Examining Grade Inflation in the ChatGPT Era
Introduction
Since the release of ChatGPT on November 30, 2022 - there is evidence that undergraduate student grade inflation is accelerating, potentially from the use of LLMs by students and instructors. This analysis will examine the following student and course attributes that contribute to this accelerated inflation: first generation status, freshman status, small course size, race and ethnicity, relative high school GPA, and 100 level courses.
Local Interest in LLMs
Google search results for “chatgpt” suggest the immediate interest in ChatGPT following its release within the Champaign-Urbana metro area. There are sharp declines in local interest when semesters end and when spring and fall breaks begin. Likewise, there are sharp increases in local interest when semesters begin and breaks end.
Accelerated Grade Inflation
Figure 1 shows that accelerated grade inflation occurred during the time of increased local interest in ChatGPT. The solid line in the chart below summarizes 3.6 million student grades from 2017-2025. The dotted line is the pre-COVID (2017-2019) trend in grade inflation, extrapolated through 2025.
Various COVID-specific policy changes and educational changes resulted in an unprecedented inflation of grades in 2020 and 2021. However, campus-wide grades immediately returned to pre-COVID levels in 2022. Finally, the grades exhibit a “shock and accelerate” pattern following the adoption of Generative AI in 2023 and on.
The 2025 average GPA of 3.55 is higher than the projected 3.47 average GPA expected from extrapolating the pre-COVID grade inflation trend to 2025. The LLM-era grade inflation rate of 0.037 grade points per year is 62% higher than the pre-COVID grade inflation rate of 0.023 grade points per year.
Subsequent sections provide more detailed breakdowns by specific student and course attributes.
Detailed Breakdowns by Student and Course Attributes
The following charts explore how these grade trends differ across various student populations.
Trends by High School GPA
This chart shows the grade trends for students based on their prior academic performance (high school GPA quartile). Most students (but not those in the top quartile) show accelerated grade inflation in the LLM-era relative to the pre-COVID inflation rate.
I use a difference-in-differences model to compare the change in grades between the pre-COVID (2017-2019) and post-LLM (2023-2025) eras for students in different high school GPA cohorts. While grades for all students increased significantly, this grade increase was largest for the bottom GPA quartile (+0.244 points) and progressively smaller for the interquartile (+0.223 points) and top quartile (+0.167 points) groups (all p<0.001).
Perhaps a ceiling effect diminishes the effects of LLMs for high-achieving students; or perhaps instructors grading with LLMs results in easier grading standards that disproportionately benefit lower-achieving students.
Trends by First-Generation Status
The average grade gap between first-generation students and the remaining students is shrinking in the LLM era. This suggests that the effects of LLMs on student grades are potentially stronger for first-generation students. However, the difference-in-differences model shows that grade increase for URM students (+0.237 points) relative to non-URM students (+0.191 points) had borderline statistical significance (p = 0.053). Since instructors are typically unaware of a student’s first generation status, this potential difference in grade inflation is unlikely to be the result of changes in grading behavior or course design.
Trends by Course Level
Before COVID, 100 level courses had a higher average grade than 200-500 level courses. During COVID, this gap disappeared. In the LLM era, both 100 level and higher level courses exhibit a shock and accelerate pattern of inflation. However, 100 level course have accelerated more than higher level courses in the last two years. The difference-in-differences model confirms this difference (p < 0.001).
There are a wide variety of course differences that may explain these differences, including but not limited to: differences in subject complexity, difference in academic skills of students enrolled, differences in course size, and differences in grading methods.
Trends by Course Size
There is accelerated grade inflation for all three sizes in the LLM era, although the inflation acceleration for large courses was small and wanes by 2025. The inflation for medium courses is also waning in 2025. Conversely, the acceleration is large and continues getting larger for small courses. The difference-in-differences model confirms this difference (p < 0.001).
Trends by Class Level
There is accelerated grade inflation for all calculated class levels in the LLM era. This acceleration is largest for calculated freshmen. This comports with the observation that 100 level courses have the most acceleration.The difference-in-differences model confirms freshmen have the largest grade inflation in the LLM era (p < 0.001). As a result, the gap between calculated freshmen and seniors is shrinking during the LLM era.
Trends by Underrepresented Status
I estimated a simple difference-in-differences linear regression model to compare the change in grades between the pre-COVID (2017-2019) and post-LLM (2023-2025) eras for URM and non-URM students. The results show that grade increase for URM students was larger (+0.237 points) than the increase for non-URM students (+0.191 points) in the LLM era (p<0.001).
Modeling these Attributes Together
A series of multilevel interrupted time series (ITS) regression models were fit to formally test the previous observations. Grades were nested by student and department. Though the full model outputs are omitted for brevity, their findings confirm the visual patterns above.
The estimated effects all of attributes were statistically significant with a sample of 3.6 million student grades. The “shock and accelerate” pattern in grades post-2022 persist after accounting for class size, class level, race and ethnicity, HS GPA quartile, first generation status, and course level. Moreover, the model supports the previous major finding: the lack of a grade inflation shock for top performing students. Also, students with a bottom quartile high school GPA had the largest grade inflation shock in the LLM era. Finally, the model quantifies the meaningful impact of both students and departments on grades; while differences between departments create significant variation, the variation between individual students is over five times larger.
Additional models, in which grades are nested by student and instructor, also confirm significant variance at the instructor level. These suggest that instructor-specific factors are a key component in explaining the overall trend. However, the most theoretically complex model specifications with instructor random effects failed to converge given the limitation of the data structure. In short, there was an insufficient number of student-instructor pairings to power these most complicated specifications.
Nevertheless, the instructor random intercept models still provide insight. While the student differences account for roughly twice as much of the variation in grades as the differences between instructors, this cannot be confidently interpreted to mean student behavior is the larger contributor to the accelerated grade inflation. Differences in grading may be reflected in student differences.
Case Study of LLM-Resistant Course Design
Dr. Challen teaches large 100-level computer science courses, with a cumulative enrollment of over 15,000 students since 2018. His courses provide a robust case study that exemplifies the instructor-level variance identified in the statistical models.
As the chart below illustrates, Dr. Challen’s courses do not exhibit the same signs of accelerated grade inflation seen campus-wide. His reliance on the Computer-Based Testing Facility (an in-person, proctored environment that restricts internet access) appears to insulate his assessments from the direct influence of LLMs. Dr. Challen’s lack of LLM grade inflation is especially noteworthy since 100 level courses with high freshmen enrollment were highly susceptible to LMM effects across campus.
Interestingly, the breakdown by high school GPA cohort within Dr. Challen’s LLM-resistant courses reveals a slight but statistically significant decline in grades for top-quartile and interquartile students during the LLM era. Because Dr. Challen has made small modifications to his courses from 2023-2025, this decline could be attributed to changes in his courses. However, this decline could also suggest a worrying possibility: that the widespread use of LLMs could be leading to an atrophy of critical academic skills, which becomes apparent only in proctored assessment environments. This potential atrophy, if real, may start occuring in high school.
Final Thoughts
The evidence presented suggests that the LLM-era acceleration in grade inflation is driven by student and course attributes. First generation status, freshman status, small course size, race and ethnicity, relative high school GPA, and 100 level courses are all identified as contributors of the LLM era’s accelerated grade inflation. And yet Dr. Challen’s case reveals how course design can overcome these drivers of LLM inflation.
Importantly, the etiology of the identified contributors to LLM inflation is unclear. These contributors could be the result of students completing coursework with LLMs; conversely, some could be the result of instructors grading coursework with LLMs.
Furthermore, this accelerated grade inflation itself is ambiguous. It may reflect authentic improvements in educational quality or simple efficiency gains from LLM usage. Conversely, as the case study hints, there is the possibility it could mask a degradation of foundational academic skills starting in high school. An analysis of final grade data alone does not provide the necessary leverage to distinguish between these possibilities. These complexities will be difficult, and perhaps impossible, to disentangle confidently in further research.