UC Berkeley Fields Medalist Borcherds drops correctness grading after AI solves homework

Borcherds shifted to completion-only grading after observing widespread AI use that rendered paid reader evaluation pointless. The change followed OpenAI claims of solving the Navier-Stokes existence problem and occurred amid Berkeley Senate guidance requiring AI-use statements in syllabi. Faculty retain discretion on enforcement, with tools like Turnitin and Rumi rejected for campus standardization due to inadequate detection rates.

Campus AI Hub supplies sample syllabus language while policies vary by course level. Borcherds noted academic inertia that preserved paper journals for decades after digital alternatives emerged, and he compared current shifts to loom and automobile displacements of prior labor. Five-year forecasts of mathematician career safety have collapsed, leaving even ChatGPT responses inconclusive on future human roles.

Operational impact centers on assessment redesign: completion metrics replace verification, decoupling homework from skill measurement. Related benchmarks such as GSM8K and MATH dataset saturation show model performance exceeding 90 percent on contest problems by 2024, aligning with Borcherds observations. Departments face pressure to replace homework with in-class or oral components that current agents cannot yet replicate at scale.

Departments retaining unchanged structures through 2026 will accumulate larger gaps between reported grades and actual competence, prompting external credentialing shifts.