A lot of ink is being spilled about the effects AI is having on learning and education.
The drawbacks are obvious. For many classroom assignments, AI is now good enough that it can simply do your homework for you. But having AI do your homework is self-defeating. You can’t learn without doing the work.
That said, the opportunities AI may provide are numerous. Alongside its drawbacks, AI also brings the possibility for cheap and patient tutoring, on-demand explanations, extra practice problems and assistance in research.
To paraphrase Dickens, when it comes to learning, it really is both the best of times and the worst of times.
This brings me to a recent essay [1] written by the always insightful Carl Hendrick. In it, he reviews recent findings on AI’s impact on learning, drawing an analogy between the introduction of autopilot and the degrading of pilot skill:
The thing I find really fascinating about this story [2] and why I think it has so much import for learning and AI is that his core diagnosis was that the autopilot was in fact, not the problem. The truth is that automated function in planes saves far more lives than it endangers and I think the same will be true of self-driving cars. (I have no interest in the hackneyed claim that the machines are bad and the humans good. The machines are, on the whole, magnificent I think). The thing I think is so interesting is that when faced with an emergency, the pilots reached instinctively for more automation not less, when the right thing to do was often to drop down a level and fly. So the danger is not automation itself but over-dependency on automation and this is quite obviously what happens when we give novice learners unfettered access to chatbots.
Hendrick doesn’t blame the students for this alarming trend; he blames educational institutions themselves. Schools have been vacillating between wanting to adopt a pro-AI stance, encouraging students to use AI in their work provided they properly “cite” the results, and embracing the language of academic honesty in a vain attempt to maintain the status quo.
Obviously, this is untenable. Education will have to change because of AI.
The answer isn’t to adopt some vague AI-forward solution which simply accepts AI output as a legitimate substitute for students doing the work themselves. Instead, it will be understanding that AI, like calculators and Wikipedia, can be both an incredible tool and a debilitating crutch.
What is the Right “Amount” of AI Use for Optimal Learning?
It’s pretty clear that the solution of having AI do your homework for you will not to result in much actual learning. Enamored as I am by the possibilities unlocked by the arrival of coding agents [3], it’s obvious to me that vibe coding actively impairs your ability to learn to write code by hand.
Whether this trade-off is valid in the realm of productive work is a more difficult question. I suspect the answer of whether vibe-coding enhances or undermines productivity depends a lot on the stage of knowledge acquisition. Beginners who haven’t built up a mental model of what code is being written are losing out on the opportunity to practice foundational skills. Experts, who can steer the process, may end up benefitting from the saved effort.
But nobody can deny that if your goal was to learn to write code by hand, getting Claude to do your work for you will result in less learning.
What’s more interesting to me is if there’s an optimal amount of or timing for AI help. Meaning, even if too much AI is obviously bad, might learning be hindered if there is too little?
If you had asked me this question in the lead up to my 2019 book, Ultralearning [4], I would have said no. I was a strong advocate for doing hard problem solving and the associated mental effort it requires as a way to learn difficult skills.
However, after the research that led to my 2024 book, Get Better at Anything [5], I’m no longer convinced. Practicing problems is certainly an essential part of learning any skill. But the cognitive load involved can easily become too high. Getting help, whether it’s in the form of a worked example, a completion problem or just a hint to get you unstuck, can be faster for learning than floundering.
Some Tentative Guidelines
In education research, conversations about different pedagogical approaches often span decades. On that timescale, AI has only been here for the blink of an eye. Therefore, I expect it will take some time before a mature consensus on AI best-practice emerges in the research literature.
However, I’d like to share my best guess of what some of those AI guidelines might look like, absent stronger evidence-based work:
1. You can’t learn from work you don’t do yourself.
This is obvious, but it needs to be restated. While AI-assistance may help or harm learning, it’s clear that anything you get a chatbot to do in toto isn’t going to allow you to learn the underlying skill.
2. Don’t use AI on your first attempt.
For many tasks, there is only one problem. In such cases, I think it’s almost always better to attempt a solution on your own before looking up the answer. This seems particularly true in creative domains where each problem is unique. Asking an AI for help with the start of an essay, for instance, will invariably shape what direction you end up following, robbing you of the crucial experience of building your own judgement around a topic and selecting your own path to research and argue.
3. Don’t use AI on your last attempt.
What if you can’t solve the problem despite your best efforts? Assuming this is a repeated problem, the research is fairly clear that seeing a worked example or explanation or getting instruction is going to be better than continuing to struggle. Thus, switching to getting help with AI after a failure seems prudent. But, if you’re going to benefit from learning, you need to then attempt another problem of the same type without AI help.
4. Do use AI to suggest alternative ideas and resources.
While starting with AI is likely to cause you to skip over the necessary hard thinking that results in learning, I do think AI can be useful in suggesting alternatives you might have overlooked. Here, the superhuman breadth of knowledge possessed by AI can be an enormous asset—surfacing methods, papers, books and ideas you may have missed in your first pass.
5. Do use AI to fix confusion and mistakes.
Similarly, after you’ve completed your first attempt at solving a problem, I would typically encourage looking at the expert solution for feedback. AI can not only generate expert-level solutions in many domains, but it can help point out the source of confusion and mistakes.
See-Do-Feedback (+AI)
In Get Better at Anything [6], I argued that much of effective learning looks like a practice loop: Seeing involves getting instruction, viewing a worked example or otherwise getting knowledge into your head to guide the search for the right answer. Doing involves practicing on your own. Feedback involves getting some relevant signal from the environment about how effective your practice was.
AI has the potential to assist with both the first and last element of that loop. What it cannot replace is the middle step—doing the work for yourself.