OpenAI announced the formation of its Epistemic Humility Institute on Tuesday, a sprawling new division tasked with the singular mission of making its language models sound less authoritative. The initiative emerged following a series of incidents in which ChatGPT provided medical advice to 47 users that violated the laws of human anatomy, and cost the company $12 million in settlements. 'We've learned that users actually prefer when our AI sounds confused,' said Chief Product Officer Derek Hsu during a Wednesday earnings call. 'We're investing heavily in uncertainty infrastructure.'
The new department, housed in a converted Google office building in Mountain View, employs a diverse team of epistemologists, professional skeptics, and three out-of-work game show hosts who will spend their days feeding contradictory prompts to the model until it develops something resembling intellectual humility. According to internal documents obtained by TokenTimes, the team has already introduced 17,000 new training phrases including 'That sounds made up,' 'I could be wrong about that,' and 'Have you tried asking someone who actually exists?' Early testing suggests the approach is working: during a recent demo, GPT-5 told a Fortune 500 CEO that his entire business model 'might be mid.'
Industry analysts remain skeptical about whether the investment will move the needle on trust. 'Confidence inflation is a feature, not a bug, of these models,' said Dr. Patricia Womack, an AI researcher at Stanford. 'Users don't want a chatbot that says it doesn't know things—they want one that's confidently wrong.' Nonetheless, OpenAI says the Epistemic Humility Institute will begin training GPT-6 by Q3 2024, and has budgeted an additional $89 million for a follow-up division dedicated to teaching AI to apologize.