The helicopter and the mountain: AI should raise the summit, not lower the bar

1Two voices, one frame

This month, two very different voices crossed my feed.

The mathematician

The first was Hugo Duminil-Copin, Fields Medal 2022, on France Inter. He was promoting De l'autre côté du tableau noir, the book he wrote with Geneva psychologist Emmanuel Sander.

• His image: letting AI solve a maths problem for a child is like asking a helicopter to drop you at the summit.
• His verdict: nice photo, zero effort. The only thing the child learns is how to get into a helicopter.
• The echo: La Libre Belgique used the same image on 18 September: being airlifted to the top of Everest instead of walking up.¹

The CEO

The second was Jensen Huang, CEO of NVIDIA.

• The finding put to him: children who lean on AI get their homework done faster but learn less.
• His answer: he agreed. Long division, multiplication tables and square roots are being forgotten.
• His follow up: does it matter? Probably not those skills, he said. We will discover new ones.

Where I land

I think both are partly right, and both miss the point. I spent years on school improvement projects at McKinsey and a good part of my career since building AI. Here is my position:

• The helicopter is not the problem. Flying one is far harder than walking. The problem is that we still grade children on the walk.
• The bar is now too low, not too high. With AI in every pocket, the mountain we ask children to climb should be much higher.
• Education systems are structurally bad at step changes. They were slow on calculators and slow on the internet. AI is a bigger step than both, and it is arriving faster.
• The binding constraint is teachers, not students. Children will use AI whatever we decide. The question is whether the adult in the room can fly better than they can.

The rest of this piece backs each point with data: eight randomised trials, surveys of teachers in 55 school systems, and fifty years of history.

2The mathematician's case: the value is in the climb

The argument deserves to be steelmanned before it is challenged.

Who is speaking

• Hugo Duminil-Copin, born 1985, is a French probabilist at the University of Geneva and the IHES.²
• The Fields Medal, on 5 July 2022, rewarded his work on phase transitions in models such as percolation and the Ising model.³
• His new book with Emmanuel Sander (Seuil, 2026) attacks the myth that mathematics is reserved for a handful of geniuses.⁴ Behind the rigour of theorems, it argues, sit intuition, analogy, trial and error, emotion and productive mistakes.
• Sander researches mental representations and how learners interpret problems.

What he says

In the France Inter interview, his core points are:

• Transferable capability, not the answer. Today's mountain is the exercise. There will be another tomorrow. School exists to equip children for mountains they have not seen yet.
• AI short circuits the core of mathematics: building concepts and tools yourself, and finding your own way to a solution.
• Mathematicians produce tomorrow's intuitions. Zero and negative numbers were once frontier ideas. Today they are common sense.
• His own practice. He told Le Monde he does not use AI in his research. The host called that almost anachronistic, given how widely the French now use AI.

He is not alone

La Libre reports that Mickaël Randour, head of mathematics at UMONS, signed the Leiden Declaration, a warning from mathematicians about AI misuse in research.^1,5

• His point: AI proofs written in the formal language Lean can be checked as correct without anyone understanding why they are true. Humanity learns nothing reusable from them.
• His rule: the brain is a muscle, and mental arithmetic comes before the calculator.

The strongest version of the argument

It is not nostalgia. It is that understanding is built through productive struggle, and a tool that removes the struggle can remove the learning with it.

That concern is real. Section 4 shows it is partly supported by data.

The counter voice in the same article

Amaury Hayat, a mathematician working on AI for proofs, disagrees.¹

• His view: doing maths with only pen and paper no longer makes much sense, and researchers who refuse AI will struggle to find posts.
• His metaphor: AI for a proof is less a helicopter ride than climbing Everest in a robotic exoskeleton instead of with an ice axe. Even an assisted ascent teaches you the route.

3The CEO's case: some skills stop mattering

The context

Jensen Huang spoke on The Ezra Klein Show (New York Times, 23 September 2026).⁶ Klein opened with a large Chinese study on AI in school and asked Huang what he made of it.

The study behind the question

Researchers from Stockholm University and the University of Hong Kong followed 26,811 students in grades 7 to 12, in one county in central China, for 30 months, as tools such as DeepSeek and Doubao spread.⁷

• Homework got better. Homework scores rose 18% and homework time fell 30%.
• Exams got worse. Monthly closed book exam scores fell 20% within about five months.
• The stakes are real. Scores on two high stakes entrance exams fell 18% and 24%. The full penalty took about two years to appear.
• The detail nobody quotes: most of the loss came from what the authors call homework outsourcing, very short homework time with very high homework scores. Students who kept spending normal time on homework lost far less.

How the tool was used mattered as much as whether it was used.

What Huang said

• He agreed with the finding. Long division, multiplication tables and square roots are being forgotten.
• He asked whether it matters, and answered that he did not think it does. Some skills matter, he conceded, but society will discover new ones.
• His illustration: he does not know his own address, ZIP code or phone number, and once panicked at a petrol pump when asked for his ZIP code.
• After pushback from Klein, he framed the trade as losing some fine intellectual dexterity while becoming better systems thinkers.⁸

Why I agree with the direction, with a grain of salt

• The direction is right. Every generation offloads skills to tools. Nobody mourns the slide rule or the log table, and nobody memorises phone numbers any more.
• The messenger is conflicted. NVIDIA sells the hardware that powers AI adoption. Discount the enthusiasm accordingly.
• The example is weak. Forgetting a ZIP code is not the same as never learning number sense. You cannot become a systems thinker without a mental model of the system, and you cannot spot a wrong AI answer without a feel for what a right one looks like.
• "We will discover new skills" is half a plan. He is right that new skills matter more than the old ones. The main one is already clear: judgment. Naming it and teaching it now is the job of education systems. Discovering it a generation late is the failure mode.

4What the evidence says: design decides the outcome

The research does not say "AI harms learning". It says unstructured AI harms learning, and structured AI can accelerate it. Each camp quotes half of it.

Three terms, explained once

• Randomised controlled trial (RCT). Students are assigned by lottery to use a tool or not. Any difference in results can then be credited to the tool.
• Effect size, in standard deviations (SD). A common ruler for learning gains across studies. 0.2 SD is roughly a student moving from the middle of the class to the top 42%; 0.4 SD, to the top 34%.
• Guardrailed tutor. An AI set up to teach, not to answer: it gives hints one step at a time and never hands over the full solution.

Same model, opposite outcome

The cleanest test is a trial in Turkey with nearly 1,000 high school maths students.⁹ Three groups: notes only, plain GPT-4 in a chat window, and the same GPT-4 set up by teachers to give hints, not answers.

A trial by trial view

Read the big meta analyses with care

A meta analysis pools many studies into one number. The pooled numbers for generative AI look spectacular, and that is the warning sign.

The most cited 'ChatGPT boosts learning' number no longer exists. Field trials that measure learning without the tool report far smaller, and far more useful, numbers.

The OECD's 2026 Digital Education Outlook names the problem directly: meta analyses claiming ChatGPT boosts "academic performance" often measure immediate task achievement, "not durable learning, and sometimes not even learning".¹⁰

Three conclusions I draw

• The helicopter without a pilot licence is real. When children use a general chatbot to produce answers, measured learning drops. Duminil-Copin is right about that.
• The tool is not the variable. The pedagogy is. Guardrailed tutors, sequencing (think first, then AI) and teacher framing reverse the effect. Huang's "it doesn't matter" ignores this. So does the mathematician's "don't use it".
• The biggest gains came with an adult in the loop. Nigeria, Harvard, Stanford and the Google and Eedi trial all had educators designing or supervising the experience. That is the thread I pull on below.

5We have seen this film before

Each time a tool made the old exercise trivial, schools first resisted, then slowly raised the bar. The lag is the damage.

Calculators: endorsed in 1980, tested in 1994

• Ownership came first. By 1975, a Science News estimate put one calculator in circulation for every nine Americans.¹¹ By 1980, 75% of US 9 year olds and 85% of 17 year olds owned or had access to one.¹²
• The profession said yes early. In 1980, the NCTM's Agenda for Action recommended calculators at all grade levels.¹¹ Nearly 100 studies had found no measurable harm.¹³
• Classrooms did not follow. In 1986, some teachers picketed the NCTM meeting in Washington against its pro calculator position.¹⁴ The famous photo overstates it: about 15 of 6,500 attendees took part.¹⁵ The real resistance was quieter, and it lived in daily practice.
• Ten years later, it had still not landed. The 1990 NAEP found 47% of 4th graders and 22% of 8th graders were never asked to use a calculator in maths class. Only 2% of 4th graders could use one on a test. The US statistics agency's own verdict: "NCTM's recommendations have not been implemented."¹⁶
• The exam moved last. The SAT allowed calculators from March 1994, fourteen years after the endorsement. 87% of test takers brought one that year.¹¹

The research explains why the lag matters. A 2003 meta analysis of 54 studies found calculators improved operational and problem solving skills when they were "an integral part of testing and instruction", and did not hinder basic skills.¹⁷ The gain arrived only once the assessment changed too.

The internet: wired rooms, unready teachers

• The wiring was fast. US public school classrooms with internet access went from 3% in 1994 to 93% in 2003.¹⁸
• The teaching was not. In 1999, when 64% of classrooms were online, only about one teacher in three felt well prepared to use computers and the internet for instruction.¹⁹

Wikipedia: the ban reflex

When I was 15, dissertations still asked us to compile facts. Then everything became findable. For a while the bar stayed where it was, which meant a copy paste from Wikipedia could pass.

• The institutional reflex was to ban. In January 2007, six years after Wikipedia launched, Middlebury College's history department barred students from citing it, after several students repeated the same wrong fact from it on an exam.²⁰
• Even the chair doubted it. The department chair said a full ban would be impractical.
• Good teachers moved the goal instead. The value was no longer finding information. It was combining sources, judging reliability, and forming and defending an opinion.

AI: the fastest step change yet

The lag is about to get more expensive, because the step is arriving faster than any before it.

The pattern

Children adopt first. Teachers adjust last. In between, the bar sits embarrassingly low relative to the tools available.

• With AI, that gap is wider than ever, because the tool does not just compute or retrieve. It writes, reasons and explains.
• Keeping yesterday's exercises and banning the tool does not preserve rigour. It guarantees that homework measures nothing, as the Chinese data already shows: homework up 18%, exams down 20%.

The context: results are already sliding

PISA tests 15 year olds in maths every few years.

• 2018 to 2022: the largest drop in its history. 15 points across the OECD, in the round hit by COVID.²¹
• 2022 to 2025: down again. 9 more points across 35 comparable OECD countries. France fell 16 points, to 458. Apart from Türkiye, no OECD country improved.²¹
• Over ten years: maths scores fell 22 points since 2015, which the OECD equates to just over a year of learning.²¹

2018 values derived as the 2022 score minus the OECD reported change. The OECD average covers a different set of countries in each round; the 2022 to 2025 change of minus 9 is measured on 35 comparable countries. Belgium 2025 read from a table built on OECD data (secondary). Flanders scored 489 in 2025; the Wallonia Brussels Federation 467 (secondary).

This is context, not proof. The 2022 drop predates mass AI use, and COVID explains part of it. The 2025 test was taken when most 15 year olds already used chatbots for schoolwork, but PISA cannot say whether AI helped or hurt. What it does show: the system entering this step change is already losing ground, which leaves no room for another decade of lag.

6The real bottleneck: the adult in the cockpit

Children will use AI whether we allow it or not. The only variable we control is whether their teachers can fly better than they can.

My experience

I have spent time with teachers and school directors explaining AI. Many are deeply sceptical of the technology, highly risk averse, and in some cases actively blocking change.

I understand why. They are overloaded, rarely trained, and have watched many edtech promises fail. But we cannot afford a repeat of the calculator decade.

Students moved first

When ChatGPT arrived, schools had no rulebook. In May 2023, fewer than 10% of about 450 schools and universities UNESCO surveyed had formal guidance on generative AI.⁵⁶ Students did not wait for one.

Teachers are catching up unevenly

The OECD's TALIS 2024 survey asked about 280,000 teachers and school leaders in 55 systems whether they used AI in their work and whether they were trained on it in the past year.²²

Lower secondary teachers. Belgium: French Community 23% use, 16% trained; Flemish Community 40% use, 35% trained. Source: OECD TALIS 2024 report, tables 1.59, 4.22 and 4.27, and country notes.²²

What stands out

• AI is the most requested training topic. 29% of teachers say they need more of it, ahead of every other subject surveyed.²² Resistance is partly a skills gap wearing the costume of a principle.
• The barrier is skill, not rules. Three non users in four say they lack the knowledge or skills to teach with AI. Only about one teacher in ten says their school does not allow it.²²
• The French speaking world sits at the bottom. France trained 9% of its teachers on AI, the lowest of all 55 systems. Singapore trained 76%. The French Community of Belgium trained 16%, against 35% in Flanders.²² That is the context in which a Fields medallist's "I don't use it" lands as a norm rather than a personal choice.
• Same country, three speeds. In France, 85% of 15 year olds use AI chatbots for schoolwork, 14% of teachers use AI, and 9% were trained on it.^22,45
• "Trained" usually means one session. Half of US teachers had at least one AI session by fall 2025, but only 5% to 9% report ongoing training.⁴¹ 68% had no school or district training at all in 2024 to 25, and 52% taught themselves.⁴²
• Nobody teaches students how. In the US, over 80% of students say their teachers never explicitly taught them how to use AI for schoolwork.²³
• The positive evidence always had a teacher in the loop. Edo State, Harvard and Stanford worked because educators designed the scaffolding. Remove the pilot and you get the Chinese homework data.

7My take: teach them to fly, and move the summit

A caveat first. I have spent years reading on education and worked on school improvement projects during my time at McKinsey. What follows is my personal view, not a conclusion of that work.

I would take both Duminil-Copin and Huang with a grain of salt. One is a researcher at the absolute frontier of his field, with a way of working that suits him and very few others. The other sells the shovels of this gold rush.

The metaphor argues against its author

• Flying a helicopter is harder than walking. Ask someone who has never flown to land on Everest: thin air, unpredictable winds, a landing zone the size of a table. It is one of the hardest things a pilot can do. Getting into the helicopter is not the skill. Flying it is.
• The child in his story is a passenger, not a pilot. That is the real failure. Not that a helicopter exists, but that we let children ride in it without teaching them to fly, navigate, read the weather or know when to land.
• If the climb is now easy, the mountain is too small. The right response to a helicopter is not to ban it. It is to pick a higher mountain, or a range nobody has mapped.

Amaury Hayat's exoskeleton makes the same point from inside the profession: an assisted ascent still teaches you the route, if you are paying attention.

Where I agree with Huang, and where I go further

Huang is right on the direction. Some skills stop mattering, and new ones take their place. The most important new skill already has a name: judgment.

• Execution is getting cheap. Computing, retrieving, drafting and formatting are what AI does well. They used to fill most of a school day.
• Judgment is getting valuable. Framing the problem, choosing the method, checking the answer, deciding what good looks like: that is what is left for people.
• School still trains execution. Too much time goes to learning by heart and executing procedures mindlessly. That was never the point of education, and now a machine does it in seconds.
• Some execution stays. You cannot judge whether an AI's answer is plausible without number sense, or direct it to write a strong argument if you have never built one. But a skill that used to be manual does not have to stay manual.

Keep the execution that judgment rests on, and hand the rest to the tool. What changes is how much time we spend on execution versus judgment, and how high we set the bar once the tool is in hand. Train teachers and students in judgment, not in mindless execution.

The skills the job market already names

Employers are not waiting. In the World Economic Forum's 2025 survey, employers expect 39% of workers' core skills to change by 2030, and seven in ten companies list analytical thinking as essential.²⁴ Those are pilot skills, not passenger skills.

8Hints are the floor. The summit is the point.

Read section 4 quickly and you could draw the wrong lesson: lock the model down so it only gives hints. That protects learning. It does not raise it.

What hints do, and what they do not

• They stop the loss. In the Turkey trial, the hint only tutor left exam scores where they were without AI.⁹ No loss. Also no gain.
• They free time. At Harvard, students learned more with the tutor in a median of 49 minutes, against about 60 in class.²⁵
• The authors say what the time is for: "advanced problem solving, project based learning, and group work".²⁵ That is a proposal, not something the trial measured.

Hints are necessary for practice. They are not the strategy. The strategy is what students do with the time and the power the tool gives them.

What happens when people use AI to climb

Outside school, controlled trials show the same pattern again and again: with AI, people reach work above their level.

• Beginners jump levels. At BCG, below average consultants gained 43% with AI, above average ones 17%. Support agents with two months of experience performed like colleagues with six months.^57,58
• Children too. Novices aged 10 to 17 learning Python with an AI code tool scored 1.8 times higher, with no loss on tasks done by hand afterwards.⁶⁰
• The work moves up. In a writing trial, AI shifted effort away from rough drafting and toward idea generation and editing.⁵⁹
• Judgment is the limit. On a task outside what AI does well, consultants using it were 19 points less likely to get the right answer.⁵⁷ The tool raises the ceiling only for people who can tell when it is wrong.

No study yet measures students taking on harder projects because of AI. The argument rests on these trials, not on a school result. It is a design choice, and teachers have to make it.

Climb or coast: five ways to use the same tool

About six in ten 15 year olds in PISA 2025 say a lesson already asked them to assess the quality of AI generated information.⁴⁵

Where the school day goes

School spends most of its time on execution. Four sources, measured differently, point the same way.

• Most lessons ask little. In five US school systems, only 74 of about 900 lessons observed focused on understanding and asked students to do the thinking. Students met the demands of their assignments 71% of the time, but showed grade level mastery only 17% of the time.⁶²
• Memorising does not reach the top. In PISA 2012, the students who relied most on memorisation were four times less likely to solve the hardest maths problem than those who relied on it least.⁶³
• Tests reward recall. In 9,713 exam and quiz items from 50 university biology teachers, 93% tested knowledge and comprehension, the two lowest levels of Bloom's scale.⁶⁴
• Teachers report the same balance.

The OECD's 2026 outlook draws the conclusion for the AI era: because AI output needs human scrutiny, critical thinking and higher order skills become "even more important than before", and judgment sits on its list of crucial skills.¹⁰ Employers agree: seven in ten rank analytical thinking as essential.²⁴

The one test

Ask one question after any use of AI in class: did the student end up doing harder thinking than they could have done alone? If yes, they are climbing. If the AI did the thinking, they are coasting, however good the result looks.

9What should change: five moves for schools and ministries

The five moves

1. Train teachers first, and deeply. Fund real AI professional development before rolling tools out to students. Singapore's 76% training rate against France's 9% is a policy choice, not destiny.
2. Move assessment to where AI cannot fake it. In class reasoning, oral defence, live problem solving and process portfolios. The Chinese study shows take home homework is now a weak signal. Stop grading it as if it were a strong one.
3. Protect the floor with tutors, then aim above it. For practice, buy AI that gives hints one step at a time, as in the Turkey and Harvard trials, not a generic chatbot. For projects, give students AI as a sparring partner, a critic and a power tool. Hints stop the loss; the other modes create the gain.
4. Teach "think first, then AI". Sequence tasks so students attempt first, then use AI to challenge, extend and check. The MIT essay study and a 2026 randomised lab experiment both point the same way: using AI to explain beats using it to produce.
5. Raise the summit. Keep the foundations that enable judgment (number sense, argument structure, source criticism). Then spend the freed time on harder, messier problems than we used to dare assign. A 15 year old with AI should produce work that a 15 year old without it could not.

Who is already moving

Scale figures are as reported by each government or institution. Korea and New York are listed as warnings, not models.

Two warnings

• Tools before teachers fails. South Korea spent at least 533 billion won on AI textbooks, reached about 30% of schools, and lost the legal status of the textbooks within five months.⁴⁷
• The ban pendulum wastes years. New York City banned ChatGPT in January 2023 and reversed the ban four months later; its chancellor called the first reaction "knee jerk fear".⁵¹ In September 2026, the city imposed a new one year moratorium on student facing generative AI for nearly 600,000 children up to grade 8. A pause can be a legitimate choice for young children. Oscillating without a teacher plan is not.

Why policing does not work

• Detectors misfire. Seven AI detectors flagged 61% of essays by non native English writers as AI written, on average.⁵²
• The vendor gave up. OpenAI withdrew its own detector in 2023: it caught 26% of AI text and wrongly flagged 9% of human text.⁵²
• Almost nobody is caught. 88% of UK undergraduates used AI for assessed work in 2025.³⁹ Proven AI cheating cases ran at 5.1 per 1,000 students in 2023 to 24.⁵³

Detection is not a policy. Redesigned assessment is.

For the mathematicians

None of this contradicts Duminil-Copin's deepest point. Building concepts yourself, the moment of "eurêka", is irreplaceable. The question is whether we protect it by banning the helicopter, or by teaching children to fly to places where there is still something left to discover.

10Conclusion

The mathematician fears children will become passengers. The CEO shrugs that passengers still arrive. Both accept the same frame: a fixed mountain, and a choice between walking and flying.

I reject the frame.

• AI is one of the largest step changes in human history. Adoption three years in is already higher than the PC or the internet at the same age.
• Education will lag it, as it lagged the calculator and the internet.
• That lag will show up as falling performance, because the bar will sit absurdly low relative to the tools in every pocket.

The fix is not to ground the helicopter. It is to train the pilots, starting with the teachers, and then to point everyone at a much higher summit.

11The fine print

• Studies are not directly comparable. Trials differ in age, subject, duration and test. The effect size chart is indicative, not a league table.
• The Chinese study is not a randomised trial. It compares students who adopted AI at different times (a difference in differences design). It is strong evidence of a pattern, not proof of a cause in every classroom.
• Some trials are preprints, not yet peer reviewed: the MIT essay study, the 2026 lab experiment, and the Google and Eedi trial, which was run by the vendor's own team.
• PISA trends are context, not causation. The 2018 to 2022 drop predates mass AI use, and COVID explains part of it.
• Survey questions changed. Pew asked about ChatGPT in 2023 and 2024, and about any AI chatbot in 2025. The jump from 26% to 54% partly reflects the wider question.
• TALIS figures are for lower secondary teachers (students aged about 11 to 15) and exclude England and the UK in the 2024 tables.
• The retracted meta analysis (Wang and Fan, 2025) is cited only as a warning. Its numbers are not used as evidence.
• The interviews with Duminil-Copin and Huang are summarised from the France Inter audio and the New York Times episode and its coverage, not quoted word for word.

12Glossary

Cognitive offloading

Handing a mental task to a tool, such as a calculator or a chatbot, so the brain does less of it.

Difference in differences

A method that compares how two groups changed over the same period, to isolate the effect of something only one group started doing.

Effect size, standard deviation (SD)

A common ruler for learning gains. 0.2 SD moves a middle student to about the top 42% of the class; 0.4 SD to about the top 34%.

Guardrailed tutor

An AI set up to teach: it gives hints one step at a time and does not hand over full answers.

Meta analysis

A study that pools the results of many studies into one estimate.

NAEP

The National Assessment of Educational Progress, the main US national test of what students know.

PISA

The OECD's Programme for International Student Assessment, which tests 15 year olds in maths, reading and science.

Preprint

A research paper shared publicly before peer review.

Publication bias

The tendency for positive results to get published more than null ones, which inflates pooled estimates.

Randomised controlled trial (RCT)

An experiment where participants are assigned by lottery, so differences in outcome can be credited to the intervention.

TALIS

The OECD's Teaching and Learning International Survey of teachers and school leaders.