Students Dumbing Down Writing AI Might Get Credit For

Your daughter writes a strong essay, reads it over, and then swaps “consequently” for “so” and cuts her best paragraph. Not because it was wrong. Because it sounded too good. Researchers at Harvard’s Center for Digital Thriving now have a name for this: cognitive withholding, students dumbing down writing AI might get credit for, so a suspicious teacher won’t accuse them of cheating. In a September 2026 report, the center described teens who deliberately underperform to avoid being flagged. It is a small signal from a small interview sample, but it points at a real cost of the AI-cheating panic: kids learning to hide their best thinking.

Key Takeaways

• Harvard’s Center for Digital Thriving (September 2026) heard from teens who deliberately do worse work to avoid AI accusations; co-founder Beck Tench calls it “cognitive withholding.”
• 74% of teachers and 69% of principals who faced an AI dilemma described one about cheating, so suspicion is the default climate in many schools.
• AI detectors are not reliable enough to settle the question: a 2023 test of 14 tools found none above 80% accuracy, and another study found they misflag non-native English writers.
• The best protection for an honest kid is a visible process: drafts, notes, version history and the ability to talk about the work.
• At home, praise the thinking you can see, and never tell a child to write “worse” to stay safe.

What “cognitive withholding” means and where it came from

Cognitive withholding is a student deliberately producing less capable work than they can, so the work won’t look AI-generated. The term comes from An AI Policy Isn’t a Playbook, a report by Beck Tench, Emily Weinstein and Carrie James of the Center for Digital Thriving at the Harvard Graduate School of Education, published in late September 2026.

The report draws on two kinds of evidence. The first is a nationally representative survey of more than 1,000 US public school teachers and principals, fielded in spring 2025. The second is more than 40 interviews with 12 educators and 31 young people aged 15 to 19, conducted from April to June 2026.

The survey sets the scene. When educators described an AI dilemma in their own words, 74% of teachers and 69% of principals described something about cheating. The interviews supply the surprise. Teens talked about fear of false accusation and pressure to keep up with peers, and, as Education Week reported, some described “deliberately underperforming on assignments so they’re not falsely accused of using AI.”

Tench put it this way to Education Week: adults are focused on “cognitive surrender,” students handing their thinking to a chatbot, which is a real concern. “But we have a signal in this report that there’s also this cognitive withholding that some students are doing in order to evade being falsely accused in environments where accusations don’t get any kind of conversation.”

Notice the last clause. The withholding shows up where an accusation is a verdict instead of the start of a talk.

Why students dumbing down writing AI might get credit for are acting rationally

The report describes a cycle it calls two-way suspicion. Teachers develop “AI-dar,” a hunch that a piece of work was machine-written. They usually can’t prove it. Students can’t disprove it either. So both sides start to guess.

For a student, the cheapest defense against a hunch is to stop producing whatever triggers it. If the teacher seems to treat polished vocabulary, tidy structure or an unusually good paragraph as a red flag, the rational move is to write less polished, less tidy, less good. One sophomore in the report called the situation “kind of like a hellscape for both students and teachers.”

That logic makes sense once you know how weak the tools behind many hunches are. AI-text detectors estimate how statistically predictable a passage is. Language models choose likely next words, so very fluent, very regular prose scores as “more AI-like.” The trouble is that careful human writers, kids taught to use transition words, and students writing in a second language can also produce regular prose.

The research on this is not encouraging:

• In a 2023 study in the International Journal for Educational Integrity, Debora Weber-Wulff and colleagues tested 14 detection tools and concluded they “are neither accurate nor reliable.” All scored below 80% accuracy, and the authors said the tools should not be used as evidence of academic misconduct.
• A 2023 Stanford study by Weixin Liang and colleagues, published in Patterns, found that GPT detectors “consistently misclassify non-native English writing samples as AI-generated,” while native writing was classified accurately.

Detectors have changed since 2023 and vendors report better numbers. Those are the vendors’ numbers. The underlying problem stays: a probability score about style is not proof of who did the thinking.

There is also a learning cost the report flags. When students use AI heavily, teachers lose sight of how a student writes and how that changes over time. When students withhold, teachers lose the same thing from the other direction. Either way, the teacher stops seeing the real kid.

How common is this? What the evidence can and can’t say

This is where honesty matters. The report does not give a percentage of students who withhold. The finding comes from qualitative interviews with 31 teens, which is the right method for discovering a pattern and the wrong one for measuring how widespread it is. We know it happens. We don’t know how often.

What we can measure is the climate that produces it.

Signal	Number	Source	What it tells you
Educators whose AI dilemma was about cheating	74% of teachers, 69% of principals	Harvard CDT survey, spring 2025	Suspicion dominates how adults think about AI in school
US teens who think AI cheating is common at their school	59%	Pew Research Center, survey Sept–Oct 2025	Kids assume their classmates are doing it
College students who used AI to solve problem sets vs. those who think peers did	11% vs. 49%	Dartmouth AI Survey 2026	Students overestimate peers’ cheating by a wide margin
College students who felt pressure to use AI because peers did	39%	Dartmouth AI Survey 2026	Belief about others changes behavior
Detection tools scoring 80% accuracy or better	0 of 14	Weber-Wulff et al., 2023	Tools can’t settle an accusation alone

Put those rows together and the mechanism is plain. Adults expect cheating. Kids expect cheating. The tools can’t sort honest from dishonest. An honest, capable student is the one with the most to lose in that system, and withholding is one way to protect herself.

Two caveats. The Dartmouth figures come from college students, not middle or high schoolers. And the Pew Research Center number measures what teens believe about their school, not how much cheating happens.

The other hidden cost: kids stop asking teachers for help

The report found something adjacent that parents should hear. Teens described going around their teachers entirely. “I don’t usually go to my teachers. I kind of just go to Chat,” one 16-year-old junior said. Tench told Education Week that punitive bans push AI use further underground, and that “the uses that are invisible to us aren’t just the bad uses. They are good uses, too.” She gave the example of a student who built an entire system with AI to manage his assignments.

So suspicion costs twice. The honest kid hides her ability. The kid who could use AI well hides his method. Neither one learns from an adult.

The report’s recommendation for schools is to move teachers “from AI refs to AI coaches,” using five free classroom exercises it calls plays. One, “Align on the Line,” has students brainstorm the ways AI could be used on an assignment and decide together which uses cross the line. Another asks teachers to move “from assuming to asking.” The good news, the authors argue, is that this requires relationship skills teachers already have, not technical expertise.

If your child’s school is still in referee mode, you can run a home version of the coaching approach.

What to do at home if your child is holding back

Ask whether it’s happening, without leading

Kids rarely volunteer this. Try: “Do you ever change your writing so a teacher won’t think it’s AI?” Ask it flatly, the way you’d ask about lunch. If the answer is yes, ask what they changed and why. You’ll learn which signals your child thinks teachers watch for, and you may learn about a specific teacher or class.

Build a process trail as a habit, not a defense

A student with drafts can answer a hunch with evidence. Encourage writing in a tool that keeps version history, such as Google Docs, and keeping brainstorm notes, outlines and scrap paragraphs. Frame it as how real writers work. Engineers keep lab notebooks for the same reason: the record shows how you got there. If your child ever is questioned, a timestamped history of the essay growing over three evenings is far more persuasive than any argument.

Rehearse talking about the work

The strongest proof of authorship is being able to explain your own thinking. At dinner, ask your child to tell you the argument of the essay in two sentences, or why they chose one example over another. Kids who can do this comfortably are the ones who come out of an accusation fine, and the practice makes them better writers anyway.

Praise what you can see them doing

Praise the hard part: “You rewrote that opening three times and it’s much clearer now.” Comment on effort and choices rather than on how impressive the final product sounds. This tells a child that their real ability is the thing you value, which is the opposite of the message withholding sends.

What not to do: tell your child to write “worse” to stay safe

It is tempting, if your kid was accused once, to advise playing it safe. Don’t. Teaching a child to hide their best vocabulary or flatten their structure trains the exact habit the Harvard team is worried about, and it can stick. If a teacher’s suspicion is the problem, the fix is a conversation with the teacher, supported by the process trail, not a smaller version of your kid. For how that conversation can go, see our guide to what parents should know before assuming a kid cheated with AI and our explainer on students’ rights when schools falsely accuse them of AI cheating.

Talk to the school about how accusations are handled

You don’t need to attack anyone’s policy. Ask one practical question: “When a teacher suspects AI use, what happens first?” The answer you want is a conversation with the student and a look at drafts, not an automatic zero. Some districts and states are writing this into policy; the recent Turnitin lawsuit coverage shows why.

What to Watch For Over the Next 3 Months

• Week 4: Notice whether your child talks about writing to “not sound like AI.” Ask to see one draft history. Is the process trail habit taking hold?
• Month 2 red flags: A strong writer whose grades drop without explanation, a child who stops sharing work, avoids a particular teacher, or says “it doesn’t matter what I write, they’ll think it’s AI.” The Harvard report calls a related feeling “why botherism,” the sense that if AI can do a task, there’s no point trying. Watch for it.
• Month 3 self-check: Can your child explain any recent assignment in their own words? Do they know what happens at their school if they’re suspected of AI use? Have you had at least one calm talk about where AI help crosses the line?

Frequently Asked Questions

Is cognitive withholding a proven, widespread problem?

Not proven in scale. The Harvard Center for Digital Thriving identified it in interviews with 31 teens aged 15 to 19, which shows the behavior exists but not how common it is. What is well measured is the suspicious climate that encourages it: most educators’ AI concerns are about cheating, and detection tools can’t reliably settle accusations.

My child writes very formally. Will a detector flag them?

It could. Detectors score how predictable and regular text is, and formal, careful prose can look “AI-like.” A 2023 Stanford study in Patterns found detectors misclassified non-native English writers’ essays as AI-generated. The practical defense is version history and drafts, plus your child being able to explain their work aloud.

Should I tell my kid to add mistakes so teachers believe it’s theirs?

No. That trains a habit of hiding ability, which is the cost the Harvard researchers warn about. If your child feels they must write worse to be believed, talk to the teacher. Bring the drafts. Ask how the class handles suspected AI use. The goal is a fair process, not a weaker kid.

What should happen if a teacher suspects AI use?

A fair process starts with a conversation, not a penalty. The teacher should talk with the student, look at drafts or notes, and let the student explain their thinking. Detector scores alone shouldn’t decide it; even detector researchers say they are unsuitable as sole evidence. Ask your school whether its policy says this explicitly.

How do I bring this up without making my kid defensive?

Make it about the system, not their honesty. Something like: “I read that some kids change their writing so teachers won’t think it’s AI. Does that happen at your school?” Talking about classmates first often makes it easier for a teen to talk about themselves.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Tench, B., Weinstein, E., & James, C., with Starks, A., & Konrath, S. (2026). “An AI Policy Isn’t a Playbook: Educators Need Agency, Not Just New Rules.” Center for Digital Thriving, Harvard Graduate School of Education. https://digitalthriving.gse.harvard.edu/an-ai-policy-isnt-a-playbook/
2. Prothero, A. (2026, September 30). “How AI Suspicions Are Undermining Student-Teacher Relationships.” Education Week. https://www.edweek.org/technology/how-ai-suspicions-are-undermining-student-teacher-relationships/2026/09
3. Pew Research Center. (2026, February 24). “How Teens Use and View AI.” https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/
4. Dartmouth College. (2026, October 30). “Survey reveals what students think about AI” (Dartmouth AI Survey 2026). Phys.org. https://phys.org/news/2026-10-survey-reveals-students-ai.html
5. Weber-Wulff, D., et al. (2023). “Testing of detection tools for AI-generated text.” International Journal for Educational Integrity, 19, 26. https://link.springer.com/article/10.1007/s40979-023-00146-z
6. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). “GPT detectors are biased against non-native English writers.” Patterns, 4(7). https://doi.org/10.1016/j.patter.2023.100779