What shifts for learning with AI?

The shift

Explanation, worked examples, and one-on-one tutoring — the things that used to gate a learner's access to a subject — are now free, instant, and available at any level for almost any topic. But the mechanism that actually builds a durable skill was never the explanation; it was the effortful retrieval, the failed attempt, the struggle to produce the answer yourself. AI makes the answer abundant while leaving the struggle exactly as scarce as it was — and now optional, because nothing forces you to do the hard part anymore.

The axioms

• To learn a skill you need access to an explanation of it — scarce because good explanation required a teacher, a book, or a peer who already knew.
• Personalized instruction that adapts to what a specific learner doesn't understand builds skill fastest — scarce because it required a human tutor's time, one learner at a time.
• Skill forms through effortful practice: retrieving, attempting, failing, and correcting — scarce because it requires the learner's own sustained cognitive effort, which nothing can supply on their behalf.
• The friction of not having the answer is what forces the practice: you struggled because you had no choice, and the struggle is what built the capability.
• A learner needs feedback on where they went wrong to improve — scarce because informed correction required someone who already knew the material and had time to look at your work.
• Motivation to keep going through difficulty is required for mastery — scarce because it comes from inside the learner and can't be manufactured externally.
• Knowing what you don't yet understand — metacognition — is required to direct your own learning — scarce because accurate self-assessment is itself a hard-won skill.
• Reaching competence in a field takes a roughly fixed amount of practice time — scarce because there was no way to compress the hours of doing the work.

Invalid axioms

1. Access to explanation is the bottleneck to learning. Expert-level explanation, worked examples, and re-explanation at any level of depth are now free and instant across nearly every subject. The habit-trap: learners, courses, and self-improvement advice still treat "finding a good explanation / the right resource" as the hard first step and organize effort around acquiring material, when the material was never the scarce input — using it under effort was.
2. Personalized, adaptive tutoring is a scarce luxury. A patient one-on-one tutor that adapts to a specific learner's gap, at any hour, on any topic, is now available to anyone at near-zero cost — the thing Bloom's "2 sigma" research treated as the unaffordable gold standard. The habit-trap: we still frame personalized instruction as the premium, the thing you pay for or can't get, rather than the default that's now oversupplied — and we haven't shifted attention to the part it doesn't fix, which is whether the learner does the effortful work between explanations.
3. The friction of not having the answer is what forces the practice. For most of history, struggle was compulsory: you couldn't get the answer without doing the work, so the work got done. That coupling is now broken — the answer is one prompt away, so the struggle that used to be forced is now entirely optional. The habit-trap: teaching and self-study still lean on withholding the answer as the engine of learning ("try it yourself first"), a mechanism that only worked when the answer was actually hard to get, and now depends on the learner voluntarily not doing the easy thing.

Unchanged axioms

1. Skill forms through the learner's own effortful practice, and nothing can do that effort for them. Retrieval, struggle, and error-correction build the durable capability; watching a perfect explanation or reading an AI's answer produces the feeling of understanding without the substrate. This is the load-bearing one: AI made the answer abundant but the cognitive work that turns exposure into skill is exactly as scarce as before — it lives inside the learner and can't be offloaded. The more fluent and complete the AI's output, the easier it is to mistake having-the-answer for having-the-skill.
2. Motivation to endure difficulty still comes from inside the learner. AI can lower friction, encourage, and gamify, but the willingness to stay in the uncomfortable zone where learning happens isn't something an abundant explanation supplies — and when the difficulty is now optional, motivation carries more of the load than it used to, not less.
3. Metacognition — knowing what you don't yet understand — stays scarce and is now more load-bearing. A confident, complete AI answer masks the learner's own gaps particularly well, because the output reads as understood even when nothing was internalized. Accurately judging "do I actually know this, or did the model just know it for me?" is a skill AI doesn't confer and actively makes harder to self-assess.
4. Feedback only helps if the learner produced something to get feedback on. AI gives instant, informed correction — genuinely abundant now — but correction on your own attempt requires you to have made the attempt. The scarce input was never the feedback; it was the learner exposing their own reasoning to be corrected, which still requires them to try first.
5. Reaching competence still takes practice time that can't be compressed by watching. AI can make each hour of practice better targeted and remove dead time spent stuck or searching, but the number of effortful reps to automate a skill is a property of human cognition, not of information access. This is worth watching: better AI could genuinely raise the quality of each practice hour (tighter feedback loops, better-sequenced difficulty), so the claim is that it compresses the search and support around practice, not the practice itself.

New axioms

1. When answers are free, deliberate practice must be chosen, not forced by necessity. The scarcity that used to compel the hard work is gone, so the effortful part now depends entirely on the learner electing to do something harder than they have to. We don't yet have reliable ways — for individuals or the tools themselves — to make people opt into productive struggle when a frictionless answer is always one prompt away.
2. The gap between feeling you learned and having learned is now wider and easier to fall into. A fluent AI explanation produces a strong sense of comprehension while bypassing the retrieval that builds retention. Learners can no longer trust the feeling of "I get it" as a signal, and there's no established practice for distinguishing borrowed understanding from your own — the failure is invisible until you have to perform without the tool.
3. Two divergent populations are forming and look identical from outside. Some learners use abundant tutoring to run tighter, faster practice loops and genuinely accelerate; others use the same tool to skip the effort entirely and offload the thinking. The tools, the outputs, and often the short-term results look the same for both — so the divergence in actual capability is real, growing, and currently unmeasured.
4. Skill formation now has to be defended against its own best tool. The same system that is the best tutor ever built is also the most efficient way to avoid learning, and the two uses are one prompt apart. Learners, tool-builders, and anyone designing practice have to solve for a resource that helps and undermines through the identical action, with no external friction left to lean on. How much of this can be designed into the tools versus how much rests on learner discipline is unsettled, and hinges on product choices that are moving fast.

Where it breaks

The old engine of learning was compulsory struggle — you practiced because the answer was genuinely out of reach (INVALID axiom 3). That compulsion is gone, and in its place sits a new requirement to choose deliberate practice when a frictionless answer is always available (NEW problem 1). Learning was built to run on necessity and now has to run on volition, and neither learners nor their tools have replaced the missing friction with anything that reliably gets the hard part done.

A second collision: the abundant tutor produces a strong feeling of understanding (INVALID axiom 1 — access to explanation was supposed to be the hard part, and now it's effortless), while the capability that only comes from effortful retrieval stays exactly as scarce (STILL HOLDS axiom 1), and the gap between the two is now easy to fall into and invisible from inside (NEW problem 2). The learner who feels most helped — everything explained, instantly, perfectly — can be precisely the one who built the least, and nothing in the experience tells them so.

ai-axioms.vercel.app/axioms/education/learning-skill-formation-ai-axioms