Learning New Skills in the Age of AI
Learning new skills should be easier now.
We have instant access to explanations, examples, tutorials, summaries, and AI-generated guidance. Much of the initial friction in learning has been removed.
But more information does not automatically lead to greater understanding.
A lot of the time, it just creates more noise.
You can read multiple explanations of the same topic, get several reasonable-looking answers, and still walk away without a clear mental model of what actually matters. That is part of why learning can feel strangely harder now. You are not lacking access. You are trying to separate the signal from a growing pile of input.
That problem gets worse when the topic is complex enough that you can recognize the surface-level topology before you really understand the underlying system. Research on the illusion of explanatory depth found that people often believe they understand causal systems much better than they actually do, until they are forced to explain them in detail.
To say it another way, you can often sound informed well before you are actually grounded.
I wrote about the same pattern in Don’t Automate Until You Understand the Work. If you skip understanding too early, the speed feels good right up until the moment you own the consequences.
The Limitiations of Surface-Level Learning
Let me be clear, surface-level learning is not useless.
You need enough context to get oriented, ask better questions, and avoid wasting time digging into details too early. AI is genuinely good at that part. It can help you translate jargon, compare approaches, and get moving faster.
The problem is that surface-level progress can look complete when it is not.
- You can recognize patterns without understanding the tradeoffs behind them.
- You can follow an example without knowing which parts are essential.
- You can get working output without understanding why it works.
- You can borrow confidence from a good explanation without being able to recreate the reasoning yourself.
That gap usually doesn’t surface until:
- The example does not match your situation.
- Two options both look reasonable.
- Something breaks, and you have to debug it.
- You need to explain the decision to someone else.
That is where shallow understanding runs out.
I made a similar point in To Get Better Feedback, Show Your Work. Once your reasoning is visible, it gets much easier to tell whether you actually understand the problem or you just arrived at something that looks correct.
Use AI to Compress the Scale
AI is most useful when it helps you get to the real work faster.
That usually means using it for things like:
- getting an overview of a topic
- translating unfamiliar terminology
- seeing a few examples side by side
- comparing approaches before you pick one
- finding the first reasonable place to start
But there is a point where more explanation stops helping.
A 2025 Microsoft Research survey of knowledge workers found that people reported less critical thinking when they had higher confidence in AI or when the task felt lower stakes. That pattern matters because it is easy to offload judgment when the tool seems good enough and the consequences feel manageable.
AI can shorten the setup phase. It can reduce unnecessary wandering. It can help you get unstuck.
What it cannot do for you is decide which details are foundational, which tradeoffs matter most in your context, or whether your understanding is durable enough to hold up without assistance.
That part still belongs to you.
I see the same problem in Stop Chasing Tools—Focus on Outcomes. When the tool becomes the center of attention, it gets easier to confuse activity with progress.
|
|
Where can AI save you time?
My friends at Big Creek Growth put together a quick survey to spot the repetitive work you can hand off to automation.
|
|
Friction is Necessary for Learning
One of the easiest mistakes to make right now is treating all friction as inefficiency.
Some friction is waste. Some of it is just poor tooling, unclear docs, or bad onboarding.
But some friction is where comprehension gets built.
The learning science on desirable difficulties is useful here. The basic idea is that certain kinds of difficulty can make learning feel slower in the moment while improving long-term retention and transfer later.
That matters because AI removes a lot of the struggle that used to expose weak understanding.
If you want a skill to actually become yours, you need at least some moments where you cannot just keep reading and nodding along.
You need to do things like:
- explain the concept in your own words
- make a choice before seeing the answer
- debug the problem before asking for a fix
- predict what will happen before you run it
- rebuild a smaller version without copying the pattern directly
This friction is what forces you to form a mental model instead of renting one.
When To Go Deep into Learning
Not everything deserves deep study. Recognizing when to dig in is an important part of learning efficiently.
Sometimes, working familiarity is enough. You may just need enough understanding to review a proposal, follow a conversation, or complete a one-off task. Going deep on everything is going to leave you struggling to keep up with the growing demands for efficiency across engineering orgs.
The better question is: where will shallow understanding become expensive?
A simple filter helps:
- Frequency: Will I use this often?
- Failure cost: What happens if I misunderstand it?
- Leverage: Does this unlock better decisions in other areas?
- Ownership: Am I the one responsible when it breaks?
If you are touching something repeatedly, taking ownership of the outcome, or relying on it when conditions get messy, shallow learning tends to come back to bite you. It slows down your judgment. It makes debugging harder. It keeps you dependent on external answers longer than you should be.
I wrote about this in How to Become a Domain Expert. Depth is not just about collecting facts. It is about understanding the system well enough to make better calls when the obvious answer is not obvious.
How to learn faster without staying shallow
A better learning loop looks something like this:
- Orient quickly.
Use AI, documentation, or articles to get the lay of the land.
- Identify the core question.
What are you actually trying to understand or do?
- Reduce the surface area.
Pick one sub-problem instead of trying to absorb the whole topic at once.
- Attempt before you outsource.
Make a prediction, sketch a solution, or explain the concept before checking the answer.
- Use feedback to refine the model.
Compare your reasoning to what actually happened.
- Repeat where the stakes justify it.
Go deeper only where frequency, leverage, failure cost, or ownership says it is worth it.
That loop works because it uses AI for acceleration but still leaves room for effort, recall, and correction.
Research on retrieval practice consistently shows that recalling and reconstructing knowledge improves learning more than re-reading or elaborative review alone. The act of pulling an answer out of your own head is part of what strengthens it.
That is a useful counterweight to the current tendency to keep consuming explanations.
Sometimes the fastest path to real understanding is to stop reading and see what you can produce on your own.
I want to be clear here: The goal is not to avoid AI. The goal is not to make learning slower than it needs to be, either.
The goal is to use the tooling available to you without relying on it to replace your critical thinking.
Use AI to cut through noise, get oriented, and accelerate the early part of learning. Use independent effort to test whether the knowledge is actually becoming yours. Stay broad where the stakes are low. Go deep where misunderstanding is expensive.
That is what efficient learning looks like now.
Not consuming the most information.
Not getting the fastest answer.
Learning enough to know what matters, and going deep enough in the right places that the skill holds up when the happy path disappears.