One of the key improvements promised by GPT-4 is the ability to allay critics’ fears about chatbots. The internal analysis shows that the new chatbot is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5.
In terms of safety and alignment updates, the chatbot relied heavily on human feedback, including user feedback. “We’ve incorporated lessons learned from real-world application of our previous models into GPT-4’s safety research and monitoring system,” according to the update.
