
We should judge analytical writing by its quality, not its source. A professional journal recently accepted an article of mine. The editor congratulated me on the acceptance and then made an unusual request. Could I submit another draft written “in my voice,” since this one had some sentences that “read like something written by AI”?
If I preferred, the editors would revise the piece themselves, ostensibly to “humanize” it, and I could approve their changes. Because I was first trained as a mathematician, I probably write more like an AI than most people. But that particular article had been through 10 drafts. I worked on it with Claude Fable 5.1 and other large language models, as I now work on nearly everything, but I reviewed every line and rewrote most of them.
By draft 10, I could not have said where the machine’s contribution stopped and mine began. The editor had no quarrel with the argument, the evidence, or the structure; she applauded all three. The complaint was about the voice, and the remedy offered was to have strangers rewrite my sentences so that they would sound more like me. This amusing episode helped me to understand a confusion that now runs through the academy.
We have begun to judge analytical writing not by its content but by its suspected origin, which no reader can verify. Because an analytical essay is an argument, and an argument can be evaluated, professional analytical writing should be judged as it always has been: by its quality. It helps to ask what each kind of written artifact is for. When OpenAI recently announced a machine-checked proof of a famous unsolved mathematical problem, the philosophers Silvia De Toffoli and Eamon Duede did not dwell on who deserved the credit.
They asked whether mathematicians could understand the proof and build on it, since understanding is what a proof is for. A machine-certified answer that teaches human mathematicians nothing valuable falls short. Writing can be evaluated along similar lines, by its teleology, or ultimate purpose. A literary work is meant to be an extended encounter with a unique human sensibility and the choices it inspires.
A student essay is meant to demonstrate the student’s learning and mastery of a given subject. A scholarly argument is meant for readers to understand, contest, and/or deploy. The literary establishment to date is not thinking along these lines. Days before my article was accepted, the Académie Goncourt removed a novel by the Haitian Canadian writer Thélyson Orélien from its longlist after an anonymous account on X ran passages through a detection tool and declared the book machine-written.
Orélien denies using AI. The Académie also cited plagiarism in his earlier work, a charge that can be settled by laying two texts side by side. The AI charge cannot: When Agence France-Presse tested excerpts on several detectors, the verdicts ran from very likely human to 100 percent AI. In a novel, the importance of voice is inescapable.
I make no claims here about creative writing, which I expect to remain safely human for the foreseeable future. But surely, it means something that so many readers and prize juries found Orélien’s creation exceptional. Analytical writing, in contrast, is largely a creature of logic and structure. For more than three decades, I have handed my political science students a three-page guide to analytical writing I obtained when I was a graduate student teaching writing in Harvard’s Government department.
It opens: “An analytical essay is a sustained attempt to prove a point: It is an argument.” It needs to have a thesis, and every paragraph must advance that thesis, since “no reader can long tolerate unemployed sentences and paragraphs.” Analytical writers need to provide evidence (“if you want to say it, prove it”) and anticipate the strongest possible objections. Nothing in the guide specifies how the essay must be produced.
Every test in it can be applied by a reader who knows nothing of the road the author took to get there. As a graduate writing tutor, I often encountered what I then called “Andover/Exeter b.s.” (Today I might call it “Andover/Exeter slop.”) The term referred to an essay that was highly polished, had the appearance of substance, and upon closer inspection said nothing at all. My response to this kind of writing at the time was not to ask students where they had gone to school but what their thesis was.
Bad writing can be measured on a common yardstick irrespective of its origins. An invented or missing citation fails the test of evidence. A smooth, empty paragraph fails the test of employment. As always, the author who signs the work must answer for it.
Why, then, does the argument over writing and AI generate so much heat? Much of it results from a mismatch of experience. Those who have worked with a subscription frontier AI model for hundreds of hours and those who have dabbled with a free one are not describing the same thing. Sustained use teaches that AI is alien intelligence, one whose strengths and weaknesses differ from human ones.
The model is formidable at logic, but it has no sense of where it stands in space and time, and so its narrative instinct is at best weak and at worst absurd. That is why the model needs far more instruction to write than to code. I deploy it for its strengths and am constantly on alert for its weaknesses, which is what makes it a tool and not a ghostwriter. How we educate our students in this brave new world is a question unto itself.
The student essay was designed to provide evidence of learning, which it no longer does; a novice writer cannot responsibly direct a tool whose output they cannot yet judge. Teaching therefore requires careful scaffolding, which can be constructed in myriad ways, including a ban on AI use. I have chosen to teach with AI, and that has meant replacing my old assignments with harder ones. We work over texts carefully together in class, so that students learn to read closely and find their own point of view.
I insist on rigorous citation in both informal and formal writing. I deploy the Socratic method in class discussions, which reveals whether students actually know the material or have simply fed my study questions to a machine. Because AI speeds up the routine work, I can assign more of it, in more varied forms: research, outlines of arguments, revisions of their own prose, oral presentations, and slide decks. Unannounced in-class quizzes can tell me what students have retained from the day’s reading.
And I have long offered an optional oral examination — the Oxbridge tutorial method — in place of the take-home final. It remains a choice, because no single format is fair to every student. For all forms of writing, readers are owed a human being who stands behind the work. A byline has always signaled ethical and legal responsibility.
But it has never been a certificate of unassisted composition, as anyone who has ever thanked a human editor knows. The Chronicle’s freelancer policy asks that submissions be “written by you, not AI” (with rare exceptions). I wrote this essay the way I wrote the article that inspired it, and I invite its editors and readers to judge it accordingly. My voice, it turns out, was never missing from the draft that journal accepted.
It was expressed in what I cut and what I rewrote across 10 full revisions. Another revision will follow from the editor’s guidance. We already know how to distinguish good analytical writing from bad, and we have been teaching the method for decades. The standard does not change because the tools have.
The author drafted this essay in extended collaboration with Claude Fable 5.1.
Texto de chronicle.com tal como lo publicó, leído de su página y mostrado aquí con el formato de El Autómata. El original, en chronicle.com · Volver a la noticia de El Autómata