After reading the report by Sweeney et al. , who found that “AI artificial intelligence-generated content was detected in 89. 8% of articles, ” I am left with only 1 question: what is going on with the other 10. 2%? Scientific authors who shun modern chatbots are not romantic holdouts like diarists who insist on fountain pens or jazz purists who listen to Coltrane exclusively on vinyl. They are more like statisticians who perform linear regression with an abacus. Chatbots are world-class editors, and failing to use them is simply silly. Let me tell you how I came to this point of view. I am not a natural writer, so if I have ever written something that another person valued, it is because I have become pretty adept at taming my rough drafts. Rewriting has been the closest thing to a secret superpower that I have had, going all of the way back to my time in college decades ago. At my alma mater, even the premedical students had to complete the Core Curriculum, which consisted of seminars requiring substantial amounts of writing. Despite the workload, it was considered polite, if not required, to submit papers typed and double-spaced. Some classmates labored on a battered Olivetti typewriter. Others outsourced the job to Miss Dee’s Secretarial Service at 1. 50 per page. I used Emacs, the LISP text editor still revered in some dorms at MIT. I handed in my essays on faded green computer paper, hand-trimmed to 8. 5 × 11 inches. The beauty of this setup was simple: the marginal cost of fixing an awkward sentence detected after printing was essentially zero. As a result, I got in the habit of submitting my nth draft, not my first. And because n could be arbitrarily large, I found myself investing more and more energy into rewriting. Over the years, as my college diploma has yellowed, editing tools far fancier than Emacs have become routine. Dedicated rewriters tend to adapt, embracing the newest tools. In general, new tools bring new risks, and the more powerful the tool, the greater the risk. With AI, ceding authority to chatbots courts catastrophe. Chatbots aim to be maximally engaging, and if fabricating citations counts as “engaging, ” they will do it without hesitation. (I am still grieving the loss of “Bernstein, J. My Nobel Prize Acceptance Speech. ”) Just as surgeons cannot dodge malpractice liability by blaming wayward scalpels, authors cannot evade retraction shame by blaming hallucinatory AIs. Now, you may say that it is cheating to get a computer’s help with your writing. Allow me to disagree. First, as Hercule Poirot points out in the Agatha Christie short story, “The Case of the Missing Will”1, knowing when to seek additional expertise is itself a form of expertise. As for the idea that using technology gives you an unfair advantage, I will simply note that when U. S. Secretary of State Henry Stimson shut down the government’s wiretapping office in 1929, declaring that “Gentlemen do not read each other’s mail”2, history proved him wrong. To be sure, using chatbots well requires the gift of discernment: the ability to recognize when the advice is good. This gift can be a hard-won skill, requiring years of reading. But there is a modern shortcut: use multiple programs. Try taking ChatGPT’s output and feeding it to Claude (Anthropic) ; then take Claude’s version and send it to Grok (xAI). Ask each program to explain its reasoning. More amusingly, solicit snide comments from each chatbot about the others. You will get to discernment faster than you think. It may not even be fair to call this a shortcut. In that sense, I am channeling the magician who balks at being called talented for doing a complicated card trick. “I’m not talented, ” he thinks. “I just practiced this a million times. ” And then the magician realizes: maybe practicing a million times is also a talent. Rewriting is a skill with a long and noble tradition. The palimpsest—a papyrus scroll in which a first draft was overwritten with a second—proves this point. Using AI is just a modern gloss on an ancient craft.
J Bernstein (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: