You Cannot Outsource Understanding
Every medium that teaches also teaches a model of what learning feels like. A book teaches one pace, a classroom teaches another, a video teaches another, and a feed teaches the fastest one of all: receive the shape of an idea, feel the click, move on.
This "click" is worth being suspicious of.
The modern internet is extremely good at making learning feel like it happened before much has actually happened. You can watch a video on ego death, another on quantum mechanics, another on politics, another on why the previous video was wrong, and after an hour feel as if your mind has been doing serious work. Sometimes it has. Often it has only been carried through someone else's thoughtfully curated path.
The promise is almost too obvious: learn, and become less trapped in yourself. Even humility has to arrive on demand. Even getting outside yourself gets turned into content for the self to consume.
An idea can be introduced instantly. It cannot be digested instantly. The difference between those two things is most of what people call understanding.
intellectual content is still content
Veritasium, Vsauce, 3Blue1Brown, and channels like them have probably brought more people into difficult ideas than most formal institutions could have on their own. A good explanation can make a subject available. It can remove unecessary difficulty. It can show the structure that textbooks sometimes bury under notation or bad writing.
A clean video is clean because someone else already did the messy part. The order of examples, the analogy, the pacing, the animation, the missing boring parts, the satisfying ending: all of it is the result of work done before the viewer arrives. The viewer receives the finished path.
That usefulness is also where the illusion starts. The idea feels fluent because the presentation is fluent.
Cognitive scientists call one version of this the illusion of explanatory depth. People often believe they understand ordinary mechanisms until they have to explain them in detail. Rozenblit and Keil found that confidence drops when people move from recognizing an explanation to producing one themselves. [[1]](https://peril.lol/blog/you-cannot-outsource-understanding#source-1)
Good intellectual media carries this danger precisely because it is good. The better the explanation, the easier it becomes to borrow the creator's fluency and mistake it for your own.
input is not thought
Learning starts after the input. A video, lecture, book, thread, class, or conversation gives material. The mind still has to metabolize it: connect it to other things, argue with it, apply it badly, fix the application, lose the decorative parts, and find the piece that survives.
This does not mean every video needs notes. Nobody needs to turn a Minecraft philosophy video into a doctoral seminar. If nothing in the idea changes shape inside your own head, then the idea passed through you more than it entered you.
A math video can make a concept feel obvious. The problem set tells you whether the obviousness was yours. A philosophy video can give words to something you half-believed already. The argument starts when you ask what would make it false. Commentary can sharpen taste, but taste that only selects more commentary has not left the feed.
The learning research points in the same direction. Karpicke and Blunt found that retrieval practice produced stronger long-term learning and transfer than additional studying with concept maps. [[2]](https://peril.lol/blog/you-cannot-outsource-understanding#source-2) The generation effect says that people remember material better when they generate it themselves instead of only reading it. [[3]](https://peril.lol/blog/you-cannot-outsource-understanding#source-3) Self-explanation research shows that students often learn more when they explain examples to themselves while working through them. [[4]](https://peril.lol/blog/you-cannot-outsource-understanding#source-4)
The common thread is not complicated. Understanding gets stronger when the learner has to produce something.
the feed takes away a place to stop and think
The feed does not just add bad ideas to a person's life. Bad ideas existed before phones. What it changes is the amount of time between contact and reaction, which is often the exact space where digestion would have happened.
An idea arrives already surrounded by reactions, corrections, takedowns, and explanations of what the whole discourse means. Each layer can be useful, but each layer can also move you farther away from the thing itself.
At some point the object disappears and the social positioning remains. The question stops being "what is true here?" and becomes "what kind of person believes this?" or "what does my reaction prove about me?" That is where intellectual life turns into sorting.
This is why brainmaxxing and similar language feels so off even when the desire underneath it is understandable. Wanting to be smarter is normal. Wanting to learn more is good. The rot starts when intelligence becomes an object of constant measurement and display, because attention moves from the work to the person trying to look like the work is happening.
That connects to my agency argument, but it is not the same claim. In that post, the problem was self-monitoring during action. Here the problem is mistaking received structure for internal structure. In both cases, the self gets inserted into a process where it mostly makes things worse.
school has the opposite disease
School has its own version of this. A class can teach real things. A good teacher can force contact with material you would never have chosen voluntarily. A curriculum can create a path through a subject that would otherwise stay opaque.
Then the institution often measures the wrong leftover. The test, the rubric, the grade, the required annotation, the discussion post, the fake reflection: sometimes these track understanding, and sometimes they train the production of evidence that understanding happened. This is why school and YouTube feel like opposite failures. YouTube can give beautiful input with no test. School can give useful input and then test a corrupted proxy.
Neither one can complete the last step. The learner still has to turn the material into something usable outside the format that delivered it.
where LLMs fit
LLMs make this more dangerous and more useful at the same time.
Used badly, they are the most efficient machine ever built for avoiding intellectual digestion. Summarize the book. Write the essay. Explain the video. Give the take. Make the friction disappear. In that mode, AI does not solve the problem of passive consumption. It industrializes it.
Used well, the same tool can sit in the gap between exposure and understanding, which is the part I care about because it matches how I actually use it. I can watch a video, then regurgitate what I think I understood and see if the explanation holds. I can ask whether two ideas that seem unrelated connect in any real way or whether I am forcing it. I can ask for a tiny site, simulation, or piece of code that makes a mechanism visible. I can throw a half-formed thought at something responsive and get enough resistance to keep thinking.
That is not outsourcing understanding so much as outsourcing some of the scaffolding around understanding. The distinction matters. If the model gives me a finished answer to repeat, I skipped the work. If it makes me explain, compare, test, and revise, it is helping with the work. The value is not that the model understands for me. The value is that it gives my own understanding something to push against.
Research on LLM tutoring is still early, but the useful direction is already visible: AI is more promising when it supports reflection, questioning, and metacognition than when it only gives answers. [[5]](https://peril.lol/blog/you-cannot-outsource-understanding#source-5) That fits the older learning research too. The learner still has to retrieve, generate, explain, and transfer. A chatbot can assist that process. It cannot replace it without hollowing it out.
the uncomfortable part
The feeling of understanding is easier to get than understanding. It is cleaner, faster, and more social. You can get it from a video, a podcast, a thread, a class, a quote, a summary, or a chatbot response. Real understanding is less flattering. It involves misremembering, restating, testing, asking dumb questions, building small models, and finding out that the beautiful explanation breaks when you try to use it.
This is not an argument against consuming things. Consuming things is how most ideas enter. The problem begins when the entrance gets mistaken for the whole thing.
You cannot outsource understanding. You can outsource exposure, explanation, examples, diagrams, summaries, code, and counterarguments. Those are useful. The final step still has to happen in your own head. The idea has to collide with what is already there and come out changed.
If it does not, you did not learn it. You watched someone else think.
- * *
**Sources:**
1. [[1]](https://peril.lol/blog/you-cannot-outsource-understanding#ref-1-1) Rozenblit, Leonid, and Frank Keil. "The misunderstood limits of folk science: an illusion of explanatory depth." _Cognitive Science_ vol. 26,5 (2002): 521-562. doi:10.1207/s15516709cog2605_1
2. [[2]](https://peril.lol/blog/you-cannot-outsource-understanding#ref-2-1) Karpicke, Jeffrey D., and Janell R. Blunt. "Retrieval practice produces more learning than elaborative studying with concept mapping." _Science_ vol. 331,6018 (2011): 772-775. doi:10.1126/science.1199327
3. [[3]](https://peril.lol/blog/you-cannot-outsource-understanding#ref-3-1) Slamecka, Norman J., and Peter Graf. "The generation effect: Delineation of a phenomenon." _Journal of Experimental Psychology: Human Learning and Memory_ vol. 4,6 (1978): 592-604. doi:10.1037/0278-7393.4.6.592
4. [[4]](https://peril.lol/blog/you-cannot-outsource-understanding#ref-4-1) Chi, Michelene T. H., et al. "Self-explanations: How students study and use examples in learning to solve problems." _Cognitive Science_ vol. 13,2 (1989): 145-182. doi:10.1207/s15516709cog1302_1
5. [[5]](https://peril.lol/blog/you-cannot-outsource-understanding#ref-5-1) Li, Yixuan, et al. "Ruffle&Riley: Insights from Designing and Evaluating a Large Language Model-Based Conversational Tutoring System." arXiv, 2024. arXiv:2404.17460