When people write about A.I. and universities, they’re usually talking about students: “What Happens After A.I. Destroys College Writing?” asked a headline in The New Yorker. “How A.I. Killed Student Writing (and Revived It),” declared one in The New York Times. “Members of the class of 2026 have had access to A.I. since they were freshmen,” announced an Atlantic header: “Almost all of them are using it to do their work.” These clarion calls ignore a major irony: the graduate students, postdocs, and professors teaching those undergraduates are themselves increasingly using A.I. to do their work.

I finished my PhD in literature a couple of years ago, and in the time since, I’ve seen a major shift in the scholarly use of A.I. The horrified rhetoric around undergraduate cheating remains the same, but it’s now accompanied by widespread adoption for writing and research. All I have are anecdata, but among early-career humanists, I’d say moderate usage is the new norm. In the sciences it ranges from high to constant. 

Last year, a colleague in philosophy told me it had revolutionized his writing process to have a form of intelligence to spar with; he now used ChatGPT almost every day. Another in physics described how ChatGPT had doubled his research output. An art historian praised Claude’s transcription accuracy and ability to structure her workflow; upon receiving a “revise and resubmit” from a journal, she gave Claude the feedback and asked for a twelve-day program to tackle the revisions.

We’re also seeing the start of A.I.’s sanctioned incorporation into the research pipeline. A colleague in economics recently had a paper accepted at a top field journal and was informed she could participate in the journal’s A.I. pilot. The program, Refine, was designed specifically for peer review and would provide feedback on the accepted paper. My colleague agreed and was impressed: Refine caught notational errors and inconsistencies that she and her three peer reviewers had missed. No wonder that when the teaching center at my institution offered A.I. training intensives for faculty, some hundred enrolled.

I’m an A.I. skeptic. For years, my friends in industry have been telling me to offload research and administrative tasks to LLMs. I responded by opening an Apple note called “What A.I. Can’t Do.” But as the university’s rhetoric came to ape that of Silicon Valley, I started experimenting.

I’ve seen the best results with what I’ll call “semantic spellcheck”: uploading near-finalized documents and asking Claude to note any minor errors. I started with my C.V., which, I learned, included a recent conference that took place February 27–29, 2026 — and 2026 is not a leap year. I then tried the proofs for a short book I had translated and was finalizing. Again, Claude caught things I’d missed: the French coastal town Boulogne-sur-Mer had become the wall-perched “Boulogne-sur-Mur,” and I’d described Guido Reni in one place as a Baroque, and elsewhere as a Renaissance painter. I was grateful and a little embarrassed. But Claude also missed a couple dozen more serious errors my own human eyes had already identified, among them: “canon shots,” “a women,” and the real howler, “Middles Ages.” Unlike a standard spellcheck, A.I. can look for issues on the level of meaning and logic rather than just orthography and grammar; like any spellcheck, it clearly requires supervision.

I’ve been less impressed with what LLMs do with Henry James. Over the past couple years, I’ve tried out three sorts of questions on Claude and ChatGPT: the generically bibliographic (Give me ten secondary sources on Henry James and sexuality); the targeted but broad (What do you think about the relation between sexuality and Italian in James’s novel The Golden Bowl?); and the factual (How far into The Golden Bowl do two of the protagonists, Charlotte and Adam, marry?). 

The first response gave half good citations, half bad or hallucinated ones. The second suggested some astute connections — or at least connections similar to those I’d made! — but then stated that the titular golden bowl was purchased in Italy. (In the text, the London-based characters actually just stroll over to Bloomsbury.) The third said it wasn’t sure but probably in the second half, when anyone who’s read even a plot summary of the novel would know it’s in the first half, about a quarter of the way through.

When I told colleagues about these experiments, some judged them unfair. They claimed that such tests didn’t assess deep thinking and relied on non-premium, non-customized LLMs. I suppose I could build a Henry James bot — gorged on Project Gutenberg novels and my own library of criticism. But what’s the point? I don’t mind an augmented spellcheck, but I’d rather train my own intelligence than an artificial one.