A new UC Berkeley working paper released this week says A grades have spiked at a large Texas research university since ChatGPT entered students' toolkits. After reviewing more than 500,000 grades, the authors estimate that the share of A's climbed about 13 percentage points, roughly a 30 percent jump compared with a 2022 baseline, with the biggest gains in writing-heavy and coding-heavy courses. The pattern, they argue, lines up most closely with students' growing use of generative AI.
Berkeley analysis and method
The paper, "Artificial Intelligence and Grade Inflation," tracks course grades from 2018 through 2025 and uses a difference-in-differences design that compares courses with more AI-exposed tasks to those with fewer. According to UC Berkeley's Center for Studies in Higher Education, the gains were larger in classes where homework counted more toward the final grade. The author reads that as evidence that AI is substituting for take-home work rather than signaling broad jumps in underlying learning.
What researchers and reporters are saying
The researchers conclude that the most plausible explanation is that students are leaning on generative AI to improve their assignments, not a sudden mass leap in mastery. Coverage of the study has described professors as effectively handing out about 30 percent more A's in AI-exposed courses and fewer A-minuses and B-pluses. As reported by The Dallas Express, the shift is sharpest in courses built around heavy take-home work, where it is easiest for students to offload tasks to tools.
Campuses are changing assessments
Colleges are already tweaking how they grade in response. Faculty are leaning more on in-class tests, dialing back the weight of homework, and bringing back proctoring to shore up trust in transcripts. Princeton's faculty, for example, voted in May to require instructor supervision for all in-person exams beginning July 1, 2026, a move the university framed as a response to easy device-based access to AI during unsupervised assessments, according to The Daily Princetonian.
Policy gaps and mixed K-12 signals
Experts warn that policy has not caught up with practice. A RAND survey found that about one-quarter of teachers reported using AI tools for instructional planning in 2023–24. A broader poll covered by Education Week, drawing on Center for Democracy & Technology data, found that 85 percent of teachers and 86 percent of students had used AI in 2024–25, highlighting both rapid uptake and sizeable measurement differences. As Education Week notes, use is uneven across schools, and district guidance frequently trails far behind what is already happening in classrooms.
Tools can help, but tradeoffs remain
There is also evidence that AI can genuinely speed up learning when used as a tutor. A randomized trial published in Scientific Reports found that an AI-based tutor produced substantially higher post-test scores, with effect sizes ranging from 0.73 to 1.3 standard deviations. At the same time, Gallup polling shows that teachers who use AI weekly save about 5.9 hours per week on average, a productivity boost that muddies the question of whether higher grades mainly reflect stronger learning or simply smarter use of tools.
The Berkeley paper drops a sharp, data-driven marker into a debate that now stretches from classrooms to boardrooms: if top marks can be boosted by tools rather than clear, demonstrable skill, colleges and employers may need to rethink how they certify competence. With the World Economic Forum projecting major AI-linked labor shifts through 2030, many educators argue that the safer bet is to redesign assessments so they prioritize in-person, demonstrable work over unproctored take-homes, as campuses scramble to preserve the signaling power of an A.
