A student sits down with a word problem, opens a chatbot, and has an answer in four seconds. That is not a hypothetical. It is the new baseline behavior in classrooms across the country, and the data on what it is doing to actual math skill is no longer speculative either. The numbers are in, and they are not flattering.
The Data Behind the Headline
A June 2026 working paper, "Faster Completion, Less Learning: Generative AI Reduced Study Time on Math Problems and the Knowledge They Build," tracked student performance on supervised, no-AI-allowed placement tests before and after ChatGPT became widely available. Students had historically answered roughly 80 percent of word problems correctly. After AI use became routine in their study habits, that number fell to around 60 percent — about a 25 percent drop in the odds of getting a word problem right, measured on tests where the AI wasn't even in the room. Tellingly, performance on graphing problems barely moved. The decline concentrated exactly where multi-step reasoning lives, not where pattern recognition does.
Norway's experience is the same story at national scale. Between 2018 and 2022, Norwegian 15-year-olds lost 33 points in PISA mathematics scores — the second-largest drop of any Nordic country, behind only Iceland's 36-point fall. That decline predates the ChatGPT era, but it was the backdrop against which Norway's government moved in 2026 to bar children under 13 from using ChatGPT and similar tools in schools starting this August. A country doesn't ban a technology for children over a hunch. It bans it after watching the score sheet for years and running out of other explanations.
Closer to home, RAND's American Youth Panel found that AI homework use kept climbing through 2025 across middle school, high school, and college — while 60 percent of the students using it admitted, in the same surveys, that they were worried about what it was doing to their own critical thinking. Students can feel the erosion happening in real time. They are using the tool anyway, because the assignment is due tonight and the tool is faster than the struggle.
Why Speed Is the Problem, Not the Feature
The mechanism researchers keep landing on is cognitive offloading — handing a mental task to an external system instead of doing it yourself. It's not a new concept; calculators and spellcheck offload too. The difference with a chatbot is that it offloads the entire reasoning chain, not just the arithmetic at the end of it. A calculator still requires you to know which operation to perform. A chatbot will figure that out for you as well, which means the part of the problem that actually builds procedural fluency — deciding how to attack an unfamiliar problem — never gets exercised.
Procedural fluency isn't a nice-to-have skill for math class. It's the scaffolding that later abstraction sits on. Skip enough of the struggle and you get exactly what the working paper found: intact performance on problems that reward recognizing a familiar shape, and a collapse on problems that require constructing a path to the answer from scratch. That gap is the whole story, and it should worry anyone who thinks "AI literacy" is a substitute for math literacy rather than a layer on top of it.
The decline concentrated exactly where multi-step reasoning lives. Graphing problems, which reward recognition over construction, barely moved.
What Actually Works: Friction by Design
The counterexample is worth taking seriously, because it proves the failure mode is a design choice, not an inevitability. Khan Academy's Khanmigo doesn't hand over answers — it asks questions back, in the Socratic mode, refusing to close the loop until the student has done the reasoning. In 2025 pilot data, students using it improved by an average of 1.4 grade levels in math, and a separate SRI International efficacy study clocked 23 percent faster mastery of algebra concepts against Khan Academy's own traditional video instruction. Two randomized controlled trials surfaced by Stanford's National Student Support Accelerator in February 2026 found AI-embedded math tutoring produced a 66 percent success rate on follow-on topics, compared with 61 percent for human tutoring alone.
Put those two bodies of research side by side and the design principle is obvious:
- Tools that answer immediately produce faster homework completion and measurably worse retention.
- Tools that require shown work, delay the answer, or interrogate the student's reasoning produce faster mastery and better retention.
- The variable that predicts the outcome isn't "AI or no AI" — it's whether the tool preserves the friction that forces a student to reason.
That's a procurement decision schools can actually make, not a philosophical debate about whether AI belongs in a classroom. The tools already exist on both sides of that line. Buying the wrong one, or letting students default to whichever chatbot is already open in another tab, is a policy failure, not a technology one.
Norway chose the blunt instrument: a ban for the youngest students while it figures out the rest. That will buy time but it won't teach anyone to reason through a word problem. The sharper instrument is the one Khan Academy and the Stanford-reviewed pilots are already running — tools built to withhold the answer until the student has earned it. The research this year says plainly which approach produces a student who can still do math without the app open. It isn't the one that's faster.
