The Perfect AI Essay. The Blank Exam. The Learning That Never Happened.

Two a.m., the student who opened ChatGPT

At 2 a.m., the dorm lights were off. A student opened his phone and typed an essay prompt into ChatGPT. Three seconds later, he had a complete answer with citations and clean structure. He copied it, pasted it, fixed the formatting, and submitted it. The next day, he got a high score.

This is not an isolated case. A survey by journalists and a research team at Minzu University of China covered 1,953 undergraduates at more than 400 universities. It found that 99.18% of respondents used AI. More than 60% were confused about where reasonable use ended. Fewer than 30% said their schools offered systematic AI courses. Students are already using AI. Teaching rules have not caught up.

At the start of the semester, the presidents of Peking University, Fudan University and other universities warned freshmen against overreliance on AI, against "cognitive offloading" and "false mastery." MIT also released a report stating that "many uses of AI are depriving students of opportunities to learn." The warning matters. But if classes, homework and exams continue as usual, warnings alone will not stop "thought outsourcing."

A polished assignment no longer proves a student has learned

Some professors have found final papers with perfect formatting, smooth logic and professional phrasing. Then, in oral defense, the student cannot explain the central argument or the reasoning. Homework accuracy is high. Closed-book exam scores fall off a cliff. This is "false mastery": looking at AI's polished answer and mistaking it for one's own understanding. Once the tool is gone, the gaps show.

MIT Media Lab's EEG experiment gives more direct evidence. Participants who wrote essays with ChatGPT showed much lower brain activity than those who used search engines or wrote by hand. Their neural connectivity was reduced. Minutes after finishing, they had trouble recalling what their essays said. When students hand the thinking step to AI, memory networks are not fully activated. The task looks finished, but the cognitive process needed to understand the material did not happen. Worse, the dependence is self-reinforcing. The more one trusts AI, the lower one's cognitive engagement. The lower the engagement, the more one depends on AI.

A Swiss survey of more than 600 people found that 17- to 25-year-olds relied on AI more than older people. Their critical-thinking scores were about 45% lower. The problem is just as visible in primary and secondary schools. Students use AI to write essays, answer reading-comprehension questions and organize wrong-answer notes. The homework looks far better than the student's level. Teachers struggle to tell the difference. Parents see full marks and feel optimistic. The gaps accumulate. The higher the grade, the sharper the drop.

Assessment is too old for AI

Blame AI alone and we miss the point. AI can replace so many learning steps because those steps were easy to replace in the first place.

Many assignments were always material assembly and format compliance. Assessments often looked only at the finished product, never at the process. Courses often centered on moving knowledge from one place to another, with little training in thought. AI simply makes the old assembly line run faster and smoother. It is a mirror. It shows the holes that were already there. When learning is reduced to submitting a polished product, AI becomes the cheapest subcontractor.

The deeper problem is the invasion of efficiency logic into education. At work, cutting repetitive labor raises efficiency. In learning, some seemingly inefficient labor is exactly how ability forms. A student reads a difficult passage closely. He revises an argument again and again. He struggles alone with a hard problem. He does not do this only to get the final product. Thinking, trial and error, review and polishing shape logical reasoning, resilience and problem-solving. Getting an answer from AI in ten seconds looks efficient, but it skips every step of growth. Cutting repetitive labor is efficiency. Cutting training is education canceling itself.

Hollowed-out skills drain the competitive edge

Overreliance on AI has a direct consequence: hollowed-out ability. Give up deep thinking long enough, and the brain learns laziness. Independent problem-solving, logical derivation and written expression keep weakening. Many students show classic "AI dependence syndrome." Without AI, they cannot write a complete essay, solve a complex problem or organize an argument. They look knowledgeable, but their foundations are thin.

AI tends to give the best standard answer. Students lose the chance to err, reflect and polish. Resilience, error correction and creative thinking weaken. Faced with problems that have no standard answer, they freeze. For the education system, overreliance turns learning into performance. Homework, papers and classroom exercises can all be delegated to AI. The learning process becomes virtual. Students receive results and submit products. Little learning settles. Schools cannot judge students' level through assignments and exams. The evaluation system gradually fails.

Top universities do not train machines that answer questions well. They train people who can think, create and solve unknown problems. When most students replace thought with AI, learning outcomes become homogeneous. Future competition in education will reward independent thought, critical judgment and creation. The warnings from university presidents are an early alarm.

Rules belong in the syllabus; assessment must verify ability

Warnings are a starting point. Institutions are the safeguard. If classes, homework and exams stay the same, moral exhortation alone cannot stop thought outsourcing. Reform must reach every course, every assignment and every exam.

Start with the syllabus. Every course should state which steps allow AI, which must be done independently, and what students must disclose after use. These rules should be clear when the task is assigned. They can differ by course. A literature course training close reading should have students read the original text and propose an interpretation first. Then they can use AI to find opposing views. A translation-revision course can treat AI output as an object of discussion. MIT's report recommends that every course state its AI policy in the syllabus and on the course website, with reasons. A blanket "encouraged" or "banned" cannot replace instructional design.

Move assessment from checking products to verifying ability. After a paper is submitted, ask the student to explain why a key quotation supports the conclusion. After code is written, change one condition and ask the student to revise it and explain why. After a research report, ask how the sample was chosen and what material overturned the first judgment. Such follow-up questions let teachers see how far the student understands. In-class writing, live demonstrations, oral discussion and out-of-class projects should be combined. They test both independent work and the ability to solve problems with tools. AI can generate a clearly argued answer. It cannot replace a student explaining his or her reasoning on the spot.

Teacher responsibility and institutional support must move together. If teachers use AI to generate slides, assign tasks and grade papers, but do not check content or give targeted feedback, then telling students to think independently lacks persuasiveness. Tools can assist teachers. The person responsible for course quality and evaluation results must still be the teacher. At the same time, reform must not become a new burden for teachers. One teacher facing a hundred students, expected to hold oral defenses and track every draft, will turn these requirements into another round of form-filling if the school does not provide teaching assistants and adjust workloads. What universities need to invest in is time for teachers to know students and guide them.

AI literacy must move from general education to depth. Many places now require universities to offer mandatory general AI courses. The Ministry of Education and eight other departments have called for "general plus specialized" AI general-education courses. But a general course cannot become a popular-science lecture, nor can it become a programming barrier. Train students, within specific disciplines, to judge the quality of AI output. Students who have not read the literature cannot recognize a fabricated citation. Those who have not done the derivation will miss a skipped step in an answer. Critical thinking grows out of concrete knowledge and repeated practice. It does not come from a few slogans about "keeping independent thought."

Beware of turning AI usage rates into a new teaching achievement metric. The number of "AI+" courses opened and questions answered by a platform does not directly prove students' judgment improved. Forcing every course to use AI may crowd out independent reading and foundational training that should be preserved. Universities must teach students to use AI. They must also help them build the professional foundation to judge AI. The more powerful the tool, the more human judgment matters. Judgment comes from solid knowledge and repeated training.

What universities cannot outsource

After AI enters the university, the danger is a kind of teaching idle. Students submit assignments. Teachers finish grading. Schools collect impressive numbers. The only thing missing is the student's ability. A university diploma should guarantee what kind of training a person has received and what kind of judgment that person can make. That responsibility cannot be outsourced.

Education aims to produce a complete person with thought, judgment, creativity and warmth. An efficient answering machine is not the goal. AI can generate answers, but it cannot experience thinking for the student. AI can save time, but it cannot replace growth through trial and error. Using AI moderately to assist learning, cut repetitive labor and broaden one's view is reasonable. But once the initiative of thought is handed over, what is saved may be education itself.

At 2 a.m., the student submitted a high-scoring paper. Three months later, in a closed-book exam, he stared at the test sheet and did not recognize the arguments he had "written." He reached for his phone to ask AI again. There was no signal in the exam room.

He put the phone down and faced a blank sheet. There was nothing on it.