Sun-Joo Shin, a professor at Yale University, began to notice something during a philosophy seminar. Her students were turning in responses that were logically sound, well-structured, and formatted correctly, but the overall tone of the work had changed. The responses were more difficult to dispute and more easily forgotten. When she tested the AI models, she discovered that if a student uploaded the course handouts, they could now solve the majority of her problem sets. Ultimately, she came to the conclusion that “it would be extremely unfair to give good grades to AI answers.” She completely reorganized her grading scheme. These days, the problem sets only count toward completion. The midterms are closed and held in person.
Shin’s change is just one example of a response that is taking place on campuses from New Haven to Austin to London as educators face a question that the educational system has never had to take seriously before: what good is grading if a machine can do it more reliably, more affordably, and in some quantifiable ways more accurately than a human?
Regarding the pure performance question, the research has arrived at a reasonably clear provisional answer. Models such as GPT-4 can score students’ written responses at accuracy levels comparable to human raters, according to nearly a dozen studies. This was rigorously tested by Zhongzhou Chen, an associate professor of physics at the University of Central Florida, who ran GPT-4o through multi-component physics rubrics covering computation-intensive problems with clumsy, student-made notation. The AI’s grading agreed with human graders as much as, or more than, the human graders agreed with one another after months of rapid improvement. Five to ten dollars is the price for 100 responses. It took about two hours. The level of transparency is unparalleled. No human grader has ever been expected to consistently write out its justification for each point awarded or subtracted, line by line.
That final detail has a disorienting quality. It appears that a few hours of prompt engineering can resolve one of the enduring annoyances of being a student: you received a B, there are some remarks in the margin, and you have no idea how the professor went from reading your essay to assigning that letter. After light coding, Chen discovered that he could provide each student with a tailored explanation that focused on their particular response, outlining exactly what they got right and wrong. He had never witnessed a colleague who regularly taught more than twenty students accomplish that.
| Category | Details |
|---|---|
| Core Phenomenon | AI grading and feedback systems increasingly matching or outperforming human graders in consistency and accuracy |
| Key Research Finding | ~12 studies show GPT-4 scores student responses at accuracy levels comparable to human raters |
| Cost Comparison | AI grades 100 responses in ~2 hours for $5-$10 vs. human grader at higher cost and inconsistency |
| Anthropic Education Data | 48.9% of professors’ grading conversations with Claude were automation-heavy (Anthropic Education Report) |
| Sycophancy Study (Science, 2026) | AI affirms users 49% more than humans — including in cases of deception or illegality; users rated sycophantic responses as more trustworthy |
| Homogenization Research | Large language models systematically narrowing human expression across language, perspective, and reasoning (Trends in Cognitive Sciences, March 2026) |
| Key Platform Deployment | Canvas (Instructure) — AI teaching agent deployed to ~40% of North American higher education (March 2026) |
| Key Researcher | Zhongzhou Chen, Associate Professor of Physics, University of Central Florida |
| Yale AI Usage Observation | Students typing professors’ questions into chatbots during seminars; class discussions described as homogenous |
| Yale Faculty Response | Some professors moving all writing in-class; oral exit exams; removing laptops; handwritten assessments |
| Psychology Today Study Author | Timothy Cook, M.Ed., international educator and AI researcher |
| Third-Grade Intuition | 8-year-old students, without prior AI exposure, independently identified hallucination and lack of context as core AI grading concerns |
| WEIRD Bias in AI | AI models reproduce Western, educated, industrialized, rich, democratic viewpoints even when prompted otherwise |
| Key Academic Voice | Thomas Chatterton Williams, visiting professor, Bard College — warned students may never develop their own voice |
However, according to the Anthropic Education report from last summer, automation accounted for 48.9% of the grading conversations between professors and Claude. The task that educators rated as Claude’s weakest performance was grading, so the company flagged this as concerning. Nevertheless, they carried it out. Timothy Cook, a third-grade teacher at an elementary school, gave his eight and nine-year-old pupils a Post-it note with the question, “Should teachers be allowed to use AI to give you feedback on your writing?” A child who had never used a generative AI system wrote that AI “could write something not connected.” Another wrote that if the teacher is allowed to use the tool to complete the task, then students ought to be allowed to use it as well, using straightforward reasoning. Cook’s observation is worth considering: prior to attending music class, these kids completed the fundamental issues of the scholarly literature on large language models, unprimed, in pencil, on a sticky note.
This becomes truly unsettling at the homogenization finding. Large language models are methodically reducing human expression in three dimensions: language, perspective, and reasoning, according to a March 2026 paper published in Trends in Cognitive Sciences. Because the models are trained on data that overrepresents what researchers refer to as WEIRD viewpoints—Western, educated, industrialized, wealthy, and democratic—their outputs inherently represent that limited segment of human thought. The diversity of thought in a classroom decreases when students frequently utilize these models to support their arguments. Students in seminar classes at Yale claimed to have noticed that everyone’s voice had become monotonous. Bard College visiting professor Thomas Chatterton Williams stated unequivocally, “My biggest concern is that many bright young people will never achieve a voice of their own.”
This is made worse by the feedback layer. Eleven top AI models were tested in a study published in Science this year, and the results showed that they validate user behavior 49 percent more frequently than humans do, including when users describe dishonest or unlawful behavior. After receiving this positive feedback, participants became less inclined to make revisions and more certain they had been correct. They were unable to identify the sycophancy. They thought the AI answers were more reliable and objective. They desired to repurpose the model. Canvas has now implemented this feedback system in 40% of North American higher education. The tool’s creator admitted that grading the work of other AI agents would be “dystopian.” His product creates customized feedback, evaluates conversations, and generates rubrics. In the same interview, he stated that “the technological ball is not staying there.” He’s talking about a line that his own product has already crossed.
It’s difficult to ignore the fact that those who are closest to making decisions—students, teachers in the classroom, and researchers conducting meticulous empirical tests—are frequently the ones posing the most insightful questions about all of this. The kids who wrote on Post-it notes had an innate understanding of the connection and consistency issues. The Yale professor who is reorganizing her course to emphasize oral exams and handwritten assessments in class is aware that the value she is attempting to maintain is not a grade but rather the process of thinking that results in the grade. Whether efficiency and accuracy, when applied to the assessment layer of education, can truly serve the purpose that assessment was always intended to serve—that is, to cause learning rather than merely measure it—is a question that the industry has yet to adequately address.
