Grade inflation much higher in ‘AI-exposed’ degrees

Degrees should be reviewed for their “exposure to artificial intelligence”, according to a researcher after his analysis showed certain courses have seen the number of A grades rise by nearly a third since the arrival of ChatGPT.

Drawing on publicly available data from a large research university in Texas, Igor Chirikov, a senior researcher at the University of California, Berkeley, examined the marks awarded to more than 500,000 students between 2018 and 2025.

When these grade patterns were compared against syllabus data on the types of writing tasks used for assessment, it revealed the share of A grades in “AI exposed” courses rose by 13 percentage points, or 30 per cent, compared with the 2022 baseline.

Overall grade point average rose by 0.12 points for “high-homework” courses in which AI could potentially complete the assessment, says the study, which was published as a working paper by Berkeley’s Center for Studies in Higher Education.

Grade inflation occurred only in homework-based writing and coding tasks and was not found to the same extent in in-person examination, explains the study, which suggests the computing power of “AI [is] substituting for student effort specifically on the unsupervised assessments where instructors cannot observe the production of submitted work”.

With AI “displacing” student skills, there is a risk that employers will increasingly distrust the grades awarded to students, explains the paper on how “grades [may soon] become less informative measures of student capability”.

However, Chirikov cautioned against a wholesale move back to in-person written exams to create “AI-proof assessment”.

“Some controlled assessments, like oral exams or in-class writing, can be useful when we need to verify independent performance. But they should not be a general solution,” he told Times Higher Education.

“However, many important skills cannot be assessed well in a short, controlled setting,” Chirikov explained.

“For example, when you write a research paper or develop an app as a class project, this often requires sustained work over time. If we move everything back to traditional exams, we may reduce AI misuse but also narrow not only what we are actually measuring but also what we are teaching,” he said.

Redesigning assessment to remove AI cheating is also “difficult and time-consuming” for academics because good practice “looks different in different disciplines”, Chirikov continued.

“With AI, instructors need to balance between at least three elements – assuring learning, protecting academic integrity, and preparing students for a world where AI will be part of professional practice,” he said.

Even monitoring grades over time to check for the potential influence of generative AI could prove difficult, Chirikov said, noting that “grades were imperfect indicators of learning even before AI, because they reflect many things, such as student ability, effort, grading standards, course difficulty”.

“My analysis shows that AI can further distort grades in a predictable way, when AI output substitutes for student work. But I would not make grade monitoring or grade quotas the main solution. If grades go up, that does not automatically mean something is wrong; and if grades stay flat, that does not mean assessment is valid,” he said.

A better approach is “to review courses by their exposure to AI and what assessments are actually measuring”, Chirikov said.

“This can be done at the course or department level. Instructors can explore which assignments are most vulnerable to AI substitution, which learning outcomes require independent performance, where should AI be restricted, and where should it be incorporated deliberately,” he said.

“Institutions and departments can track progress in assessment reform over time,” said Chirikov, adding that “the ultimate goal is to make sure courses are producing credible evidence of what students know and can do”.

jack.grove@timeshighereducation.com