Impromptu: Amplifying Our Humanity Through AI little of it. But they have released some, and the picture is actu- ally fairly encouraging. For example, in January 2022, OpenAI published a paper about a “sibling” model to ChatGPT, called InstructGPT, noting that fine-tuning involving human feed- back significantly reduced the model’s toxic outputs and hallu- 7 cinations. And, when problems do occur, companies are often well-incen- tivized to correct them. Microsoft was roundly criticized for rushing Bing Chat to market when Bing’s alter ego Sydney emerged and started inter- acting with some users in obviously undesirable ways. Within days, Microsoft administered a blunt fix by limiting users’ ses- sions with Bing to six questions before requiring a reset. That apparently solved the immediate issue; I think it also underscored one of online software distribution’s key virtues. It’s why I think regulation in this realm will likely be much different than regulation for automobiles or pharmaceuticals: when you need to issue a recall in the digital world, you can do so with great speed and completeness. Just the facts, LLM? There is another aspect to LLMs’ potential for generating non-factual material. When someone prompts GPT-4 to write a fictional story, or create a speculative business plan, or even just to make up six potential tag lines for a new product, these are “made-up” texts that would rightly be described as produc- tive and even creative. They amplify our human ability to think imaginatively and conceptually. 7 OpenAI reported that InstructGPT’s hallucination rate was 17 per- cent, compared to 41 percent for GPT-3. 162