During the school year, most of my weekends include a stack of student essays. I teach AP World History and AP US History, and the essay pile is the part of my job that AI companies are most eager to take off my hands. One grading company advertises that AI graders can cut grading time from fifteen to thirty minutes per paper down to a few, and another says more than 200,000 teachers use its grader.
Two university studies released this year tested a more careful version of that promise, where AI drafts the feedback and a person decides what the student sees. Neither study took place in a high school, and both used teaching assistants instead of classroom teachers, so I’d treat them as early evidence. Read side by side, they point to where a teacher’s time goes when AI joins the grading.
The Michigan trial
At the University of Michigan, a team led by doctoral student Xinyi Lu and assistant professor Xu Wang built a tool called FeedbackWriter and tested it in an introductory economics course with 354 students and 11 teaching assistants. The tool read each essay, judged whether it met each rubric criterion, and drafted suggested comments. The TAs could accept, edit, or reject any suggestion, or write their own comment instead. Each student wrote two essays, and the researchers swapped the groups between them, so every student got AI-assisted feedback on one essay and human-only feedback on the other.
Students wrote better revisions after the AI-assisted feedback. The effect size was 0.50, which the researchers compare to moving a student from the 50th percentile to the 70th. The AI-assisted comments also covered more of the rubric and were more actionable. The paper was presented this spring at CHI, a major conference on human-computer interaction.
Over the semester, the TAs used FeedbackWriter on 1,366 essays and agreed with the AI’s rubric judgments about 88 percent of the time. They corrected the rest, roughly one call in eight. In a class where nobody reads behind the tool, those mistakes would go straight to students, and if the AI’s call set the grade, into the gradebook too. Lu said the team is “not trying to replace human instructors.”
The Princeton experiment
A second study, posted in June by researchers at Princeton, asked a different question. In a 300-level machine learning course in fall 2025, TAs had to grade homework, but written comments were optional. The researchers randomly assigned submissions so that, after grading, TAs saw an AI-drafted comment on some and nothing extra on others. The study covered 11 TAs and 88 students.
With a draft in front of them, TAs wrote feedback 10.8 percentage points more often, and their comments ran about 40 characters longer. Students rated the AI-assisted comments as useful as the others. The TAs reworked the drafts instead of pasting them in, and the time they spent per character of feedback didn’t drop. The authors concluded that the drafts lowered the barrier to starting a comment without reducing the overall effort of writing one.
Where the hours go
Put the two studies next to each other and they describe something the grading-tool websites leave out. In both, a person read the student’s work, checked or reworked what the machine drafted, and decided what the student would see. At Michigan the AI helped that person cover more of the rubric, and at Princeton it made them more likely to write a comment at all. Neither study measured reading time, but nothing in either design let the TAs skip the reading, and the feedback can’t be right without it.
I wrote about this in Chapter Thirteen of my book, The AI Doesn’t Know Your Students. When I started using AI as a first-draft feedback writer on AP essays, rewriting its comments in my own voice took about a third of the time that writing them from scratch would have. The Princeton TAs didn’t get faster per character, but a draft made them more likely to start a comment in the first place. That fits my experience, where the saving came from reacting to a draft instead of composing from an empty comment box, not from reading the essays any faster. I wouldn’t want any of it to come from the reading.
The two settings also differ in a way that applies to high school. At Michigan, feedback was part of the job, and the AI made it better. At Princeton, feedback was optional, and the AI made it happen more often. Most of us work somewhere in between: we have to grade, we want to comment, and over a full class set it’s easy for the comments near the bottom of the pile to come out shorter than the ones at the top. If the Princeton result carries over to high school, the strongest case for AI drafts may be the student whose essay landed near the bottom, who might otherwise get a score and a sentence.
Both studies also tested feedback that students could still use. The Michigan comments went to students who were about to revise, and the Princeton comments came back on homework. Neither tested AI on a final exam essay or any grade that follows a student into a transcript. In the book I argue that AI is most useful on drafts and practice work, where a student can act on what it says, and riskier on summative judgment.
The corrections and the classroom
I’d want to know more about the kinds of judgments the TAs had to fix. In the book, I described AI flagging problems in student essays that weren’t problems once you understood what the student was trying to do. A tool reading against a rubric sees a thesis that doesn’t match the expected pattern. A teacher who knows the student can sometimes see an argument more ambitious than the rubric anticipated, or a student writing in a second language whose idea is sharper than the sentence carrying it. Many of my students think partly in another language, and that second case comes up often enough that I wouldn’t send a rubric judgment back unread.
My own approach is to split AI feedback into two layers. The surface layer covers grammar, phrasing, and mechanics, and it goes to students labeled as AI feedback, because a comma splice doesn’t need my judgment. The argument layer covers the claim, the evidence, and the reasoning. The AI drafts that section too, and I rewrite it before any student sees it. The Michigan design put a person between the AI and the student on every comment.
The wider research is more cautious. A meta-analysis by Rogers Kaliisa and colleagues, published in 2025 in Educational Psychology, pooled 41 studies covering 4,813 students and found no statistically significant difference in learning between AI-generated and human feedback. Taken with the two experiments, that suggests AI-assisted feedback can hold its own against human feedback and may help, but neither experiment tells us what it does for sixteen-year-olds.
Questions to ask before your school buys a grader
Many schools are being offered AI grading tools this year. The Michigan and Princeton designs give you a way to test any sales pitch. Does a teacher see and approve every comment before a student does? Is the draft easy to edit, or is the workflow built around clicking accept? Does the tool show its rubric judgment separately from its comment, so you can check the call without rereading the whole essay? And is the company promising better feedback or less of your time?
Takeaway for Teachers
Pick the next class set you collect and run a small version of the Princeton experiment on yourself. Write feedback on half the essays the way you normally would. For the other half, start from an AI draft built on your rubric, using a tool your school has approved and with student names removed, and rewrite every comment before it goes back. Keep track of how long each half takes and how many students in each half get more than a score. Then compare the AI’s rubric calls on the second half with your own and count how many you changed. That count tells you how much of the reading you can’t hand off, and the timing tells you whether the drafts are buying you better comments, more comments, or neither.
If you’re thinking through questions like this one — how to use AI on the essay pile without handing over the judgment — my book goes deeper. The AI Doesn’t Know Your Students is available on Amazon and at shouldiuse.ai/book.
David Jacobson is a high school history teacher. He writes about AI, education, and the messy intersection of the two at shouldiuse.ai.
