UC Berkeley Math Professor Drops Online Tests and Homework as Generative AI Use Spikes

The intersection of higher education and artificial intelligence has reached a critical inflection point, forcing major academic institutions to rethink foundational methods of student evaluation. Richard Borcherds, a distinguished professor of mathematics at the University of California, Berkeley (UC Berkeley), has announced that he has entirely eliminated graded online homework and remote tests from his curriculum. The decision stems from the proliferation and sophistication of generative artificial intelligence tools, which Borcherds argues have rendered unsupervised take-home assignments obsolete and counterproductive to genuine learning.

Borcherds, a renowned mathematician and Fields Medalist recognized globally for his contributions to algebra and mathematical physics, stated that distinguishing between original student work and text or code generated by advanced large language models (LLMs) has become virtually impossible. Consequently, the educator has pivoted exclusively to traditional, proctored in-person examinations where electronic devices are strictly prohibited, ensuring that student performance metrics accurately reflect individual comprehension.

The shift highlights a growing dilemma within global academia. As generative artificial intelligence models become increasingly accessible, free, and capable of solving complex mathematical equations, writing code, and composing analytical essays within seconds, universities face profound challenges in academic integrity. Borcherds’ decision underscores a broader, systemic reassessment of how professors assign work, grade performance, and maintain standards of excellence in an era where automated assistance is ubiquitous.

The Evolution of the Academic Integrity Crisis

The integration of generative artificial intelligence into everyday student workflows has accelerated dramatically since the late 2022 public release of OpenAI’s ChatGPT. Initially, academic institutions responded with widespread bans, panicked warnings about plagiarism, and the rapid adoption of AI detection software. However, these early mitigation efforts quickly proved ineffective. AI detectors have repeatedly been shown to produce high rates of false positives and false negatives, making them unreliable tools for disciplinary action. Furthermore, as models evolved to handle complex symbolic logic, advanced calculus, and multivariable math, students found themselves equipped with tireless, 24/7 digital tutors capable of executing entire problem sets instantaneously.

For mathematics departments, the challenge is particularly acute. Modern AI models are no longer limited to basic arithmetic or algebra; they can process complex proofs, solve differential equations, and explain multi-step logical derivations with fluency. When students are given days to complete take-home assignments, the temptation to bypass the cognitive struggle of problem-solving in favor of algorithmic shortcuts is immense.

Borcherds noted that the quantitative indicators of this shift are unmistakable. While individual student capabilities vary widely, online assignment averages have experienced sudden, anomalous spikes that do not align with historical performance distributions or subsequent in-person test results. This divergence exposes a systemic flaw in unsupervised digital assessment: remote homework no longer measures what a student knows, but rather how effectively a student can prompt an artificial intelligence system.

Inside the Classroom Policy Shift

The decision by the UC Berkeley professor to strip away online testing and graded homework was not made lightly, but out of logistical and ethical necessity. According to statements provided to media outlets, Borcherds emphasized that continuing the status quo would actively penalize academic honesty.

"Online exams are now a waste of time as several students just use AI to answer the questions," Borcherds explained. "I cannot tell which students are using AI, but I notice the average score increases a lot in online exams."

Beyond the integrity issue, Borcherds pointed to the perverse economic and pedagogical incentives created by traditional grading models in the age of automation. When assignments are graded strictly for correctness without real-time verification of the student’s thought process, the system inadvertently encourages cheating. Students who spend hours struggling through difficult mathematical proofs on their own may receive lower scores than peers who simply feed the prompt into an AI model.

Furthermore, the professor highlighted the administrative burden placed on universities. Allocating academic resources, graduate student instructor (GSI) hours, and institutional funding to grade assignments that may have been entirely generated by software represents an inefficient use of academic capital. By reallocating those resources and shifting the evaluation burden entirely to secure, proctored environments, the department can better assess true mastery of the subject matter.

Broader Implications for Higher Education

The move by a faculty member at an elite institution like UC Berkeley serves as a bellwether for the rest of the academic community. Higher education is currently divided into distinct camps regarding how to handle generative artificial intelligence. Some educators advocate for fully embracing AI, redesigning curricula to teach students how to prompt, collaborate with, and critically evaluate machine-generated outputs. Others argue for a return to classical pedagogical methods, emphasizing offline, handwritten, and face-to-face evaluations to preserve fundamental cognitive development.

The implications of Borcherds’ policy touch several key areas of academic administration and student development:

• The Decline of Formative Assessment: Homework has traditionally served as low-stakes formative assessment—a way for students to practice concepts, make mistakes, and learn from feedback before facing high-stakes exams. If take-home practice is outsourced to AI, students lose the vital cognitive friction necessary for long-term memory consolidation and skill acquisition.
• The Return to Analog Evaluation: Across STEM fields, universities are witnessing a renaissance of blue books, handwritten exams, and oral defenses. Institutions are finding that the only reliable way to measure individual human capability is to remove digital interfaces entirely from the examination process.
• Equity and Access Concerns: Relying entirely on high-stakes, in-person exams places immense pressure on students who experience test anxiety or perform poorly under rigid time constraints. Conversely, allowing unsupervised take-home work favors students who are adept at leveraging expensive, premium AI tools over those who rely solely on foundational study habits.
• Redefining the Value Proposition of Tuition: As automated tools handle routine problem-solving, the justification for university tuition increasingly hinges on human-centric mentorship, peer collaboration, and the rigorous, proctored validation of knowledge.

Institutional Responses and Future Outlook

UC Berkeley, like many top-tier research universities, has encouraged faculty to establish their own classroom policies regarding artificial intelligence, recognizing that a one-size-fits-all mandate is impractical given the diverse nature of academic disciplines. While humanities departments grapple with AI-generated essays, STEM departments face the distinct challenge of automated problem-solving and code generation.

Educational technologists suggest that the long-term solution may involve hybrid models—such as interactive, oral examinations, project-based learning where the process is documented in real-time, or specialized secure testing browsers. However, as AI models become more multimodal, capable of real-time reasoning, and integrated into everyday computing environments, maintaining secure digital boundaries becomes exponentially more difficult.

For now, Professor Borcherds’ approach represents a pragmatic, no-nonsense retreat to foundational principles. By prioritizing verified individual performance over unsupervised digital metrics, the mathematics department at UC Berkeley is drawing a firm line in the sand. As universities worldwide watch to see how this experiment unfolds, the message is clear: the age of easy, unmonitored online grading is drawing to a close, forcing educators and students alike to re-evaluate what it truly means to learn.