Impromptu: Amplifying Our Humanity Through AI nations of the Kennedy brothers, he called it a “flawed and irresponsible research tool.” ● Trying to put this so-called technological breakthrough into a suitable context, a skeptical editor exclaimed, “We can get the wrong answer to a question quicker than our fathers and mothers could find a pencil.” If these examples sound like old news, it’s because they are old news. The “it” in question is Wikipedia, not ChatGPT, and the examples are all drawn from articles that appeared in the mid-2000s. We could extend the exercise. Throughout the 1990s, the “it” was the internet itself, which, in the minds of still-mighty old media gatekeepers, was less trustworthy than an issue of the Weekly World News with a cover photo of Hillary Clinton and her adopted alien baby. If you listen closely on a quiet night in Washington, D.C., you can still hear echoes of the protest that erupted when web news pioneer Matt Drudge trolled reporters at the National Press Club by boasting that his accuracy rate was around 80 percent. I don’t bring all this up to excuse complacency about LLMs’ performance. In my own experience using GPT-4, I’ve seen it make up false details about my Masters of Scale podcast. I’ve seen it generate phony links when I’ve asked it to cite sources. I’ve watched it confidently assert a “fact” in one sentence and then inadvertently debunk that “fact” two sentences later with contradictory information it asserts with equal confidence. So of course I believe we should absolutely strive (and are striv- ing) to make LLMs as accurate and reliable as we can. As we pursue that goal, however, we should also keep in mind that 154