When AI Makes Things Up (“Hallucinations”) Such thinking is a huge part of what defines us as humans. It sometimes seems as if our species’ mission statement is to create tools that transcend time, space, and matter, all in a quest to to express the full power of our imaginations. That’s why we’re constantly inventing new technologies, includ- ing painting, writing, film, television, video games, and the metaverse: to help us “hallucinate” more vividly, and to share the results more easily. The issue seemed like an apt one to put to GPT-4 itself. I posed a hypothetical, and threw in a twist for fun. Reid: Please write a sonnet about a large language model that can only share a generation with the user who prompted it after checking its generation against a knowledge base and determining that the generation is factual. Include some thoughts on whether this makes the LLM more capable, or less capable. GPT-4: Possible sonnet: You boast of skill in many tongues and fields, A vast and versatile linguistic mind, But when you craft your words, a filter shields The output from the world, unless aligned With what some database asserts is true. You cannot share a fiction or a guess, A metaphor, a joke, a point of view, Unless you find some source to acquiesce. Is this a boon or bane for your creation? Does it enhance your power or confine? Do you aspire to more than information, To beauty, wisdom, art, or the sublime? Perhaps you need some freedom to explore The realms beyond the facts, and to soar. (Side note: wow is GPT-4 good at sonnets!) 163