I've used the car comparison on this blog before, but there's a part of it I skipped.
Cars let us carry more and travel further than we ever could on foot. They also, almost certainly, made us less fit. In the olden days you walked to get anywhere, and you carried what you needed or pulled a wagon, so exercise came built in. Now we barely walk, and we've had to invent deliberate versions of it. Ten thousand steps is something you do on purpose. So is the gym.
I think AI is doing the same thing to thinking. The slow, tedious mental work that used to build skills as a side effect is the walking. AI is the car.
Does using AI make you worse at learning?
The honest answer is that nobody knows yet. Chatbots haven't been around long enough for anyone to run proper long-term studies, especially on young children. What we have is early, partial evidence. It's worth looking at carefully, without panicking.
The clearest I've found is a randomised trial at a high school in Turkey. About 1,000 students in grades 9 to 11 took part in maths lessons. One group studied from textbooks and notes. One group got a standard ChatGPT-style assistant. A third group got a version set up by teachers to give hints rather than answers. Then everyone sat an exam with no help at all.
- The standard AI group scored 48% higher during practice, then 17% lower on the exam than the textbook group.
- The hint-giving group scored 127% higher during practice, and on the exam did about the same as the textbook group.
Most students in the standard group had simply asked for the answer, and that tool was right only about half the time. Both AI groups also thought they'd learned more than they had. [1]
It's one school, one subject and four lessons, so it's a warning rather than a verdict. But it's exactly what the car comparison would predict: it feels great while you're being driven, and you find out how fit you are when someone asks you to walk.
Closer to home, Australia's results in the OECD's PISA 2025 tests came out in September. Maths and reading scores for 15-year-olds fell by almost 2%. Students who used chatbots to summarise texts had lower average science scores, and half of Australian students now use AI tools for learning at least weekly. Education Minister Jason Clare said "the children that are online more are getting lower results." [2] That's a link, not proof that screens or AI caused the fall. Plenty else changed over the same period.
Why does AI speed up learning for some people and not others?
Because it depends on what you bring with you. I'm an example of both ends.
I've done enough blogging and SEO to know what I want from a page. When AI drafts something, I can tell when it's off, steer it, and see what impact each change has. I'm not a professional, but I have enough of the basics to guide it, and AI makes me faster.
Coding and Git are the opposite. Claude Code does that work, and it's still taking me forever to understand it myself. I can build things I'd never manage alone, but I'm building a house without fully understanding how it's put together. If something breaks, I often can't tell why. That's the black box: I'm not directing the work, I'm approving it.
Same tool, two very different outcomes. The difference is the basics. If you have them, AI speeds you up. If you don't, it hides the gap until the day it matters.
There's a third way. You can ask AI to teach you the basics. That's roughly what the hint-giving version in the trial did, and at worst it didn't hurt exam scores. Used as a tutor, it can be a shortcut to the basics. Used as an answer machine, it skips them.
What can you do about it?
None of this means giving up the car. I'm not going to, and neither are you. It means choosing your steps on purpose.
- Have a go first, even badly. Write the rough version, work out the answer, or sketch the plan, and then bring in AI. The struggle is the walking.
- Ask it to teach, not tell. "Don't give me the answer, give me a hint" or "quiz me on this" is a very different request from "do this for me."
- Explain the result back in your own words. If you can't, you don't own it yet.
- Pick the skills you don't want to lose. Choose a few things you'll keep doing by hand, even though AI could do them faster. That's your 10,000 steps.
Why does this matter most for kids?
Adults who use AI well usually already have the basics. Kids are still building them. That's why the question of when to start matters so much for them, and I've covered it in a companion post, When Is It Appropriate for Kids to Start Using AI?
Where this leaves us
The car isn't the enemy. It gave us things walking never could. But we learned, eventually, that staying fit meant walking on purpose. Learning with AI will be the same. The tool is fine. The question is whether you still get your steps in.
Sources
- PsyPost, Unrestricted generative AI harms high school math learning by acting as a crutch, summarising "Generative AI without guardrails can harm learning: Evidence from high school mathematics" (PNAS). Randomised trial, about 1,000 students, one school in Turkey.
- Information Age (ACS), Govt blames technology as student scores hit record low, reporting the OECD's PISA 2025 results and Education Minister Jason Clare's comments.
alltools.solutions builds free and paid tools, sometimes in the same space as what's discussed here. Nobody mentioned in this post paid for or reviewed it, and nobody pays us to mention them. We hold our own tools to the same rubric we'd apply to anyone else's.