AI Detection Pushed My Students to Use AI

This fall, a student told me she began using generative AI only after learning that stylistic features such as em dashes were rumored to trigger AI detectors. To protect herself from being flagged, she started running her writing through AI tools to see how it would register.

AI detection is failing in ways that go beyond accuracy. It is reshaping student behavior by making not using AI feel risky, especially at open-access institutions like the City University of New York, where students encounter inconsistent AI policies across courses. In 2023, researchers at Stanford University warned that detection tools were biased against nonnative English writers. Their study, published in Patterns, found that detectors misclassified more than half of TOEFL essays written by nonnative speakers as AI-generated, while nearly all essays by native speakers passed. The authors anticipated that such bias could encourage defensive or evasive behavior. Two years later, that prediction is taking shape in my writing classrooms, extending even to native English speakers whose polished prose has begun to feel suspect.

What matters here is less the accuracy of detection than the lesson it teaches students. Detection tools communicate, even when instructors do not, that writing is a performance to be managed rather than a practice to be developed. Students learn that style can count against them, and that fluency invites suspicion. In that environment, AI becomes a way to smooth uncertainty. It offers reassurance that language will appear ordinary enough to pass.

This is a subtle shift with real consequences. Writing begins to function as a source of risk rather than evidence of learning. Students adapt accordingly. They seek tools that reduce exposure and normalize tone. Over time, detection begins shaping student behavior in ways that resemble instruction, even when no explicit guidance is given.

A few examples. One student, a native English speaker, had long been praised for writing above grade level. This semester, a transfer to a new college brought a new concern. Professors unfamiliar with her work would have no way of knowing that her confident voice had been earned. She turned to Google Gemini with a pointed inquiry about what raises red flags for college instructors. That inquiry opened a door. She learned how prompts shape outputs, when certain sentence patterns attract scrutiny, and ways in which stylistic confidence trigger doubt. The tool became a way to supplement coursework and clarify difficult material. Still, the practice felt wrong. “I feel like I’m cheating,” she told me, although the impulse that led her there had been defensive.

After being accused of using AI in a different course, another student came to me. The accusation was unfounded, yet the paper went ungraded. What followed unsettled me. “I feel like I have to stay abreast of the technology that placed me in that situation,” the student said, “so I can protect myself from it.” Protection took the form of immersion. Multiple AI subscriptions. Careful study of how detection works. A fluency in tools the student had never planned to use. The experience ended with a decision. Other professors would not be informed. “I don’t believe they will view me favorably.”

Detection culture had already begun shaping how students think about writing. One afternoon, four high-school students enrolled in college courses through Hostos Community College’s partnership with the Health, Education, and Research Occupations High School stopped me two blocks from campus. They wanted to understand how professors decide when student language sounds suspicious. “If you’ve never heard what AI actually makes,” one said, “you don’t know which phrases draw attention.”

Another explained that some students try the technology early so nothing in a paper invites scrutiny. A third asked whether professors who assign essays on music videos expect discussion of synthetic vocals or autogenerated visuals, and whether that kind of knowledge itself might raise concerns. Standing on a sidewalk, they tested scenarios against each other, reasoning through a problem none of them had been taught to navigate. They had not yet formally entered college.

I have taught at Brown University, at the John Jay College of Criminal Justice, at Hostos Community College, and in early-college sections serving high-school students through Hostos’s dual-enrollment partnership. That range has sharpened my attention to how different populations manage institutional uncertainty. At CUNY, many students work 20 to 40 hours a week. Many are multilingual. They encounter a different AI policy in nearly every course. When one professor bans AI entirely and another encourages its use, students learn to stay quiet rather than risk a misstep. The burden of inconsistency falls on them, and it takes a concrete form: time, revision, and self-surveillance. One student described spending hours rephrasing sentences that detectors flagged as AI-generated even though every word was original. “I revise and revise,” the student said. “It takes too much time.”

Midway through the semester, I stopped requiring students to disclose their AI use. My syllabi had asked for transparency, yet the expectation had become incoherent. The boundary between using AI and navigating the internet had blurred beyond recognition. Asking students to document every encounter with the technology would have turned writing into an accounting exercise. I shifted my approach. I told students they could use AI for research and outlining, while drafting had to remain their own. I taught them how to prompt responsibly and how to recognize when a tool began replacing their thinking.

The atmosphere in my classroom changed. Students approached me after class to ask how to use these tools well. One wanted to know how to prompt for research without copying output. Another asked how to tell when a summary drifted too far from its source. These conversations were pedagogical in nature. They became possible only after AI use stopped functioning as a disclosure problem and began functioning as a subject of instruction.

I have watched detection produce the behavior it was designed to prevent. At open-access institutions like CUNY, where students already carry the weight of work, family, and policy inconsistency, the cost is tangible. It appears in the silence of students who never used AI until fear of accusation gave them a reason to start.