AI Participant, Not ConsumerWhat happens when a student lets the machine do the homework — and what has to be true of a system that doesn't let it
The framework has always distinguished the student who receives AI output from the student whose own thinking AI amplifies. That distinction is no longer only a philosophical one. There is now a preregistered randomized trial, peer-reviewed in PNAS, that tested both stances on about a thousand students and measured what each did to learning. This page is what it found, what it does not show, and the mechanism built in response.
A child who uses AI to finish homework is getting better marks and, on the best available evidence, learning measurably less — while feeling that they have learned more. That combination is what makes it hard to catch. It does not look like failure from inside.
The trial that settles the near-term question
Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman ran a preregistered classroom experiment in a large Turkish high school across the autumn of 2023–24. Roughly 1,000 students across about 50 classrooms, grades 9 to 11, 2,848 observations, randomized at classroom level. Four ninety-minute sessions covering about 15% of the semester's mathematics.
Each session ran the same way: teacher review, then assisted practice — the part that was randomized — then a closed-book, closed-laptop exam on conceptually parallel problems, graded by independent graders. Three arms: no AI, an ordinary ChatGPT-like interface on GPT-4, and a guardrailed tutor version.
The result, against a control practice mean of 0.284:
| Condition | Assisted practice | Unaided exam (the preregistered primary outcome) |
|---|---|---|
| GPT Base — ordinary chatbot | +0.137 = +48%, p<0.01 | −0.054 = −17%, p<0.05 |
| GPT Tutor — guardrailed | +0.361 = +127%, p<0.01 | −0.004, not significant |
Read the top row twice. In the same session, the same students got 48% better at practice and 17% worse at the exam. Performance and learning did not merely diverge — they moved in opposite directions, simultaneously, in the same ninety minutes.
The students did not perceive any reduction in their own learning. From the inside, it felt like it was working.
The interaction logs show the mechanism plainly. GPT Base users asked for solutions and copied them. GPT Tutor users asked for help and then attempted the answer themselves.
What the guardrail actually was
Not a personality. Three concrete engineering decisions:
One — hints only, one at a time, never the answer. Two — teacher-authored correct solutions injected into the prompt, so the model was not generating the mathematics. Three — teacher-written common-mistake feedback, so the system could respond to the specific error rather than re-explain the topic.
Those three moved the exam effect from −0.054 to −0.004. An order of magnitude.
Why nobody notices
The reason a parent cannot detect this by watching is that the signal a parent has access to — how well the work is going — is the signal that has been corrupted.
This was established well before AI existed. Deslauriers, McCarty, Miller, Callaghan and Kestin (2019, PNAS) ran a randomized crossover in Harvard introductory physics with N = 149, every student experiencing both conditions. Active learning against passive lecture:
| Measure | Effect of active learning |
|---|---|
| Feeling of learning | −0.56 SD — they felt they learned less |
| Test of learning | +0.46 SD — they actually learned more |
The mediator was fluency. Perceived instructor smoothness predicted felt learning at +0.51 SD, and once fluency was in the model, the type of instruction no longer predicted felt learning at all. Fluency was doing all the work. And students with less prior expertise were worse calibrated — the novices, the ones who most need the signal, had the least reliable one.
A second study shows the gap opening in a live AI setting. Fan and colleagues (2025, British Journal of Educational Technology) gave 117 university students either ChatGPT, a human expert, a checklist tool, or nothing. The ChatGPT group's essays improved by 3.600 against 1.630 for control — outperforming even the human expert at 1.480. Knowledge transfer told a different story: F = 0.019, p = 0.996, η² = 0.000. Group means were 0.785, 0.782, 0.771, 0.777. The output improved dramatically. Nothing transferred.
What the AI Consumer stance actually costs
Three further findings fill in the shape, and each carries a real limitation.
The compounding question. Strömberg, Lei and Wu (CEPR Discussion Paper 21577) tracked 26,811 Chinese secondary students, grades 7–12, using staggered adoption. Homework scores rose 18% and completion time fell 30%; monthly exam scores fell 20% within six months; entrance exams fell 18% and 24% — with the full penalty emerging only after about two years, concentrated among the roughly 80% of users whose pattern looked like homework outsourcing. This is a working paper, not peer-reviewed, and unreplicated. Its most useful implication if it holds is methodological: anything measuring under two years is measuring too early.
The disclosure gap. Mak and Walasek (2025, peer-reviewed, preregistered) analysed 4,820 empirical reports by 2,000 psychology undergraduates across ten years at one UK university. AI lexical markers surged after 2022, peaked in 2024, then declined in 2025 — which the authors suggest indicates masking rather than abstention. Writing quality, indexed by grades and feedback, showed no discernible shift. And zero students disclosed AI use. Not a low number. Zero.
The usage baseline. Pew's probability-based panel of 1,458 US teens (February 2026) found 54% have used chatbots for schoolwork — up from 13% in 2023 and 26% in 2024 — and 10% say they do all or most of their schoolwork with chatbot help.
The parallel from fields that have been here longer
Aviation and medicine have been delegating skilled work to automation for decades, and both show the same shape.
Budzyń and colleagues (2025, Lancet Gastroenterology & Hepatology) examined 19 endoscopists, each with over 2,000 prior procedures and a mean of 28 years' experience, across four Polish centres. In their non-AI colonoscopies, adenoma detection fell from 28.4% to 22.4% after routine AI exposure — an absolute drop of 6.0 percentage points (95% CI −10.5 to −1.6, p = 0.009), adjusted OR 0.69.
Haslbeck and Hoermann (2016, Human Factors) had 126 randomly selected airline pilots fly a manual raw-data approach in a full simulator. Ten of 57 approaches exceeded stabilized-approach or licensing limits — all of them on the long-haul fleet, none on short-haul. Their conclusion isolates the variable precisely: recent flight practice predicted fine-motor performance better than total experience or type rating.
And what nobody knows
There is no evidence — none — on students who leaned on AI in school and then went to college, or to work, or into adult life. ChatGPT launched in November 2022. The longest observation window anywhere in this literature is about thirty months, in one country, in an unreplicated working paper. The longest delayed post-test in higher education is 45 days.
The framework's reading is that a decade of consumer-stance use during the years when the instrument is still forming will prove costly. That is a prediction, and it is labelled as one here. It is not a research finding, and this page will not dress it as one.
Where AI tutoring has worked — and the pattern inside it
The counter-evidence is real and must be carried honestly, because it contains the most useful finding on the page.
Kestin, Miller, Klales, Milbourne and Ponti (2025, Scientific Reports) ran a randomized crossover in Harvard physical sciences, 194 students, AI tutor against in-class active learning. Median learning gain 1.75 against 0.75 — over double — at z = −5.6, p < 10⁻⁸, standardized coefficient 0.63, in less time (median 49 minutes against 60).
A World Bank randomized trial in Benin City, Nigeria (De Simone et al., 2025, working paper) found +0.31 SD overall and — the important number — +0.21 SD on the regular third-term curricular English exam, taken on paper without AI. That is genuine transfer, which almost nothing else in this literature demonstrates.
Now look at what those two had in common, because it is the whole argument:
| Trial | What held the line |
|---|---|
| Kestin (Harvard) | Instructor-authored step-by-step solutions injected into the prompt for every question — the authors deliberately did not let GPT-4 generate the physics. Scaffolding enforced in software, because they found the system prompt could not reliably do it. Months of build time. |
| World Bank (Nigeria) | Teacher introduces the topic; a teacher-built prompt library explicitly designed around desirable difficulties and retrieval practice; students working in pairs; monitors checking fidelity. |
| Tutor CoPilot (US, Title I) | The AI never touches the student — it coaches the human tutor. 900 tutors, 1,800 students, +4 points of topic mastery; students of less-experienced tutors gained up to 9. |
In every trial where the gains were large, the AI was never working alone and never improvising. These are structured programmes with AI inside them — not AI tutors.
How JyoLing Academy is built, and why it is built that way
The three things that removed the harm in the PNAS trial — hints instead of answers, human-authored solutions rather than model-generated ones, and named common mistakes — are not features that were added to JyoLing Academy after reading the study. They are the architecture the system was already built on, for the reason the study demonstrates: an AI that answers is not a tutor, it is a shortcut with a friendly voice.
Here is the actual mechanism.
The explanations are written by humans, not generated
Every subtopic opens with a foundation file — human-written, served before any video. The authoring standard requires the writer to first identify who currently teaches that topic best anywhere on the internet, study their sequence and their canonical worked example and where they slow down to name the trap, and then answer one question in writing: what does our version do that theirs does not? "More detailed" and "more engaging" are rejected as answers.
The bar is explicit: better than the best explanation this student could find free on the internet in ten minutes. If they can search their way to something clearer, the file has failed.
Why it matters: this is the second guardrail condition from the trial. The model is not the source of truth about the subject. It cannot hallucinate the mathematics, because it was never asked to produce the mathematics.
Every wrong answer is tagged to a named error
Each concept declares a list of misconception codes — MC_GAS_HAS_NO_MASS and its kind. Every distractor in every question is tagged to one. Every code maps to a hint written to name the error the student made, not to supply the next step.
The rule governing those hints is strict, and it exists because of a specific failure: a hint that only makes sense for the item it was written on will, when the same code appears on a different item, walk the student directly into a distractor — teaching the very error it was meant to diagnose. So each hint must survive the test: if this code were attached to a different item in the same concept, would this sentence still be true? If it needs that item's numbers, direction, or sign, it is not a hint. It is a solution step, and it belongs somewhere else.
Why it matters: this is the third guardrail condition, built as infrastructure rather than as a prompt. A miss is diagnostic. The system responds to the specific misunderstanding rather than replaying the lesson.
The sequence is enforced in code, not requested in a prompt
Curriculum is a directed acyclic graph — topic to subtopic to concept, each with declared prerequisites, validated for cycles before anything ships. A concept does not unlock until its prerequisites are complete. A diagnostic warmup, sampling from the topics on the student's chosen goal path and tagged with the same misconception codes, decides where the student starts.
Why it matters: Kestin's team found the same thing at Harvard and reported it plainly — the system prompt "could not reliably provide enough structure to scaffold problems with multiple parts," so the platform had to do it. Scaffolding that lives in a prompt is a request. Scaffolding that lives in the code is a constraint.
The author of the content is never its auditor
Content is written by one agent and adversarially checked by a separate one briefed to find errors rather than confirm quality. This exists because of a documented incident: on 16 August 2026, thirty-three foundation files were written, and every one passed its validator with zero errors and zero warnings. A separate reviewer then found two real defects in a trigonometry file — an explanation misnaming a ratio, and a distractor attributed to the wrong arithmetic error. Both answer keys were correct. What was wrong was the attribution of a distractor to a named misconception — which is exactly the thing the product claims to do better than anyone, and therefore the worst possible place to be sloppy.
Why it matters: no validator catches a misconception attached to the wrong item. Every referential check stays green while the student is taught the wrong error. Only a reader looking for blood finds it.
What this does and does not entitle anyone to claim
The honest statement is narrower than a marketing one, and more useful.
What can be said: the three conditions that removed the measured learning harm in a preregistered randomized trial of about a thousand students are present in this architecture, and present as enforced structure rather than as instructions to a model. That is a design-alignment claim, and it is verifiable by reading the system.
What cannot be said: that this produces better learning outcomes. JyoLing Academy has not been through a randomized trial. When it has, the result will be published here whichever way it falls. And the ceiling on what such a trial could show is set by the evidence above: in the PNAS study, guardrails removed the deficit without producing a gain. The largest, longest independent trial of a coach-not-answer tutor found 0.06–0.08 SD per year and named engagement, not model quality, as the binding constraint.
That last finding is the one worth taking most seriously, because it is the hardest. Students skipped the tutor in 83% of the sessions in which they had just made a mistake. A system can be built correctly and still lose at the moment of the miss, because being wrong is uncomfortable and closing the tab is easy. Misconception coding is a partial answer — a hint that names your specific error is a different experience from a hint that restates the lesson — but nobody should pretend the problem is solved by architecture alone.
The harm is a property of interface design, not of AI. The same model, in the same classroom, in the same ninety minutes, produced a 17% deficit or none at all — depending entirely on whether it handed over the answer.
What a parent can actually do
Four things the evidence supports, and one it does not.
Watch the unaided work, not the finished work. The finished work is the corrupted signal — it improved 48% in the arm that learned least. What discriminates is performance when the tool is absent: a closed-book attempt, an explanation given aloud, a problem worked without a device in the room.
Distrust fluency, including your child's. They are not being dishonest when they say they understand. In a Harvard crossover, students felt they learned less in exactly the condition where they learned more — and the least experienced students were the worst calibrated.
Treat "asked for an explanation" and "asked for the answer" as different activities. They are the difference between the two arms of the trial. It is a question about the interaction, not about screen time.
Expect the mistake to be the moment that matters. The two-year trial found students skipping help in 83% of the sessions where they had just erred. Whatever support exists has to be present, and tolerable, at that exact point.
And what the evidence does not support: banning it. Nothing here shows that students without AI access outperform students with a well-designed one — the guardrailed arm matched the no-AI arm, and in Nigeria and at Harvard, structured programmes with AI inside them beat conventional instruction. The finding is about design, not abstinence.
Questions people are actually asking
Does using ChatGPT for homework hurt learning?
In the one preregistered randomized trial that tested it, yes. About 1,000 Turkish high school students practised maths with an unguarded GPT-4, a guardrailed tutor, or no AI, then sat a closed-book exam with no AI. The unguarded group scored 48% higher during practice and 17% lower on the unaided exam than students who had no AI at all.
Why do students not notice?
Because performance during practice is a poor guide to learning, and often an inverted one. In a randomized crossover at Harvard with 149 students, active learning lowered how much students felt they had learned by 0.56 SD while raising how much they actually learned by 0.46 SD. AI makes practice feel fluent, and fluency is what students read as understanding.
Do AI tutors actually help?
Sometimes, and never when the AI works alone. In every trial with large gains, something outside the model held the line — instructor-written solutions injected per question and scaffolding enforced in software, or a teacher-built prompt library and paired students. The largest and longest trial, across two years and 18 schools, found only 0.06 to 0.08 SD per year.
What happens when these students reach college?
Nobody knows, and anyone who says otherwise is guessing. No study follows students who relied on AI in school into college or work. The longest observation window is about 30 months, in an unreplicated working paper. The longest delayed post-test in higher education is 45 days. Any claim about the long run is a prediction.
How is JyoLing Academy different?
The explanations are human-written rather than model-generated, so the AI is not the source of truth. Every wrong answer is tagged to a named misconception, and the hint names the error rather than supplying the step. Concepts unlock along a prerequisite graph enforced in code, not by prompt. Those are the three conditions that removed the harm in the PNAS trial — built as structure rather than instruction. It has not yet been through a randomized trial of its own.