A UC Berkeley math professor is rethinking homework grading now that students can use AI

The role of homework in education is undergoing a fundamental reconsideration as AI tools become capable of solving complex problems instantaneously. This shift raises deeper questions about what homework is actually for, how learning happens, and what assessment can tell us about understanding. What homework is supposed to do Homework exists in modern education for several interrelated purposes. The traditional view holds that repeated practice reinforces skills and builds fluency. A student solving twenty algebra problems is assumed to internalize the mechanics of factoring or solving for x through repetition. Beyond drill, homework also serves diagnostic functions: instructors review completed work to identify where individual students are struggling, which shapes subsequent lessons. There is also a behavioral or motivational component: homework extends learning beyond class time and can signal to students that mastery requires effort and engagement outside formal instruction. The older question about homework's necessity Even before AI entered classrooms, educational researchers questioned whether homework actually delivers on these promises at scale. Meta-analyses of homework's effects have historically shown modest correlations between homework completion and achievement, with significant variation by grade level and subject. High school students show stronger correlations between homework and performance than elementary students, and some studies suggest diminishing returns: more homework does not necessarily produce proportionally better outcomes. The quality and design of assignments matter far more than their sheer quantity. These findings existed for decades, yet homework remained deeply embedded in educational practice. How AI changes the equation An AI tool that can solve a math problem in seconds transforms the nature of the assignment itself. If the purpose was fluency through repetition, and the student receives a solution instantly, then the homework cannot accomplish that pedagogical goal. If homework existed to provide the instructor with diagnostic information about student understanding, an AI generated response provides no such information about that particular student's thinking. This represents a genuine structural problem, not merely an enforcement challenge. The instructor cannot simply catch cheating and move on; the assignment's core justification has shifted beneath them. Understanding versus procedural completion This tension highlights a long standing but often obscured question in mathematics education: the difference between solving a problem and understanding why the solution works. Early mathematics education emphasizes procedural fluency; students learn steps and apply them correctly. More advanced mathematics requires conceptual understanding: grasping why a method works, how it connects to other concepts, when it applies and when it fails. Homework designed for procedural practice looks very different