitting false information to a chatbot AI program to generate fake cases. This highlights the current problem with hallucinations in automated language generation. These false outputs pose a significant risk to individuals who rely on the accuracy of the information received from these systems. OpenAI has attempted to solve this issue by developing the process supervision approach, which trains AI models to mimic human-like thinking processes. This should help eliminate hallucinations and provide a safer and more trustworthy result.
