Picture a student working through a set of word problems with a chatbot open in the next tab. She isn’t copying answers outright, just asking for a hint, a nudge, a way to see the setup. The problems get done fast. Correctly, mostly. She closes the laptop feeling like she just had a good study session.
Then a researcher takes the chatbot away and asks her to do it again, alone.
That’s roughly the shape of a study run on ALEKS, a math platform used in schools across the country. Researchers looked at millions of student interactions, from before ChatGPT existed and after, and the picture that came back is almost exactly what you’d predict once someone lays it out for you, and almost nobody predicts going in.
Students got faster. They also got worse — 25% worse odds of solving a word problem correctly, but only once the AI was gone and they were tested solo. Speed went up. Skill, tested honestly, went down.
Here’s the detail that makes it hard to argue with. Graphing problems, the kind that are annoying to copy into a chatbot, didn’t decline at all. Same students, same class, same semester. The only thing that moved was whether a given problem happened to be easy to hand off.
Researchers have started calling this cognitive surrender. The tool does the thinking, the answer shows up fast, and the whole exchange feels exactly like learning while you’re inside it. There’s no warning bell. You did understand something in that moment, just not the something the exam is going to ask for. You understood how to ask a good question. You never built the part where you answer one cold, and nothing about the experience tells you the two aren’t the same thing until a test does it for you.
I’ve sat across from enough students holding a transcript full of A’s that stopped meaning much by December to recognize this up close. It almost never looks like laziness. It looks like relief.
And none of this actually started with AI. The international PISA testing had already caught reading and science scores sliding for over a decade before a chatbot existed to blame for anything. Then COVID landed, and math scores alone lost something like three-quarters of a school year in one stretch. Whatever AI is doing now, it walked into a slide that was already underway.
What’s worth noticing is what actually predicted which schools held up through all of that. Not the ones that banned phones outright. The ones where teachers stayed engaged and parents stayed involved. Even the raw technology numbers refused to sort into “screens bad.” Kids distracted by a classmate’s device in math class scored 15 points lower. Kids using tech for their own schoolwork, in modest daily doses, scored higher. Same object in the room. Opposite result, depending on who was actually in charge of it.
The pattern doesn’t stop at graduation. A Stanford team studying the job market found that AI is remarkably good at replacing what researchers call codified knowledge, the textbook material, the formal stuff a classroom is built to hand you. The trouble is that entry-level workers arrive carrying mostly codified knowledge and not much else, because that is more or less what entry-level means. The other kind, tacit knowledge, the judgment you only build by doing the actual thinking yourself, in real situations, over and over, takes time nobody just starting out has had yet. Early-career employment in the roles AI automates most is already down 16% relative to everywhere else.
Put next to each other, none of this is really a story about phones, or chatbots, or any one app. It’s a story about a shortcut that feels identical to the real thing for exactly as long as nobody checks. The tool never took anything away from these students. It just made it easy to hand something over without noticing that’s what happened.
So here’s the question worth sitting with the next time an answer arrives before the effort does, at your laptop tonight or ten years from now in a job nobody’s invented yet: if the feeling of understanding and the fact of it are this easy to mix up, how exactly would you know, in the moment, which one you’re standing on?
Sources
Barshay, J. (2026, July 6). Faster solutions, lower test scores: How AI is eroding math skills. The Hechinger Report.
Brynjolfsson, E., Chandar, B., & Chen, R. (2025, November 13). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab.
OECD. (2023). PISA 2022 results (Volume I): The state of learning and equity in education. OECD Publishing.
Tully, S. (2024). The U.S. spent $30 billion to ditch textbooks for laptops and tablets: The result is the first generation less cognitively capable than their parents. Fortune.