Yes, some students are becoming dependent on AI, but the risk sits in how a task is set up, not in the tool itself. In a field experiment with nearly 1,000 high school math students, those who practiced with an unrestricted chatbot scored about 17% lower on a later exam than students who never had one. A tutor built to hold answers back largely removed the damage. A principal's job is to decide, task by task, which of those two tools the student is actually using.
What does the evidence say about students becoming dependent on AI?
Students who use a general-purpose chatbot to practice tend to perform better while it is open and worse once it is closed. The clearest test is the study by Hamsa Bastani and colleagues, published in the Proceedings of the National Academy of Sciences in 2025. Students with an ordinary GPT-4 interface improved their practice grades by about 48%. Students with a version prompted to guard learning improved by about 127%. When access was removed for the exam, the first group scored about 17% below students who had never had the tool. The authors' reading is that the first group used the AI as a crutch. The guarded tutor largely protected against that.
The OECD reached the same conclusion across a wider body of research. Its Digital Education Outlook 2026, published January 19, 2026, reports that students using general-purpose tools often produce better work, but those advantages tend to disappear, or even reverse, when AI access is removed. In its words, completing a task with generative AI "does not necessarily translate into learning." It also warns of "metacognitive laziness," where handing the thinking to a chatbot lowers engagement and long-term skill development. Tools designed or used for teaching showed more durable benefits.
Principals already suspect this. In an IEA brief published in August 2026, drawn from the 2023 International Computer and Information Literacy Study, over three quarters of 1,172 principals across 11 countries thought it likely that students would come to rely on these tools instead of learning independently. That is an opinion survey, collected in 2023 and early 2024, so read it as what school leaders expect, not as a measurement of what students do. A US survey commissioned by IBM and run by Morning Consult in July 2026 found a similar worry among 1,019 K-12 professionals: 52% of educators named student dependency as a concern, ahead of cheating at 47%. IBM sells AI education products, so weigh the framing, though the question and sample are plain.
Students say it too. A Year 11 student at Rongotai College in Wellington told 1News in August 2026: "I've become really dependent on asking ChatGPT for help."
One limit on all of this: the strongest experiment covers one subject, math, and one kind of practice. It is the best evidence available, not the last word, and it says nothing about whether a student who writes with AI loses the ability to write.
Why is "dependence" the wrong thing to ban and the right thing to check?
Dependence is not what a student feels about a tool. It is what the student cannot do when the tool is gone. That distinction changes the decision. A ban attacks the tool and misses the cause, which is substitution: the effortful part of a task, the part that was meant to build a skill, gets done by the machine instead of the student.
Think of a ninth-grader using a calculator in a physics lab. Nobody worries, because the point of the lab was never long division. Now put the same calculator in front of a second grader learning to add. The device is identical. The task is different, and the task is what decides whether the tool helps. My governing principle is "outsource the doing, not the thinking", and a school has to add one clause for children: sometimes the doing is the thinking.
What is the Unaided Check?
The Unaided Check is my suggested way for a school to test any student use of AI before it spreads. It is a way of thinking, not a validated instrument.
The Unaided Check is three questions I ask about any student use of AI before a school endorses it: What was the effort in this task meant to build? Does the AI version keep that effort or remove it? And when would anyone find out, without the tool in the room? A use that survives all three is support. A use that fails any of them is substitution, and dependence is what substitution produces when nobody checks.
A use that fails the check does not have to be banned. It gets one of three settings, and the settings are the useful part. They apply to the task, not to the student and not to the tool.
| Setting | What AI does | Use it when | Hypothetical example |
|---|---|---|---|
| Closed | Nothing. The tool is off. | The effort is the skill, and it has not formed yet. | First attempts at factoring in an algebra unit. |
| Guarded | Asks, hints, checks reasoning. Does not hand over the answer. | The student is stuck and needs a push, not a rescue. | A tutor-style assistant after two unaided attempts. |
| Open | Anything the student chooses. | The student has shown the skill unaided, and the task now asks for something bigger. | Using AI to test a finished argument for gaps. |
The order matters. Most schools write a rule for "AI use" in general, usually some version of "allowed with acknowledgment", and that rule is a single setting applied to every task. Closed, Guarded and Open let a department say "Open" for a research project and "Closed" for a first draft in the same week, and defend both.
What do I see in practice?
The mistake I see most often is a school that writes one rule for all work and never runs the third question. I have trained more than 150,000 educators across 30+ countries, and in the leadership sessions I run, the third question, when would anyone find out, is the one that draws silence. Leadership teams can say what they allow. Few can say how they would know, a week later, whether a student can still do the thing.
Across the leadership teams I work with, the pattern runs like this. The policy permits AI for homework. Homework quality rises. Nobody looks at what happens in class on Monday, with the tool out of reach. By the time the first assessment shows the gap, a term of practice has gone by with the effort removed.
How does a principal run the Unaided Check this term?
A principal can run the check in five steps, and none of them needs a new policy.
- Pick three tasks. Choose the tasks where students in one year group use AI most. A quick student survey or a conversation with a department head is enough. Take the real tasks, not the ones in the handbook.
- Name the effort. For each task, write one sentence on what the struggle was meant to build. If nobody can say, the task may be worth redesigning, with or without AI.
- Set the setting. Closed, Guarded or Open, for each task, for now. A setting can change once the skill has formed.
- Schedule the unaided check. Ten minutes, a week later, low stakes, in class, no tool. A short version of the same task. This is the step that nearly everyone skips, and it is the only one that answers the third question.
- Read the result as information about the task. If half a class stalls, the task is substituting, not building. Move it toward Closed or Guarded and run the check again. Do not read the result as a verdict on the students, and do not publish it as a league table.
A hypothetical example: a ninth-grade math team finds that most students use a chatbot to produce worked solutions on problem sets. The effort the sets were meant to build is choosing a method. Under the Unaided Check, the first two attempts at each new problem type move to Closed, and a Guarded assistant is allowed after that. A ten-minute unaided problem, a week later, shows the team whether the method stuck. That is close to the structure of the guarded tutor in the Bastani experiment, though I am drawing the parallel, and the study did not test this routine.
What if students are already dependent?
Dependence shows up as an inability to start. A student who cannot begin an essay paragraph without the tool open, or cannot explain a piece of their own submitted work, is telling you the effort has been removed for some time. Do not start with a confrontation.
Start by moving the first five minutes of a task back to Closed, so that students attempt something before the tool opens. Then ask the students what they notice. The Read-Back Test, a set of questions students answer about a draft policy, works well here, because students know exactly where in their week the tool has become a crutch. Tell them what you are looking for, and tell them it is about the task.
Doesn't every student need to use AI anyway?
Yes, and the Unaided Check does not argue otherwise. The counterargument is that students will work with AI throughout their careers, so building the habit early is the skill. It is right about the destination and wrong about the order. Open is the correct setting once a student has shown the underlying skill unaided, because that student can tell when the tool is wrong. A student who never formed the skill cannot check the output, and checking is the job AI leaves for people. The setting moves toward Open as the skill forms. It does not start there.
Where does this leave a principal's decision?
The decision is smaller than the debate makes it look. You do not have to rule on whether AI is good for students. You have to name, for the three or four tasks where students use it most, what the effort was for, and then find out a week later whether it still happens. Everything else, the rules about acknowledgment, the age limits discussed in what age students should use AI, and the assessment questions in assessment in the age of AI, sits on top of that answer.
The same logic runs through the adult version of the problem in building capability, not dependence, and through what AI literacy actually means in a school. For the coursework end of this, how a head of department handles AI in coursework takes the Unaided Check into assessed work, and what to do when a student is accused of using AI covers the case where the check was never run and the question arrives as an accusation. Staff need the same judgment you are asking of students, which is the subject of AI skills for staff beyond basic training.
If your leadership team is working through these questions, this is the kind of work I support through AI training for schools, and for a conference or a board session I speak on it through my education keynotes.
Sources and further reading
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O., and Mariman, R. "Generative AI without guardrails can harm learning: Evidence from high school mathematics." Proceedings of the National Academy of Sciences, 122(26), 2025. Record.
- OECD. "OECD Digital Education Outlook 2026." Published January 19, 2026. Report page. Quotations are taken from the OECD's published summary of its findings.
- Meinck, Holtmann, Fraillon and Strietholt. "School Principals as Change Agents in an Era of AI-driven Transformation: Insights from ICILS 2023." IEA Compass: Briefs in Education No. 31, International Association for the Evaluation of Educational Achievement, August 2026. PDF.
- IBM and Morning Consult. "New IBM study finds AI adoption is outpacing K-12 readiness." IBM Newsroom, September 2, 2026 (survey of 2,048 US adults, July 2026). Release.
- 1News. "Students say they're becoming 'dependent' on AI for schoolwork." August 14, 2026. Article.
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and an international keynote speaker who has trained more than 150,000 educators. More about Dan.
