Students are turning in better homework. They’re bombing the exams that actually test whether they learned anything.
That’s the quiet crisis unfolding in college classrooms four years after ChatGPT arrived. Professors who once worried about cheating now face something deeper: large language models are eroding the hard, uncomfortable work that produces real thinking.
The pattern is consistent. Students use AI for polished essays, clean code, and solved problem sets. Then they freeze when asked to explain their own ideas, write without the chatbot, or solve similar problems on a closed-book test. MIT’s Eric Klopfer calls it “the illusion of learning.” It feels like mastery. It isn’t.
Early studies back the classroom observations. Brain scans show weaker connectivity when students write with AI. People who lean on the models give up faster once the help disappears. Surveys find most students already use the tools—and two-thirds suspect the more they rely on them, the more their critical thinking suffers.
Faculty are watching the consequences in real time. At Carleton, first-year humanities students handed in dissertation-level outlines they couldn’t discuss. At Berkeley, computer-science students aced automated test cases on AI-generated code, then scored near zero when asked to explain it orally. Attendance and office hours have dropped. Failure rates on in-person exams have spiked.
The response is scrambling. Some professors have banned take-home writing, gone tech-free, and brought back blue books and oral exams. Others redesign assignments so AI becomes a sparring partner instead of a ghostwriter. A few are redesigning entire courses around the skills machines still can’t fake: judgment, struggle, and the ability to think without a prompt.
The technology is moving faster than the data. But the early signal is clear. When students outsource the struggle, they also outsource the learning. What remains is fluent output and hollow understanding—the perfect résumé of a generation that never had to think for itself.