Choosing to Stay Human

An essay in One Useful Thing argues that people should make conscious choices about when to use AI rather than defaulting to it. The author writes that AI-generated writing has become pervasive on social media, in academic papers, opinion articles and short stories, and says it can offer little meaning per word while consuming readers' attention. The author, who has written for decades, says developing a personal style required effort from teachers, rewrites and critical comments, and that delegating writing to AI risks giving up something important to their career and happiness. The essay does not reject AI writing help, citing uses such as checking drafts and roleplaying reader perspectives, but condemns using AI as a default or without thinking.

Two papers with an overlapping research team, including peers at Wharton, are cited. In one experiment at a high school in Turkey with about a thousand students learning math, a group using plain ChatGPT did homework better and believed they learned more, but underperformed classmates without ChatGPT on tests. In a second paper, many of the same authors ran a five-month Python course across ten high schools in Taipei with close to a thousand students. Students given a personalized sequence of problems by an AI tutor scored 0.15 standard deviations higher on a final exam taken without AI help, which the essay says some estimates equate to six to nine months of additional schooling without added instruction time or teacher workload.

The essay describes tools for tutor-style use: Gemini's Guided Learning, typing "/learn" into ChatGPT, and selecting a learning style in Claude, noting Anthropic has announced the approach is changing but has not documented the change. It cites a Wharton term, "cognitive surrender," and an experiment with 758 Boston Consulting Group consultants, half given GPT-4, in which AI users vastly outperformed others but were significantly less likely to solve a problem the AI failed at. It also cites a small Anthropic study in which programmers who let AI do the work could not answer questions about what they had done, while those asking the AI to explain or using it for only part of the work seemed to avoid that outcome.