How Do I Keep Students from Over-Relying on AI?

A student aces the take-home problem set and leaves half the exam blank. Someone hands in a lab report they cannot walk you through, or a block of code they cannot explain line by line. A colleague wants the tool banned in the first-year course and you are weighing what that costs the students who lean on it most.

The bottom line: over-reliance is mostly a placement problem. The designs that reduce it put the student's attempt ahead of the tool, configure the AI to hint instead of answering, make the check a required deliverable, and keep one graded moment where the tool is absent.

The short version

1. Put the student's own ideas, hypothesis, or rough draft in front of the tool before it opens.

2. Configure the AI to hint, ask, and give examples wherever the target skill is what you grade.

3. Make verification a deliverable with its own line on the assignment.

4. Keep the graded moment tool-free, and compare the assisted work against it.

5. Time the reflection prompt to the phase of the task it can still change.

6. Look at who your design leaves behind, by prior knowledge rather than by average.

Decide whether you have a problem, or just heavy use

Cognitive offloading is the transfer of a cognitive demand to an external tool, which frees limited mental resources for other work. Guo and Ye (2026) review the construct through Nelson and Naren's dynamic metacognitive model, where monitoring of difficulty drives a decision to offload to an internal or an external strategy. Offloading is a self-regulatory choice, and the failure worth worrying about is miscalibration; volume on its own tells you little.

Biswas (2026) models reliance as three actions: solve alone, accept the AI's answer unverified, or use it and verify. Calibration fails in both directions. Over-reliance is accepting wrong output, and under-reliance is discarding useful AI after it errs. Collective over-reliance is a population that has abandoned verification, and because raw over-reliance diverges from regret, heavy use is compatible with good outcomes. The student who checks every AI answer against your materials is not your problem. The one who submits a model's first draft is.

Lodge and Loble (2026) describe the mechanism as "fluency on demand": coherent, confident output that bypasses the desirable difficulties of retrieval, elaboration, and generation through which knowledge is consolidated. Their performance paradox is that assisted work feels fluent and students perform well in the moment while retaining less, which they call an illusion of competence. They name metacognitive laziness (after Fan et al. 2024), where convenience lets learners abdicate self-regulatory processes they need to develop.

Telling them apart in your own course takes a comparison rather than a number. Look for work that is better than the student's earlier work and worse than their explanation of it: a strong problem set followed by a weak exam, a polished report the student cannot discuss, a correct answer that cannot be reproduced on a similar problem. No single pattern proves anything. Two or three together justify changing the design.

Most of what you can measure here is self-report. Gerlich (2025) surveyed 666 UK participants with 50 interviews and found AI use negatively correlated with critical thinking (r = -0.68), with offloading partially mediating (total effect b = -0.42; indirect b = -0.25). Gao, Sun and Khan (2026) used three-wave time-lagged survey data from 623 Chinese students plus educator interviews and found that effective AI use raises sustainable learning performance and over-reliance at the same time. Both designs are correlational, and both say so.

Behavioral instruments are thinner. PAUSE (Alam, 2026) is a browser-only self-check with no reliability or validity evidence, and its authors warn about item validity. The items record when and how often AI enters a workflow, not whether the student's own reasoning stayed engaged, so a student who scaffolds deliberately with AI early will honestly score as an offloader. PAUSE also cites Padmakumar et al.'s (2026) Offloading Score, which estimates the fraction of effort offloaded from behavioral logs (n = 40 developers). What you want to know is what the student can do unaided.

Put the student's attempt in front of the tool

The clearest tested sequence is "think first, ChatGPT later." Wong and Qiu (2026) assigned N = 196 students to work independently, to use ChatGPT freely, or to work in a regulated condition: generate your own ideas, collaborate with ChatGPT to improve and evaluate them, then independently refine and submit one solution. The free-use group produced more creative work on the assisted task and fell back to human-only levels on a later, harder task done without ChatGPT. The regulated group showed no immediate advantage and outperformed both other groups on independent creativity afterward.

Prompt data explains the gap. In the regulated group, 88.6% of prompts were collaborative, and that was the only prompt type significantly correlated with later independent originality. In the free-use group, 70.9% of prompts were non-collaborative and 59.6% simply asked ChatGPT to generate ideas outright. The same tool, asked two different ways, produced two different learners.

On your assignment sheet this becomes an ordering instruction: ideas, hypotheses, or a rough draft first, then the tool. A sentence at the top of the page is enough, provided the next step collects what the student wrote.

Two classroom details decide whether the sequence holds. The first is time. If the independent phase happens outside class, some students will open the tool first and write the ideas down afterward. The second is collection. If nothing is handed in from that phase, the sequence exists on paper only. Half a page of notes, collected at the end of class, is enough to make the order real.

Configure the tool to hint instead of answer

Brcic and Frljic (2026) report the sharpest version of the placement argument. An unguarded AI helper left high-school students roughly 17% worse on an unaided exam than peers who had no tool at all. The same model rebuilt to withhold answers erased the harm, and a well-engineered tutor roughly doubled learning. Their diagnostic is usable in any course: if letting AI in makes the task feel effortless, the tool is in the wrong place.

Zohar, Bloom and Inzlicht (2026) give that argument a shape. They argue the link between effort and meaning is an inverted U, so the target is a gradient: remove overwhelming obstacles and preserve the struggles that produce comprehension and ownership, with assistance as a supplement. Their argument is a conceptual Comment with no new data, and the inverted-U relationship rests on a single empirical anchor without being quantified, so nobody can tell you where the optimum sits for a given learner.

Practically, that means guarded AI in the middle of the task, where hints, worked examples, and practice items do the work, and a secured check at the end. It also turns an AI-use policy into a per-skill placement rule, which is easier to write and easier to defend than a list of prohibitions.

A hint is a question about the student's next step, a worked example of a similar problem, or a pointer to the relevant passage. An answer is a finished sentence the student can paste. Most general-purpose chatbots produce the second by default, in fluent prose that reads like help, which is why the configuration matters more than the student's intentions.

Require the check, and make it visible

Verification is the cheapest lever with real causal support. Biswas (2026) found that making verification visible triggered a counter-cascade to near-complete verification: over-reliance fell to 0.00 and regret to 0.07. Reducing the friction of checking was the weakest lever in the same model, because the barrier is social rather than mechanical. A link to the library does not make a student check a claim.

Viberg and colleagues (2026) found the students who already do this well check output against course materials, peers, or instructors when accuracy is uncertain. That is a behavior you can require. Put one line in the assignment: what you checked, and against what. Read it as part of the grade, and the norm becomes visible to the whole class rather than to the conscientious few.

That line also gives you something to follow up. When a student writes that they checked a claim against the course text, you can ask which page. When they write that they compared with a peer, you can ask what the peer said. It costs a sentence. The follow-up question is what makes students treat it as work.

Verification also needs a scope. Checking that a claim matches the course text is a different act from rereading a sentence for grammar, and the first is where the model fails more often. Say which one the line asks for, or you will get proofreading.

Demonstrate one confident failure

Set a task whose answer the model gets confidently wrong, then have the class check that answer against the course text. It lands because the error arrives in the same voice as the correct answers. Add the evidence afterward: an unguarded helper left students roughly 17% worse on an unaided exam, and students given GPT-4 solved more problems with the tool and performed worse than controls once it was removed (PAUSE reports Bastani et al., 2025). The same source adds that Liu et al. (2026) found assistance reduced persistence in randomized controlled trials (N = 1,222).

Choose a failure that is modest and checkable, such as a citation that does not exist, a date that is wrong, or a calculation that does not close. Then ask the class to write the correction and the check they used to find it. That written correction is the artifact, and it tells you whether they can do the check without you standing over them.

Keep the graded moment tool-free

The assisted product is a poor proxy for capability, so at least one graded moment has to happen without the tool. Wong and Qiu's (2026) later unassisted task is that measurement. Their human-only group also declined on the harder follow-up, which rules out the simple version of the fix: removing the tool without supporting the learner is not an intervention.

Alternating modes has support. PAUSE reports Kosmyna et al.'s (2025) session-four result, where brain-only participants who later used ChatGPT outperformed sustained LLM users. Accountability works through explanation as well: Makransky et al. (2025) found that a tutoring chatbot which prompted students to connect ideas and explain their reasoning produced better assessment performance than traditional instruction.

For a course, alternate: assisted work in practice, unassisted work where it counts. Read the assisted piece as a record of performance under good conditions, and the unassisted piece as the evidence about learning. Comparing them is also the most useful conversation you can have with a student who cannot understand a mark.

Scheduling is the constraint. A tool-free moment needs a room, a time, and a task that fits in it, so the cheapest version is a short in-class piece attached to an assignment you already set. A separate exam is a larger commitment and usually not necessary for the first attempt at this.

The unassisted piece earns a second use as well: give it back at the start of the next unit, and both of you can see what survived the break.

Separate instrumental help from executive help

Viberg, Feldman-Maggor and Wong (2026) interviewed 20 STEM university students and found a four-stage process: deciding whether help is needed, choosing a source, choosing the type of help, and judging the help received. Stronger students favored instrumental help, meaning hints and explanations, over executive help, meaning direct solutions. The same study warns that using LLMs for debugging or cross-language programming can bypass independent problem solving even when students never ask for an answer.

That maps onto cognitive offloading cleanly. Instrumental use is the duplicative offloading that holds up when the tool goes away, because the student still encodes the reasoning. Executive use is the substitutive offloading that collapses when the external store is removed. Guo and Ye (2026) draw the boundary in exactly those terms: remove the store and substitutive offloaders decline severely, while duplicative offloaders keep their accuracy through internal encoding.

Write the distinction into the assignment. Ask students to label each AI use as instrumental or executive, require at least one collaborative prompt, and grade the reasoning they added after the response. In Wong and Qiu's free-use condition, the default drifted toward executive use without anyone deciding to drift.

Not all offloading is suspect. PAUSE cites Hong et al. (2025), where deliberately offloading lower-order writing tasks to free attention for analysis and revision produced larger critical-thinking gains. The test is whether the part being offloaded is the skill you are grading.

Those labels also diagnose your task. If most of a class reports executive use, the assignment is inviting it, whatever the policy says, and the next change belongs in the task rather than in the wording of the rule.

Time the prompt to the phase

Feedback that targets a stable belief works before a task starts. Feedback about immediate correctness and difficulty works during it. Guo and Ye (2026) call this timing-component matching, and it explains why a reflection due at the end of a project changes so little: by then the choices are made.

Lodge and Loble (2026) recommend integrated metacognitive prompts that make learners pause, reflect, and assess their understanding as they work, alongside Load Reduction Instruction, which manages cognitive burden while enabling progressive independence. A one-minute prediction before a hard unit, written down and compared afterward, fits that window and costs nothing to mark.

The wording of the prompt matters less than its position. Before a unit, ask what the student expects to be hard and why. During it, ask which step they are stuck on and what they have tried. An end-of-project reflection mostly records what happened, which is why it changes so little. Keep the prediction on the same page as the work, so the comparison at the end costs no filing.

Expect the risk to cluster

Lodge and Loble (2026) locate the risk in prior knowledge and self-regulation and name a metacognitive equity gap: using AI well requires resources that novices lack, so the students who most need the practice are the likeliest to delegate the learning itself. They report that 80% of Australian students already use AI and that two-thirds of early secondary teachers do (OECD 2025).

Age and habit correlate with the same pattern. Gerlich (2025) found participants aged 17-25 showed higher AI dependence and offloading and lower critical thinking than those aged 46 and over, and that attainment predicted better critical thinking regardless of AI use (r = +0.34), with a significant interaction indicating that attainment mitigates the negative effect. Gao, Sun and Khan (2026) found polychronicity, a tendency to multitask, moderates the pathway, with high-polychronicity students at greater risk.

For a section, that argues for watching the students with the weakest preparation most closely, and for checking whether your own task has created the conditions where checking stops. A demanding task paired with a capable model is the combination Biswas's model flags for the most regret, and it is easy to create by accident.

Context matters as much as person, and your course controls part of the context. Biswas (2026) shows that task difficulty and AI quality set the baseline: over-reliance rises from about 0.02 to 0.38 with difficulty, and on hard tasks it runs at 0.38 with poor AI against 0.16 with good AI. The highest regret comes from high-quality AI on hard tasks (0.441), because agents over-defer and rarely self-rely. Social proof compounds it: as visible use rises from 0 to 0.6, verification collapses from 0.29 to 0.002. Checking stops exactly where the model is strongest and the task is hardest, which is rarely where an instructor looks first.

Decide how much weight the evidence carries

Nothing in this corpus tests whether a specific over-reliance intervention holds across settings or semesters. Think-first rests on one experiment with N = 196. The withholding result comes from Brcic and Frljic's (2026) synthesis rather than from a trial of its own. The friction argument is a conceptual comment with no new data. Biswas (2026) states the limits of the model: exogenous stationary AI quality, a fixed network, stylized verification. Estimating the real values would need longitudinal traces of per-task verification rates, social-proof strength, and how trust updates after verified as against unverified use.

Measurement remains the weakest link. Survey designs dominate, and both Gerlich (2025) and Gao, Sun and Khan (2026) call for longitudinal and experimental follow-up. PAUSE has no psychometrics and was built for adults. Whether the equity gap can be closed by instruction is theorized rather than demonstrated.

That leaves you with a reasonable bet, made cheaply. Change one assignment, keep the unaided measure, and check whether the gap between assisted and unassisted work narrows. Do not rebuild a course on the strength of these numbers.

What instructors get wrong about this

The students who use it most are the ones in trouble. Frequency is what surveys capture and dashboards display, and it is not the construct. A student who checks every answer and cites your materials can use AI daily and be fine. A student who submits the first output once a week has the problem.

A ban removes the risk. The strongest results here come from placement, and the human-only arm of Wong and Qiu's experiment also declined on the harder follow-up. Removing the tool without supporting the learner leaves the difficulty where it was.

Cheaper checking will fix verification. Making the check easier was the weakest lever in Biswas's model, because the obstacle is social. A better plagiarism checker does not make a student read the source.

A usage dashboard tells me who is offloading. PAUSE's items record when and how often AI enters a workflow, and the authors flag the validity problem themselves. A careful scaffolder scores as an offloader, so acting on those numbers punishes the students you want to keep.

Fluency and satisfaction show the design is working. That is the illusion Lodge and Loble (2026) describe in their performance paradox: work feels fluent and students perform well in the moment while retaining less. Unaided performance and delayed transfer are the measures that answer the question.

Teach AI literacy to students and the problem goes away. Lodge and Loble (2026) caution that over-investing in student-facing AI literacy may be the wrong allocation, given what novices lack.

What will come back at you

"This is just good pedagogy with a new name." Partly. Scaffolding, formative feedback, and productive struggle are old ideas. What is newer is the ordering: visible verification beat cheaper verification, and the two moments worth protecting are the first hard attempt and the final unaided check. Do the familiar things in the new order.

"I cannot police it." You cannot, and the evidence says not to try. Placement produced the strongest result. Design the task so that using the tool at the wrong moment shows up in the work itself, as an answer the student cannot explain or a verification step that is missing.

"My course is too large for a sequenced unit." The cheap levers scale: a verification line, a one-minute prediction, and moving the graded check to an AI-free room. The expensive item, a fully sequenced think-first unit, can start as one assignment. Regenerate one problem set into assisted and unaided halves and see what the gap tells you.

"My students are novices. They need the help." They need hints more than answers, and the equity finding is the reason. The students with the least prior knowledge are the most likely to delegate the reasoning, and a tool that withholds answers keeps the practice where it belongs.

"Will the weaker students not just fail the unaided check?" Sometimes they will, and that result is information rather than a verdict. A tool-free check that much of a class fails says the practice was carrying the performance, which is the thing you needed to know. Weight it as evidence, reteach what it exposes, and offer a second attempt on a parallel task. No study here tests whether reteaching after a failed unaided check recovers the learning, so that part is your professional judgment.

"Some of my students use AI as an honest scaffold. Am I supposed to stop them?" No. A student who reads a hint, tries the step, and checks the result is doing the duplicative offloading that survives the tool being removed. What you want is a task that keeps that use available and makes the executive version less rewarding, which is a matter of what the task collects and what the rubric pays for.

"The evidence is thin, so why change anything?" It is thin in specific places, which the limits above name. What is consistent across the studies is the direction: placement, visible checking, and an unaided measure. One changed assignment, with the unassisted result kept, is a small enough bet to run on that basis.

What to do this week

1. Add a verification line to the next assignment: one sentence naming what the student checked and against what.

2. Move the graded check to a moment without the tool, and grade that one.

3. Put the first hard attempt before the tool, and collect the ideas, hypothesis, or draft that comes out of it.

4. Change the tool setting you recommend so it hints rather than answers, and tell students that is what you did.

5. Run one confident AI failure live, have students catch it against the course text, and keep the correction they wrote.

Where to go next

The neighboring question, whether AI produces learning at all, is covered in Does Using AI Actually Help My Students Learn?. How Can I Reduce AI Cheating in My Course? handles integrity and the limits of detection, and How Do I Redesign Assessment So That a Grade Still Tells Me Something Defensible About What the Student Knows or Can Do? covers the assessment architecture this page assumes. For the underlying ideas, see Cognitive Offloading, Metacognition, and Desirable Difficulties.