OpenAI has announced that it is working on a new system to tackle the problem of AI hallucinations where machines generate false facts. The proposed new method’s approach is to train AI models to reward themselves for each individual correct step of reasoning when arriving at an answer rather than just rewarding a correct final conclusion. OpenAI’s researchers believe that the new approach, process supervision, could lead to better explainable AI. To achieve this goal, OpenAI plans to release an accompanying dataset to the public to enable the wider tech community to develop a better understanding of the issue. Although some independent experts are sceptical about the effectiveness of this proposed method, OpenAI hopes to apply it to its popular ChatGPT chatbot technology, which now has over 100 million monthly users.
OpenAI is a research organization researching artificial intelligence, machine learning, and robotics. The company was founded in San Francisco, California, United States, in 2015 by some of Silicon Valley’s most prominent people, including Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, and John Schulman. To date, Microsoft has invested over $13 billion in OpenAI, more than any other investor.
Ben Winters is a senior counsel at the Electronic Privacy Information Center and leader of its AI and human rights project. He expressed skepticism about the proposed approach, stating that he would prefer to see the full dataset and accompanying examples before judging the effectiveness of the process. He also raised concerns about transparency, suggesting that OpenAI had not provided enough detail about the data used to train and test GPT-4 to evaluate the new method properly. In response, OpenAI did not comment on whether the research had been peer-reviewed or independently reviewed.
