The research report Think and Create with AI (6 November 2026), on one page
We teach thinking before the AI is switched on
Students who used AI first scored higher in practice and lower on the exam. Here are twelve studies, their numbers and their limits, and why every CIT lesson moves from thinking to making and improving.
Every number comes from a published study, and every card states what that study does not show.
E1
Unrestricted AI raised practice scores and lowered exam scores
Practice and exam scores versus the control group
Students with unrestricted ChatGPT-style access scored 48% higher on practice problems, then 17% lower on the no-AI exam than students who never had AI. With a tutor version built to give hints instead of answers, practice scores rose 127% and the exam harm disappeared.
What this study does not show The tutor version prevented the harm but did not raise exam scores. One high school in Turkey, mathematics, four 90-minute sessions.
Population
Nearly 1,000 grade 9 to 11 students at one Turkish high school, randomized controlled trial
Unrestricted AI raised practice scores and lowered no-AI exam scores (Turkish high-school maths RCT, about 1,000 students).
A hints-only tutor prevented that harm; it did not raise exam scores.
Do not say
Generalising to all AI use ("AI lowers grades by 17%")
"A hints-only tutor raises grades" (exam scores did not differ)
Any claim about CIT students until CIT has measured it
E2Preliminary
Brain first, then AI, stayed more engaged than AI first
Brain first, then AIhigher recall and activation
AI first, then brain onlystayed under-engaged
54Participants
Brain connectivity measured by EEG during essay writing fell as reliance on external tools rose. When conditions were swapped in a fourth session, the group that had written unaided first and then added AI showed higher recall and brain activation; the group that had started with AI stayed under-engaged.
What this study does not show 54 participants (18 in the fourth session), one essay-writing task, an arXiv preprint not yet peer-reviewed. It does not show brain damage, lower intelligence or permanent change.
Population
54 participants in three groups (brain-only, search engine, LLM); 18 completed the fourth session
An MIT Media Lab preliminary study found differences in brain activity and recall depending on the order in which AI was used (54 participants, essay task).
The group that wrote unaided first and then added AI stayed more engaged.
Do not say
"AI damages the brain", "makes you dumber", "brain damage", "permanent"
"MIT proved" (a preprint, not peer-reviewed)
Applying it directly to children
E4
A tutor built on pedagogy beat an active-learning lecture
2x+x or moreMedian learning gain (AI tutor ÷ active lecture)
In a Harvard physics course, median learning gains with a purpose-built AI tutor were more than double those from the same material taught as an in-class active-learning lesson.
What this study does not show 194 undergraduates and a tutor the researchers built on pedagogy principles. The authors write that an AI tutor should not replace in-person teaching.
In a Harvard physics experiment (194 students), a tutor built on pedagogy principles more than doubled median learning gains versus active lecture.
The authors say it should not replace in-person teaching.
Do not say
"Studying with ChatGPT doubles learning" (a custom tutor, not a generic chatbot)
"CIT lessons double learning"
E5
A structured, teacher-supervised AI programme gained 1.5 years of learning
0.3SDOverall assessment
In nine public senior secondary schools in Edo State, Nigeria, students who studied with GPT-4 after school for six weeks, in pairs and under teacher supervision, scored about 0.3 standard deviations higher overall and 0.24 in English: nearly 1.5 years of typical learning, outperforming around 80% of education interventions evaluated by randomized trials in developing countries.
What this study does not show The gain came from the whole package: designed prompts, teacher facilitation and pair work, not from a chatbot alone.
Population
Nine public senior secondary schools in Edo State, Nigeria; six-week after-school programme; randomized
In a World Bank trial in Nigeria, a structured six-week GPT-4 programme under teacher supervision gained about 0.3 SD, nearly 1.5 years of typical learning.
Do not say
"1.5 to 2 years" (the source says nearly 1.5)
"A chatbot alone gained 1.5 years" (teachers, structure and pairs were part of it)
See all twelve studiesShow fewer
E3
The more people trusted the AI, the less they checked it
319Respondents
Across 936 real AI use cases reported by 319 knowledge workers, higher confidence in the AI went with less critical thinking, while higher confidence in one's own ability went with more.
What this study does not show A self-report survey of working adults, not an experiment with students; it shows association, not cause.
In a survey of 319 knowledge workers, more trust in the AI went with less checking of its output.
Do not say
"AI destroys critical thinking" (association, not cause)
Presenting it as student data
E6
Teaching students to plan and check their own learning adds about 8 months
Teaching students to plan, monitor and evaluate their own learning (metacognition and self-regulation) is rated at +8 months of additional progress on average in the EEF Teaching and Learning Toolkit (355 studies). The guidance report's second edition, published 13 November 2025, gives recommendations in seven areas.
What this study does not show The toolkit itself says this impact can be difficult to realise in practice; it is the average when teachers explicitly teach and model the strategies.
Population
EEF Teaching and Learning Toolkit, 355 studies (review updated May 2025)
The EEF toolkit rates metacognition and self-regulation at +8 months of additional progress on average (355 studies).
Always with the caveat: when taught well.
Do not say
"+7 months" (the toolkit, updated May 2025, says +8)
Promising 8 months to every student
E9
Employers rank analytical thinking as the top core skill
Employers surveyed by the World Economic Forum named analytical thinking (69%), resilience, flexibility and agility (67%) and creative thinking (57%) as core skills for 2025, and expect 39% of workers' core skills to change by 2030.
What this study does not show An employer survey (1,043 employers representing 14.1 million workers), not a study of student outcomes.
In the WEF Future of Jobs Report 2025, 69% of employers named analytical thinking a core skill.
Do not say
"70%" (the report says 69%; the digest says seven in ten)
Presenting it as a study of student outcomes
E10
AI textbooks rolled out without pedagogy lost more than half their schools in one semester
4,095 Korean schools adopted AI digital textbooks in the first semester of 2025; by the second semester the number had fallen 58.8% to 1,686. The drop followed the National Assembly's amendment of 4 August 2025 that reclassified them from official textbooks to supplementary materials.
What this study does not show Not a criticism of the ministry or schools. We keep it as a lesson that AI tools without lesson design rarely stay in classrooms. Figures are Ministry of Education data via Rep. Baek Seung-a's office; a different tally (provincial offices, 4,146 to 2,095) also exists.
Population
All Korean schools, Ministry of Education data (October 2025 parliamentary audit)
Schools using AI digital textbooks fell from 4,095 to 1,686 between the two semesters of 2025 (Ministry data via Rep. Baek's office).
The lesson is lesson design over tools.
Do not say
Criticising the ministry, schools or teachers
Rounding to 59% (it is 58.8%)
Quoting the numbers without the source (other tallies exist)
E11Quote
Critical thinking cannot be taught apart from domain knowledge
The processes of thinking are intertwined with the content of thought (that is, domain knowledge).
Cognitive psychologist Daniel Willingham explains that critical thinking is not a stand-alone skill: the processes of thinking are intertwined with domain knowledge. That is why CIT teaches thinking on real content such as Python, data and AI.
What this study does not show A review essay by a cognitive psychologist, not a new experiment.
Critical thinking cannot be taught apart from domain knowledge (Willingham, 2007), so thinking is taught on Python, data and AI content.
Do not say
Describing any CIT course as "thinking skills only"
E12Opinion
Move grading into the classroom; verification matters more
The verification ability is especially important in the case of AI, which is presently a lot more fallible in a great variety of ways compared to calculators.
AI researcher Andrej Karpathy, summarising advice he gave a school board, wrote that AI use in homework cannot be detected, so most grading has to move into the classroom, and that verification matters especially because AI is far more fallible than a calculator.
What this study does not show A practitioner's opinion, not a study.
Karpathy argues for moving assessment into the classroom and building verification skills (an opinion post).
Do not say
"Research shows" (it is an opinion)
"OpenAI said" (a personal post)
For staff: safe to say and do not say
Organised by card. The copy button always includes the caveat.
ID
Safe to say
Do not say
E1Unrestricted AI raised practice scores and lowered exam scores
Unrestricted AI raised practice scores and lowered no-AI exam scores (Turkish high-school maths RCT, about 1,000 students).
A hints-only tutor prevented that harm; it did not raise exam scores.
Generalising to all AI use ("AI lowers grades by 17%")
"A hints-only tutor raises grades" (exam scores did not differ)
Any claim about CIT students until CIT has measured it
E2Brain first, then AI, stayed more engaged than AI first
An MIT Media Lab preliminary study found differences in brain activity and recall depending on the order in which AI was used (54 participants, essay task).
The group that wrote unaided first and then added AI stayed more engaged.
"AI damages the brain", "makes you dumber", "brain damage", "permanent"
"MIT proved" (a preprint, not peer-reviewed)
Applying it directly to children
E3The more people trusted the AI, the less they checked it
In a survey of 319 knowledge workers, more trust in the AI went with less checking of its output.
"AI destroys critical thinking" (association, not cause)
Presenting it as student data
E4A tutor built on pedagogy beat an active-learning lecture
In a Harvard physics experiment (194 students), a tutor built on pedagogy principles more than doubled median learning gains versus active lecture.
The authors say it should not replace in-person teaching.
"Studying with ChatGPT doubles learning" (a custom tutor, not a generic chatbot)
"CIT lessons double learning"
E5A structured, teacher-supervised AI programme gained 1.5 years of learning
In a World Bank trial in Nigeria, a structured six-week GPT-4 programme under teacher supervision gained about 0.3 SD, nearly 1.5 years of typical learning.
"1.5 to 2 years" (the source says nearly 1.5)
"A chatbot alone gained 1.5 years" (teachers, structure and pairs were part of it)
E6Teaching students to plan and check their own learning adds about 8 months
The EEF toolkit rates metacognition and self-regulation at +8 months of additional progress on average (355 studies).
Always with the caveat: when taught well.
"+7 months" (the toolkit, updated May 2025, says +8)
Promising 8 months to every student
E7Struggling first, then being taught, deepens understanding
Problem-solving before instruction had an effect size of 0.36 on understanding, up to 0.58 with high fidelity (53-study meta-analysis).
It does not suit younger elementary pupils.
Rounding to "about 0.4" (it is 0.36)
Claiming it works for younger elementary pupils
E8Growth-mindset training has a small effect. We do not sell it
Growth-mindset training has a small average effect on achievement, d = 0.08 (43 effects, 57,155 students), so CIT does not sell it.
"Change the mindset and grades rise"
"It has no effect" (small, not zero)
E9Employers rank analytical thinking as the top core skill
In the WEF Future of Jobs Report 2025, 69% of employers named analytical thinking a core skill.
"70%" (the report says 69%; the digest says seven in ten)
Presenting it as a study of student outcomes
E10AI textbooks rolled out without pedagogy lost more than half their schools in one semester
Schools using AI digital textbooks fell from 4,095 to 1,686 between the two semesters of 2025 (Ministry data via Rep. Baek's office).
The lesson is lesson design over tools.
Criticising the ministry, schools or teachers
Rounding to 59% (it is 58.8%)
Quoting the numbers without the source (other tallies exist)
E11Critical thinking cannot be taught apart from domain knowledge
Critical thinking cannot be taught apart from domain knowledge (Willingham, 2007), so thinking is taught on Python, data and AI content.
Describing any CIT course as "thinking skills only"
E12Move grading into the classroom; verification matters more
Karpathy argues for moving assessment into the classroom and building verification skills (an opinion post).
"Research shows" (it is an opinion)
"OpenAI said" (a personal post)
What we do not oversell
Effect sizes on one scale, so large reads as large and small reads as small.
Effect size (0 = no difference). Hedges' g and Cohen's d read on the same scale.
E7
Struggling first, then being taught, deepens understanding
g = 0.36Effect size (conceptual, transfer)
Letting students attempt a problem before the concept is taught (productive failure) beat teach-then-practice on conceptual understanding and transfer by an effect size of 0.36 (Hedges' g); studies faithful to the design principles reached 0.37 to 0.58.
What this study does not show A meta-analysis of 53 studies, 166 comparisons and 12,000+ participants. For grades 2 to 5 the direction reversed: instruction first worked better.
Population
Meta-analysis of 53 studies, 166 comparisons, 12,000+ participants
Problem-solving before instruction had an effect size of 0.36 on understanding, up to 0.58 with high fidelity (53-study meta-analysis).
It does not suit younger elementary pupils.
Do not say
Rounding to "about 0.4" (it is 0.36)
Claiming it works for younger elementary pupils
E8Not sold
Growth-mindset training has a small effect. We do not sell it
d = 0.08Effect size (achievement)
Growth-mindset interventions moved academic achievement by an average d = 0.08, a small effect.
What this study does not show 43 effects, 57,155 students. CIT does not sell mindset training; this card is here as an example of what we do not oversell.
Today you do not open the AI when the problem arrives. Take five minutes and write your approach on paper; wrong is fine. In a study of about a thousand Turkish high-school students, those who used AI from the start scored higher in practice but 17% lower on the exam they took alone. Thinking first removed that loss. Five minutes, go.
R2
Predict-Then-Query
When Before asking the AI anything
How Write your predicted answer, then ask. Underline where the AI differs and settle which one is right.
One line before you ask: 'I think the answer is...'. Then ask the AI and underline every difference. In a survey of knowledge workers, the more people trusted the AI, the less they checked it. A written prediction gives you something to check against.
R3
Assumption Audit
When Early in a project, before the plan is fixed
How List three assumptions the result depends on and ask, one by one, what breaks if each is wrong.
Before the plan: what has to be true for this result to hold? Write three things. Now flip each one. What if the data is biased? What if the measurement is off? Critical thinking is not a separate skill; it happens on the content in front of you, so we do it on your project.
R4
Five Whys / Inversion
When When a result is wrong or the work is stuck
How Ask why five times in a chain, or invert it: write how to fail for certain, then do the opposite.
When you are stuck we do not fetch the answer yet. Why is it wrong? And why is that? Go five deep. Fifty-three studies found that students who struggled first understood the concept more deeply once taught. This is why the struggle is not wasted time.
R5
Explain-to-a-10-Year-Old
When When a concept feels finished
How Explain it in 90 seconds with no technical words and record it; where you stall is what you do not yet know.
Understood it? Then explain it to a ten-year-old in 90 seconds, no technical words. Checking your own learning like this is the habit the EEF rates at about eight months of extra progress. Recording starts now.
R6
Verify-or-Reject Log
When Every time the AI returns something
How Log accept or reject, one line of reason, and how it was checked. AI output without a log entry is not submitted.
You do not paste what the AI gave you. Fill three boxes: accept or reject, why, and how you checked. Today's AI is wrong far more often than a calculator. Verification is the real skill you take away.
R7
Brain-Only Retrieval Check
When At the end of a unit and in the first ten minutes of a lesson
How With no tools, write the key points of the last unit on paper. It finds the gaps; it is not a grade.
Laptops closed, ten minutes. Write what you learned last week on paper. In a preliminary MIT study, the group that recalled unaided first and used AI later stayed more engaged. What you could not write is what we study today.
R8
Reflection Log
When The last three minutes of a lesson
How One line each for today's plan, check and evaluation; the next lesson starts from these three lines.
Last three minutes, three lines: what I planned, what actually happened, what I change next time. A synthesis of 355 studies says students who plan and check their own learning make more progress. Next lesson starts from these lines.
R9
Mini-Viva
When At every project deadline and every two weeks
How A five-minute question session: explain one line of your own code and say what happens if it changes.
Today is a viva. I pick one line of your code; you tell me what it does and what happens if I delete it. Work done at home may have had AI help, which is why assessment belongs in the classroom. If you can explain it, it is yours.
R10
Build-to-Finish
When At the end of every unit and in the last week of every project
How Use what was learned to finish one working thing, however small: a program, a model or a report, taken to the point where someone else can try it.
Today does not end with the explanation. You use what you learned to finish one small thing, and it does not have to be perfect. In a meta-analysis of 53 studies, students who struggled with their own attempt first and were taught afterwards understood the concept more deeply. Thinking does not grow apart from content; it grows on what you make. Write in one line what you will build, and start.
R11
Second-Version Rule
When After the first finished version, before it is presented or submitted
How Try it yourself or hand it to one classmate, find what to fix, pick one thing and make a second version. One line in the verify-or-reject log says what changed and why.
Finishing is not the end. For the next ten minutes, hand your work to the person next to you and write down where they get stuck. Pick one of those, fix it, and hand in a second version. Teaching students to plan, monitor and evaluate their own work is rated at about eight months of extra progress in the EEF toolkit (355 studies), with the caveat that this is the average when it is taught well. So one line in your log says what you changed and why. Go.
Seven AI+X studios and two tracks
The seven studios share one capstone pattern: define, build, verify, improve, defend.
Define · Build · Verify · Improve · Defend The capstone pattern every studio shares: define the problem, build, keep a verification log, improve it on feedback, defend it in a viva.
A research report that answers one school-subject question with data
Tools
Python, pandas, 과목 데이터
Fits
Korean STEM, International
Hours
12h
International track and Korean STEM track
International track
IB / AP / IGCSE Computer Science
EC portfolio
IAI²O (AIIC, AI4SCI, AI4BIZ)
USAAIO stays off the map until eligibility for Seoul-based students is confirmed.
Korean STEM track
Science and gifted-school research
KOAI rounds 2 and 3
Subject-data studio
KOAI (Korean AI Olympiad)
The bridge both tracks cross. KOAI round 1 is prepared on the same Python and machine-learning base by students of both tracks; round 3 leads to IOAI selection.
Shared spine: thinking rituals, Python, AI tech
How CIT differs
Six rows side by side, then the nine things CIT does not do, each tied to its evidence.
CIT
A typical coding academy
Evidence
When AI comes in
CITThink and predict first, then ask the AI.
A typical coding academyAI is banned, or the work starts with AI.
Banning AI did not work, and handing the work to AI from the start did not either. Across the studies we collected, what separated the outcomes was not whether students used AI but the order in which they did: think and predict first, then ask, then verify what comes back, then build something with it, improve it on feedback, and finally explain it in your own words. Every CIT lesson repeats those six steps, every unit ends in something the student made, and assessment happens in the classroom without tools. Every number on this page comes from a published study, and next to each one we note what that study does not show.
Everything this page cites. Accessed 6 November 2026.
What the research shows
E1 Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://www.pnas.org/doi/10.1073/pnas.2422633122 · open access (PMC)
E2 Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I. & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872 (preprint). https://www.media.mit.edu/publications/your-brain-on-chatgpt/ · arXiv
E4 Kestin, G., Miller, K., Klales, A., Milbourne, T. & Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15. https://www.nature.com/articles/s41598-025-97652-6
E8 Sisk, V. F., Burgoyne, A. P., Sun, J., Butler, J. L. & Macnamara, B. N. (2018). To What Extent and Under Which Circumstances Are Growth Mind-Sets Important to Academic Achievement? Two Meta-Analyses. Psychological Science, 29(4), 549-571. https://doi.org/10.1177/0956797617739704