College grades are supposed to signal what students know and can do. A UC Berkeley study now suggests that signal may be getting weaker in the age of ChatGPT, especially in courses built around writing and coding assignments.
The study analyzed more than 500,000 grades at a "large, selective public research university" in Texas and found that grades rose sharply after ChatGPT launched in November 2022. The pattern was strongest where AI tools are most useful: unsupervised writing and coding work.
What the grade data shows
The research tracks grade trends across eight fall semesters, from 2018 through 2025. It covers 319 courses across 84 departments, giving the study a broad view of how grading changed before and after ChatGPT became widely available.
To estimate which courses were most exposed to AI, the study used the assignment mix from fall 2022 syllabi. That timing matters because those syllabi came before ChatGPT existed. The key measure was how much of a course depended on writing and coding tasks, the kinds of work where AI tools perform especially well.
In those high-exposure courses, the share of A grades jumped by 13 percentage points, about 30 percent above the 2022 baseline. Average GPA rose by 0.12 points, and the spread of grades became narrower. In practical terms, some A-minus and B-plus grades appear to be moving into straight A territory.
That shift does not automatically prove that students learned less. Higher grades could reflect better study habits, better teaching, or some other change. But the study looks closely at where the gains appear, and that is where the AI explanation becomes harder to ignore.
Homework is where the effect concentrates
The central question is whether AI is helping students learn or simply helping them produce higher-scoring submissions. Author Igor Chirikov tested that by comparing courses where homework counted more heavily toward the final grade with courses where it counted less.
If AI were broadly improving learning, the grade increase should show up across different assessment formats. Students who learned more with AI help should also perform better when tested in settings where they cannot easily outsource the answer.
Instead, the increase was strongest in courses where homework counted for more than the median share of the grade. In those courses, A's rose by an extra 16 percentage points compared with lower-homework courses with the same AI exposure. In courses below the median for homework weight, the effect was small and not statistically significant.
The study says this result is "difficult to reconcile with broad learning gains or sorting effects alone." A placebo test points the same way: grades did not move for oral presentation assignments, where AI is far less useful.
The implication is direct. When work happens outside supervised settings and depends on writing or coding, AI can raise the quality of what gets submitted without necessarily raising the student's underlying skill by the same amount.
Why this grade inflation is different
Grade inflation itself is not new at US universities. The study notes that at Harvard, the share of A grades rose from 24 percent in 2005 to 60.2 percent in 2025. Earlier explanations included teaching evaluations that reward leniency, competition between universities, and institutional grading policies.
AI changes the mechanism. Those earlier forces affected grading after students had completed the work. AI can alter the work before an instructor sees it.
That difference matters because grades are used outside the classroom. Employers and graduate programs may rely on grades to judge ability, effort, or readiness. If grades in writing- and coding-heavy courses increasingly reflect AI-assisted output, those decisions may become less reliable.
The concern is not only about selection. Chirikov also warns of a feedback loop. If AI performs the very tasks students are meant to practice, graduates could become weaker in the areas where AI is already strong. That could accelerate automation and deepen skill gaps in the job market.
What schools may need to change
The study does not suggest that simply moving everything to proctored exams is enough. It also says that approach is not simple. Instead, the proposed direction is to rethink assessment design.
That could mean assignments that limit AI use, or assignments that deliberately include AI while requiring students to show their process. The source points to documentation of the work process and follow-up interactions that demonstrate understanding as examples.
The distinction is important. A course can ban AI, integrate AI, or do some mix of both. But the assessment has to make clear what is being measured: the final artifact, the student's independent skill, or the student's ability to work with AI while still understanding the material.
The wider education warning
OpenAI CEO Sam Altman has also raised concerns about education's response to AI. In a recent interview, he said that three and a half years after ChatGPT's launch, the education system has barely changed in a systemic way. He had expected a year of cheating followed by an overhaul, but said he cannot identify meaningful systemic change.
Without that change, Altman warned that critical thinking skills risk "significant atrophy." He still believes education can adapt, as it has during earlier technological shifts. But he also argues that skills such as writing and coding should continue to be taught because they train how people think.
Norway has already taken a more restrictive path for younger students. The country recently mostly banned AI tools from elementary schools and limited their use in secondary schools. Prime Minister Jonas Gahr Stoere said uncritical AI use can tempt students to skip important learning steps.
The UC Berkeley findings bring the issue back to higher education. If grades are rising most where AI can quietly do the work, universities face a difficult task: preserve the value of grades while teaching students how to live and work with tools that are not going away.