Calculators Made Better Mathematicians by Refusing to Do It All. Can AI?

Every generation gets a tool that promises to make thinking easier, and every generation panics that it will make thinking disappear. Calculators went through this in the 1970s and 80s. AI chatbots are going through it now. But there’s a real difference between the two panics, and it’s worth taking seriously.

The Calculator Precedent

The fear was straightforward: give kids a machine that does arithmetic, and they’ll never learn arithmetic. Decades of research say otherwise. A synthesis of nearly 200 classroom studies by the National Council of Teachers of Mathematics (NCTM) found that calculator use does not damage skill development or procedural proficiency; if anything, it improves conceptual understanding and student attitudes toward math. A separate meta-analysis of 79 precollege studies (Hembree & Dessart, 1986) found calculators improved basic problem-solving skills in nearly every grade level studied.

But the research comes with a condition, and it’s the condition that matters here. A Vanderbilt University study (Rittle-Johnson & Kmicikewycz, 2008) found that calculators only helped when students already had foundational number sense in place first. As the study’s co-author put it, kids need to learn to calculate on their own before the calculator becomes “a fine thing to do.” A UK review commissioned by the Education Endowment Foundation (EEF, 2018) reached the same conclusion: calculators boost fluency only when they’re layered on top of mental math skills students already have, not used as a replacement for building them. One study out of a private school in Bulacan (Beros et al., 2024) even found a measurable negative correlation between calculator dependence and students’ confidence in their own fundamental skills. The more they leaned on the tool, the less sure they felt about the basics underneath it.

In other words, calculators didn’t fail to erode skills because they were harmless. They failed to erode skills because of a structural limitation: a calculator can’t do the whole problem. Punch numbers into a calculator and ask it to “solve this integral,” and it stares back blankly. The student still has to know that this is an integration problem, break it into steps, decide where a derivative is needed, decide where a definite value needs computing, and only then hand the arithmetic itself to the machine. The tool assists execution. The human still owns the reasoning.

What Changes With AI

That structural limitation is exactly what’s missing from an AI tool. Ask ChatGPT to solve the whole integral, steps and reasoning included, and it will. Ask it to write the essay, not just check your grammar, and it will do that too. There’s no point at which the tool hits a wall and hands control back to you. The barrier that accidentally protected students from over-reliance on calculators simply isn’t built into general-purpose AI assistants.

The data on outcomes reflects that difference. Researchers at MIT’s Media Lab ran a controlled study titled “Your Brain on ChatGPT” (Kos’myna et al., 2025), asking 54 participants to write SAT-style essays using either ChatGPT, a search engine, or no tools at all, while measuring brain activity via EEG. The ChatGPT group showed the weakest neural connectivity of the three, had the hardest time recalling what they’d just written, and reported the lowest sense of ownership over their own essays. The researchers coined a term for the pattern, “cognitive debt”: mental effort deferred now, paid back later as weaker recall, weaker independent reasoning, and, in their words, greater vulnerability to manipulation.

This isn’t a fringe pattern. A RAND (2025) report tracking the American Youth Panel found that the share of students using AI for homework climbed from 48% in May to 62% by December. Over that same stretch, the share of students who believed AI use was harming their own critical thinking rose from 54% to 67%. Notably, even the students still using the tools admitted the cost: 60% of AI-using students said it was hurting their critical thinking, a view shared by 78% of students who’d opted out of using AI altogether. Separately, College Board (2025) research put AI homework use among students at 84% as of October. Whatever the exact number, the direction is the same across every survey: adoption is climbing faster than confidence in the outcome.

The Missing Guardrail

Here’s the mechanism worth sitting with. Nobody picks up a calculator and starts trying to solve an integral cold. You don’t reach for a calculator unless you’re already mid-problem and already know which specific operation you need help with. The tool only ever gets involved after the human has already done the conceptual work of framing the problem. That sequencing, human frames it and tool executes a piece of it, is what the research above suggests actually builds skill rather than eroding it.

AI assistants don’t require that sequencing. You can open a blank chat window with zero understanding of a topic and walk out the other side with a finished product, having supplied none of the framing yourself. That’s a genuinely new failure mode, and it’s the one the MIT researchers were flagging when they warned about consequences for learners still developing foundational skills.

So the idea worth taking seriously: what if AI tools deliberately reintroduced that barrier? Not dumbing the tool down, but gating full solutions the way calculators are structurally gated already, requiring a demonstration of foundational understanding (a rough attempt, a stated approach, an explanation of what’s being asked) before the assistant will carry a task the rest of the way, rather than handing over the whole answer to a prompt with no framing behind it at all.

It’s not a new idea in disguise, either. It’s closer to how good calculators already worked without anyone designing it that way. Nobody sat down in 1975 and decided calculators would only help you after you understood the problem. It just happened to be built into what the device was capable of. With AI, that same guardrail would have to be a choice, because the technology itself no longer forces it.

The Trade-Off Nobody Wants to Say Out Loud

Gating AI this way has an obvious cost: it makes the tool less convenient, and convenience is most of the reason people adopted it in the first place. It also raises a fair question the MIT researchers themselves flagged: some AI use isn’t replacing thinking at all, it’s enabling work that would’ve been impossible without it, the same way a calculator now lets engineers run computations no one would do by hand. Restricting that kind of use for the sake of protecting foundational learning would be throwing out real capability to guard against a risk that mostly applies to students still building basic fluency, not to someone already competent using the tool to go faster.

The honest version of the argument, then, isn’t “AI should always make you prove yourself first.” It’s narrower: for learners who haven’t built the fundamentals yet, an assistant that can complete the whole task with no framing required is a different, and so far worse-performing, tool than the one previous generations grew up on. Whether the fix is gating, better classroom design, or just teaching people to use AI the way the EEF review says students learned to use calculators (thoughtfully, and only after the basics are in place) is still an open question. But pretending the two tools carry the same risk profile isn’t supported by what the data actually shows.

References

Beros, J., Bono, K. A., DeChavez, M. C., Labrador, L., De Jesus, R., Datiles, M. R., & Tus, J. (2024). Calculator usage and its relationship on student’s perception of their fundamental mathematical skills. Psychology and Education: A Multidisciplinary Journal, 18(3).

College Board. (2025). Student AI usage research, October 2025.

Education Endowment Foundation & Nuffield Foundation. (2018). Improving mathematics in key stages two and three [evidence review with UCL Institute of Education and University of Nottingham].

Hembree, R., & Dessart, D. J. (1986). Effects of hand-held calculators in precollege mathematics education: A meta-analysis. Journal for Research in Mathematics Education, 17(2).

Kos’myna, N., et al. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task [preprint]. MIT Media Lab.

National Council of Teachers of Mathematics. Calculator use [Research Brief].

RAND Corporation. (2025). American Youth Panel survey, May-December 2025.

Rittle-Johnson, B., & Kmicikewycz, A. O. (2008). When generating answers benefits arithmetic skill: The importance of prior knowledge. Journal of Experimental Child Psychology, 101(2).