Trading the Grading Grind for Human Connection: A Berkeley Case Study

## Overview Instructor: Dr. Joshua Grossman, Assistant Teaching Professor, Department of Statistics, University of California, Berkeley At the University of California at Berkeley, Data 100 is more than just a class. It is a massive operation that serves between 800 and 1,200 students every single semester. To keep a course of that size running, Professor Josh Grossman and his co-instructor Ramesh Sridharan require a teaching staff of up to fifty TAs each semester. For Josh, the challenge was never just about getting through the semester. It was about how to use those 50 highly educated teaching assistants in a way that actually helped students learn, rather than just forcing them to act as manual data entry clerks. ## The Problem with the Grading Bottleneck Josh came to Berkeley with a background in product design, having worked in educational technology before starting graduate school. That experience gave him a healthy skepticism of tools that claim to "teach" students. Like many faculty, Josh believes that the most valuable part of education is human interaction. However, when you have a thousand students turning in written work and needing actionable feedback, that human interaction is often the first thing that teaching staff has to sacrifice. The TAs usually end up spending significant time grinding through stacks of student submissions. It is repetitive and exhausting mental labor. In a direct tracking of one assignment, it took the staff 65 hours to finish grading without help from AI. That is 65 hours in a single week where some of the brightest minds at Berkeley were looking at screens instead of talking to students. ## The Pensive Reinvestment When Josh integrated Pensive, that 65-hour grading window dropped to an estimated 27 hours. That is a 60% reduction in the time TAs spent on grading. When you look at the math, the impact on the student body is massive. In the world of private data science tutoring, students often pay \$60 or more an hour for 1-on-1 help. By saving 38 hours of TA labor on a single exam, the course is essentially generating \$2,280+ in instructional value. With multiple exams and more than two dozen assignments, that adds up to a lot of hours of grading time turned into high-value tutoring time. Instead of those TAs being stuck in a room grading, they are now available for extra office hours and specialized tutoring sessions. The students are getting thousands of dollars worth of professional support that was previously buried in administrative overhead. ## Closing the Gap on Grading Equity Efficiency was the initial goal, but the most profound shift happened around the idea of "grading equity." Josh conducted a study where he compared human-assigned versus Pensive-assigned grades for 100 randomly selected student submissions. In the case on one problem, in the 12 cases where the human grader and the AI disagreed, the AI was correct every single time. This is where the "Confidence Tax" comes into play. When a human grader is tired, distracted or hungry, they make mistakes. Usually, it is the most confident or vocal students who notice the error and file a regrade request. Students who are less sure of themselves, or those from underrepresented backgrounds, often assume they were the ones who made the mistake. They take the lower grade and move on. By using a system that stays highly accurate regardless of how many papers it has seen, Josh is helping to mitigate that bias. Students get the credit they actually earned, which levels the playing field for those who might be too intimidated to argue with a TA over a few points. ## A New Type of Transparency Josh is completely open with his students about this. He tells them on the first day of class that they are using AI grading as a tool to improve efficiency. He explains that the goal is to make the staff more available for them. Because the AI does not get tired, the team has also been able to change how they write rubrics, including much more detailed feedback grading items. This has turned the grading process from a stressful evaluation into a supportive feedback loop. By automating the mechanical parts of the course, Data 100 has actually managed to become a more human experience.