Oct. 13, 2026
AI can be an efficiency multiplier when used to automate something you already know how to do well.
It can be a fantastic tutor, if it’s well-calibrated.
But it cannot replace the learning process.
The process of entangling yourself with a concept (or web of concepts), by working through it slowly, is how you come to understand something deeply.
Here are some BAD examples of AI applications for learning…
Mistake 1: Dropping documents into an LLM (like Gemini Notebook) and then reading the summary rather than the documents themselves.
(Same goes for tactics like generating a transcript of a video, which is then summarized or searched through for key terms.)
Opting for a high-level summary (alone) bypasses learning.
Now, to be fair, this approach is perfectly fine when you don’t need to understand what the documents say…
- When you just want a surface-level overview, this works great.
- When you want to convert the information into a different format (e.g., Notebook can create slide presentations, ad hoc podcast episodes for you to listen to, etc.), this is one of the best approaches available.
- When you need to find a needle in a haystack, this might work phenomenally—provided you follow up and verify whatever is found by checking the actual document yourself.
- When you want the big picture first, before diving in more deeply.
In other words, when the goal is NOT learning or deep understanding, then by all means drop the documents into an LLM and have it summarize or conduct related tasks.
But note that none of these things replaces the experience of actually reading the documents yourself.
There is no Matrix-style download of the information into your brain.
Even if you walked away with a decent summary of the key points, you still wouldn’t be changed by the experience in the same way you would’ve been had you actually read the documents yourself.
A change in who you are through experience can only occur through that experience. There’s no way around this.
The summary of the thing is not the thing.
Mistake 2: Having AI replace your active role in the knowledge generation process, such as through…
- Generating code that you run without really understanding
- Debugging code for you before you’ve tried to troubleshoot it yourself
- Writing your school assignment for you
- Solving your homework problems for you
- Fabricating responses to prompts during online trainings
- Et cetera
This feels too obvious to warrant saying, but it’s one of the most prevalent pitfalls!
Take the example of having AI write your code for you… This isn’t wrong per se, but it’s the worst imaginable way to learn to code.
It’s like walking around with a translator in a country where you don’t speak the language. The translator can help you get what you want, but you’re not going to pick up much of the language along the way.
Additionally, if you ever fail to convey the correct message to the translator, then the outgoing message is wrong, and you may have no way of knowing.
Again, we should ideally use AI for applications that we have already mastered ourselves. Its purpose should be to automate processes we already know how to do well and for which we can quickly verify the outputs.
Mistake 3: Treating the AI like an omniscient entity, whose proclamations are accepted unquestioningly.
Large language models like ChatGPT are basically the distilled voice of the Internet. That includes a lot of good, useful information, as well as a lot of terrible, inaccurate information.
Would you believe that a sentence formed using your smartphone’s predictive text was necessarily true?
Neither would I.
Obviously, LLMs are far more advanced than predictive text on a phone, but this is a pretty helpful simile to keep in mind.
No one who understands the mechanics of LLMs, even at just a basic level, would ever regard them as omniscient—or as an entity, for that matter.
By the way, I know it’s popular to say that “nobody understands how LLMs work,” but that’s not really true. The people designing these things do understand how they have written the code.
No one was ever writing the code that produced ChatGPT and said, “What’s happening? How are my fingers hitting these keys? I don’t understand!”
Nor was the code underlying these technologies found on alien stone tablets in the desert.
Humans wrote those instructions. The mechanics of LLMs are fully understood.
What’s NOT 100% understood is exactly how any particular LLM output is produced. For example, Why did it say X rather than Y? The process is too complex to follow at present. (Although researchers are continuously working on this.)
Anyway, the point is that we don’t need to feign a level of ignorance that turns AIs into deities.
The Upshot: If you want to learn something at a deep level, then you have to put in the reps.
There is no other way.
Use the AI as a supplemental tool on the edges; use it as a personalized tutor; just don’t use it to offload your entire brain.
BY THE WAY…I wonder what effects there might be of cognitive offloading during key brain development periods (e.g., adolescence)?
