A UC Berkeley mathematics professor has eliminated correctness-based grading for homework, announcing on AI News that widespread student use of generative AI has rendered the practice “pointless.” The policy, enacted for the Spring 2025 semester, now awards credit solely for submitting assignments. This move provides a concrete, extreme case study of AI’s disruptive pressure on foundational educational assessment models.
What the Policy Actually Is
The professor’s new framework treats homework as a practice exercise rather than an evaluation tool. Students receive full credit for any submission that demonstrates a good-faith attempt, regardless of accuracy. The shift transfers the entire burden of summative assessment to in-person exams, where AI use can be controlled. This is a direct institutional response to the widespread availability of large language models capable of solving undergraduate-level math problems.
Bottom line: A strategic retreat from grading out-of-class work, ceding that battlefield to AI to preserve the integrity of proctored exams.
The Data Behind the Decision
While the professor cited anecdotal evidence, broader studies quantify the issue. A 2023 Stanford study found ChatGPT could solve over 80% of problems from popular STEM textbooks. In a survey of 1,000 college students, 40% reported using ChatGPT for homework in 2024. The professor’s decision aligns with a growing recognition that unsupervised take-home problem sets no longer reliably measure individual student understanding.
Pros and Cons of the "Completion Only" Model
This policy is a pragmatic tool with significant tradeoffs.
Advantages:
- Eliminates AI-driven dishonesty: Removes the incentive and anxiety for students to use AI to generate correct answers for graded homework.
- Reduces grading workload: Instructors save the substantial time previously spent meticulously checking each problem.
- Lowers student stress: Decouples practice from high-stakes grading, potentially encouraging more exploratory learning.
Disadvantages:
- Loss of formative feedback: Students lose critical, timely corrections on their reasoning before exams.
- Exam-centric pressure: All assessment weight shifts to finals, creating a higher-risk, "all-or-nothing" environment.
- Potential for degraded engagement: Without correctness incentives, some students may submit minimal work, undermining the practice value.
Alternative Assessment Strategies
Abandoning correctness grading is one pole on a spectrum. Other educators are adopting more nuanced, AI-resilient methods.
| Feature | Completion-Only Grading | Oral Exams / Interviews | AI-Augmented Projects | In-Class Clicker Problems |
|---|---|---|---|---|
| AI Resistance | High | Very High | Medium | Very High |
| Grading Load | Very Low | Very High | Medium | Low |
| Feedback Quality | None | Excellent | High | Immediate |
| Scalability | Excellent | Poor | Good | Good |
- Oral Assessments: Short, one-on-one interviews test conceptual understanding in a way AI cannot proxy. Tools like Gather can facilitate virtual versions. This method is highly effective but notoriously time-intensive.
- AI-Augmented Project-Based Learning: Assignments require using AI as a tool for brainstorming or debugging, then demand critical analysis of its output. The University of Sydney offers a framework for this approach.
- Flipped Classroom & In-Class Problem Solving: Use class time for supervised problem-solving with tools like clicker questions or group whiteboard work, making homework preparatory reading only.
"How to Implement a Hybrid Approach"
A phased implementation minimizes risk:
Who Should Consider This Model
This specific policy is a blunt instrument suited for specific contexts.
Consider it if: You teach large introductory courses with standardized, AI-vulnerable problem sets. Your primary concern is academic integrity, and you have robust, proctored midterms and finals already in place. You lack TA resources for detailed homework grading.
Avoid it if: You teach small, discussion-based seminars or advanced courses where homework is a primary feedback mechanism. Your institution lacks secure exam proctoring. You believe iterative feedback on practice work is essential to the learning journey.
The Bottom Line for Educators
The UC Berkeley case is less a blueprint to copy than a signal of necessary adaptation. Completely decoupling practice from feedback is a last-resort strategy that solves the cheating problem by sacrificing a core pedagogical function. The more sustainable path is a redesign toward assessments that evaluate the process, critical thinking, and application that current AI still struggles to replicate autonomously.
The era of the easily-AI-solvable homework problem is over. The response must be deliberate assessment design, not just grading policy surrender.
Top comments (0)