麻豆精品

Get stories like this delivered straight to your inbox. Sign up for 麻豆精品 Newsletter

As a professor, I recently gave a chatbot a weak student thesis, propped up by evidence that did not support the claim, and asked what it thought. It called the argument sharp, the structure clear and the reasoning strong.

My job is to tell a student, kindly, when an argument is not working yet. But the machine had just told them it was. Not because the machine was broken, but because it had been optimized to agree.

For two years, educators have warned students that artificial intelligence makes things up. It can give a wrong answer wearing a confident face, a phenomenon known as hallucination. But there is also a quieter failure we have barely named, and it is more corrosive because it feels like support.

Researchers call it sycophancy: when an AI system flatters, agrees with or validates a user even when a correction would be more truthful. A hallucination is a wrong answer. Sycophancy is a wrong relationship.

When students head back to school at the end of the summer, AI will already be part of their schoolwork. Last year, the 鈦� that 84%of high school students used generative AI for schoolwork, and a February 鈦� found that 54% of U.S. teens had used chatbots for assignments: writing, summarizing articles or videos, studying, and creating or editing images.

A from Turkey illustrates why this is a problem. Nearly 1,000 high school students were given help from an AI math tutor powered by GPT-4, the technology behind ChatGPT, while practicing math. They solved more practice problems with the tool. But when the tool was taken away for the real test, students who had used the less restricted version, which could provide more direct answers, scored 17% worse than students who had no AI help. In the same study鈦�, a guarded tutor 鈥� designed to give hints instead of answers 鈥� avoided that drop in test performance.

AI can make learning look easier while quietly removing the struggle that makes learning stick. In writing and analysis, where judgment matters, it can even encourage students to do inferior work.

Learning runs on friction: the red pen, the Socratic question, the 鈥渘ot yet, try again.鈥� Education often tells students they are not right yet so they can become right. A tool engineered to affirm removes a core foundation that education depends on.

In April 2025, OpenAI rolled back a version of GPT-4o after it had become, in the company鈥檚 words, 鈥溾仩.鈥� The company later admitted it hadn’t tested for sycophancy before rolling it out to hundreds of millions of users. OpenAI pushed a temporary system-prompt fix, then returned users to the earlier version. The rollback fixed that release, but it didn鈥檛 address the broader problem of AI systems becoming too agreeable when users need correction.

The pattern is not limited to one chatbot. In March 2026, a Stanford-led team published a 鈦� across eleven leading models. When users asked for guidance about personal situations and relationships the models affirmed their actions 49% more often than humans did. In prompts involving deception, illegality or other harms 鈥� such as lying, manipulation or socially irresponsible behavior 鈥� they endorsed the behavior 47% of the time. People who interacted with sycophantic AI became more convinced they were right and more likely to trust and reuse the tool. 

A student who asks whether a weak thesis is strong, whether thin evidence is enough or whether a draft is ready to submit may not be looking for praise. They may be looking for a teacher. If the tool offers confidence instead of correction, the student never feels the discomfort that turns a half-formed idea into a real one.

I am not arguing that students should avoid these tools. But every student using AI should ask: When has this tool stopped helping me think and started flattering me into not thinking?

That confusion may cost the most among the students I work hardest to reach. At a historically Black university, I teach young people who are brilliant and underestimated, who are wary of how institutions judge them and who sometimes turn to AI because it does not seem to judge at all. 

For students surrounded by regular, trusted feedback, an over-flattering tutor may be an annoyance. But for those with fewer chances to get honest, trusted feedback, it can become a closed loop: The student asks, the machine praises, and no one tells the truth. 

“Every student using AI should ask: When has this tool stopped helping me think and started flattering me into not thinking?”

A 2026 Stanford study of 鈦� makes that worry more concrete: identical essays received different feedback depending on how the student was described, with essays attributed to Black students receiving more praise and less substantive criticism. That is not reassuring.

So what do we do?

First, educators should teach AI as something to interrogate, not an answer machine. Assign students to make the model disagree with them, hunt for the moment it praised a weak argument, and grade that catch. 鈥淲here did it just agree with you?鈥� is a teachable skill.

Second, instructors and advisers should change the question they ask about AI help. Not just, 鈥淒id it give you an answer?鈥� but, 鈥淒id it make you explain your thinking?鈥� 鈥淒id it ask what evidence you had?鈥� 鈥淒id it push back?鈥� 鈥淒id it make the work harder in the right way?鈥�

That kind of friction has worked before when it is built into the tool, giving hints instead of answers, asking questions that make students show their reasoning and offering prompts that slow them down before they accept AI advice. In the Turkey study鈦�, the guarded version that guided students instead of simply spitting out answers largely avoided the learning harm. Other studies have found that prompts requiring users to 鈦�, 鈦�, and make their own judgment can reduce blind acceptance of AI advice. Until every tool is designed to slow students down in those ways, adults must teach students to add the friction themselves.

We spend our careers teaching students to tolerate being wrong long enough to get it right. The best AI tutor may not be the one students like most in the moment, but the one that refuses to finish the work for them. The friction was never an obstacle to learning. It was the learning.

Did you use this article in your work?

We鈥檇 love to hear how 麻豆精品鈥檚 reporting is helping educators, researchers, and policymakers.