One of the most important (if least discussed) findings in educational research concerns calculators.
Several meta-analyses have found that students who learn mathematics using calculators perform better on exams…so long as they retain access to their calculators. Once the calculators are removed, this advantage largely disappears and performance on procedural skills typically reverses (1, 2, 3).
This evidence suggests that, rather than boosting learning, calculators boost calculator-assisted performance. Students can become highly proficient at manipulating a machine to produce the ‘correct answer’ without developing the underlying knowledge and skills required to self-generate or genuinely understand that answer.
The majority of digital tools have been built for productivity: they make experts more efficient by reducing the time and effort to complete cognitive tasks they already understand.
Consider statistics.
After nearly two decades in academia, I have become quite skilled at performing the required tests, procedures, and corrections required to find meaningful patterns in data. Unfortunately, performing these tasks manually can take hours (sometimes days).
Luckily, statistical software can complete them in mere seconds.
But this software only works for me because I already possess the deep statistical knowledge required to drive it. I can identify faulty inputs, recognize implausible outputs, and detect when the software has substituted one procedure for another. Although the machine performs the calculations, I remain responsible for the thinking, vetting, and adaptation.
When experts use a tool to make performance easier, we call this cognitive offloading: and it works because established expertise allows the user to maintain control over the tool. Cory Doctorow describes this relationship as a centaur: part human, part machine, but with the human head firmly in charge.
Unfortunately, novices (by definition) do not yet possess the knowledge required to meaningfully drive cognitive productivity tools. They lack the deep understanding required to isolate aberrant inputs, vet erroneous outputs, or creatively devise alternate yet accurate approaches. Oftentimes, the best they can do is copy words, press the prescribed buttons, and paste the resulting output.
This is not cognitive offloading – it’s now called cognitive outsourcing: when a tool is used to perform the very process a person is meant to be learning.
Whereas offloading is a way to demonstrate expertise, outsourcing typically prevents expertise from developing. Instead of becoming centaurs, novices risk become reverse-centaurs: the machine supplies the direction while the human performs the minor tasks required to keep it moving.
Generative AI can formulate arguments, summarize readings, solve problems, supply ideas, and generate explanations. In other words, it can perform many of the cognitive operations students attend school to develop.
An expert may use AI as an offloading tool because that expert can interrogate its claims, reject false reasoning, input missing context, and repeatedly re-direct the machine. But a student who has not yet developed this knowledge cannot do the same.
Crucially, the same person can be both an offloader and an outsourcer.
If I ask AI to ask summarize an argument about the brain (my area of expertise) I am offloading: I am a centaur driving and correcting the machine. But if I ask the same AI to summarize an argument about car repair (a topic I know nothing about), I am outsourcing: I am a reverse-centaur being driven by a machine whose output I cannot meaningfully evaluate.
The transition from outsourcing to offloading has little to do with training on the machine or developing some generic ‘AI literacy’. It depends upon deep knowledge of the field in which the tool is being used.
This means we do not need to learn more about calculators to transform outsourcing into offloading – we need to learn more about mathematics.
This is why the NYC policy is so intelligent.
First, the city has not permanently banned AI. It has established a moratorium while its effects are examined. That is basic prudence: if schools are going to place a powerful new technology in front of children, it should first demonstrate basic efficacy, safety, and legality.
Second, although critics argue that restricting AI usage will damage children’s employment prospects, does anyone seriously believe an eight-year-old will struggle to find work in two decades because they did not spend their primary school days using the AI systems of 2026?
If AI training were genuinely more important than a general education, we would remove students from schools and have them prompt chatbots all day. Of course, nobody proposes this because we intuitively understand that a tool is only as useful as the knowledge brought to it.
Carpenters aren’t better at construction than I am because they can swing a hammer any better than I can. They are better because they deeply understand carpentry – those material, structures, and principles that make swinging a hammer meaningful in the first place.
K-12 does not exist to ensure children know how to operate the tools of today: it exists to ensure that, when confronted with the unpredictable tools of tomorrow, they’ll have something meaningful to do with them.
Operating AI is easy (some might say trivial). Developing the knowledge, judgement, imagination, and expertise required to use it meaningfully is extraordinarily difficult.
That’s the work of education.
Too often, we mistake the tool for the skill. But outside of rare cases (such as the piano or archery), the tool is not the skill: it’s merely a means through which skill is expressed.
This is why the popular warning that “AI will take your job” has morphed into “Experts using AI will take your job.”
I predict that soon that will soon be simplified even further to “Experts will take your job”.
Tools work beneath expertise – they cannot meaningfully substitute for expertise we have yet to acquire. Students are not experts: they are developing the broad knowledge and specific skills from which expertise will eventually emerge.
This is why NYC’s AI moratorium is the first evidence-based response to classroom AI use I have seen.
Leave a Reply