This is Tabbit's "Tinsight · Insight" column, where we share in-depth articles worth spending time on. Previously we shared "The Real Divide in the AI Era" from offbook.press. The new series "Rebuilding Individual Learning" focuses on providing a string of practical, actionable solutions. The full text is about 19,000 words and will be delivered in 6 installments — this is Chapter 5. In this chapter, the author focuses on the boundaries of what AI can and cannot do in personal learning, helping us understand how to use AI tools properly to improve learning quality, rather than falling into the trap of passively consuming information.
Author: Dawei Geng
Source: https://offbook.press/essays/rebuilding-learning/
Chapter 5 · What AI Can and Cannot Do
The previous three chapters were all about doing things with AI: having AI review you, having AI mine your experience, having AI help you survey a new field. It sounds like "you can outsource a lot of things to AI."
But that sounds wrong.
What this chapter deals with is: exactly which parts can be outsourced, and which parts cannot. Once this boundary blurs, all the earlier methods will, within a few weeks, degrade into advanced versions of the three patterns described in Chapter 1. You chat with AI a lot, read a lot, "understand a lot" — but nothing has actually happened to you.
▍The 80% AI Can Do
The volume of things AI can do is enormous.
Organizing and structuring information. Turning your scattered observations into models, mapping a field's knowledge into a map, structuring multiple documents into a coherent understanding — AI does all of these extremely well, and tens of times faster than you.
Reviewing and rebutting. Using any thinking tool to review your judgments, challenging your assumptions from multiple perspectives, pointing out the flaws and blind spots in your reasoning — in many scenarios AI's review quality approaches that of a serious expert.
Role-playing. Playing a seasoned practitioner in a certain field, playing your opponent, playing a rational but picky peer — this kind of role-play lets you see problems from different angles at nearly zero cost.
Recording and tracking. Writing down your hypotheses and your predictions, then months later automatically reminding you "you made prediction Y at time X — what's its verification status?" In the past, this kind of tracking required you to maintain a log yourself; AI can do it for you.
Knowledge integration. Connecting what you've learned with what you already have, discovering contradictions, building new connections — this kind of integration used to require a lot of your own reflection; AI can accelerate it.
Draft generation. Writing out the ideas already in your head in some format, doing preliminary polishing, offering several versions of wording — these execution-level tasks AI does faster than most people.
Taken together, these are the 80% of your learning process that "can be outsourced." Hand them to AI, and your time and energy are completely freed up.
▍The 20% AI Cannot Do
But the remaining 20%, AI cannot do — and this 20% is the center of gravity of everything.
One: Forming your own initial hypothesis
AI can generate many hypotheses. Ask it to list possible causes of a phenomenon, and it can list a dozen or more. But these are AI's hypotheses, not yours.
Your own hypothesis is the starting point you raise based on your specific situation, your existing experience, and your implicit intuition. It carries your unique perspective on the matter. The hypotheses AI generates don't have this "you" — they are averaged, standardized hypotheses that fit most situations.
The starting point of the learning process must be your hypothesis. Without your hypothesis, the entire loop has no driving force, because the "hypothesis being reviewed," the "viewpoint being rebutted," the "judgment being verified" — these "hypotheses," "viewpoints," "judgments" must all exist first, and they can only come from you.
Two: Choosing the direction
AI can tell you "based on this information, there are three possible directions: A, B, and C." But which direction is worth pursuing is your decision.
AI has no preferences, no mission, no sense of cost. It takes no stance on "which direction to go" — to it, all directions are equivalent possibilities.
You are different. You have time, energy, and opportunity costs. You must choose one among A, B, and C, and this choice is necessarily based on your value judgment — the one thing AI does not have.
Three: Making real predictions
AI can list "possible predictions": "this could develop toward A, or B, or C." But that is not a prediction. The essence of a prediction is betting on which one will happen, and being willing to bear the consequences of that judgment.
AI has no capacity to "bet." It has no win or loss, no consequences, nothing it must bear. The "predictions" it offers are all probabilistic statements of possibility — they sound comprehensive, but they are not decisive.
You have to make the call yourself: "I bet A will happen." This call is the key link in the learning process, because only when you truly commit does the subsequent verification become meaningful.
Four: Applying it in real situations
AI can tell you "this method works well in similar situations." But your specific situation is always more complex than any description: your specific users, your specific team, your specific constraints, your specific timing.
Applying a method to your specific situation requires judging the uniqueness of that situation, identifying the method's boundary conditions, and adjusting the method's concrete operations. These judgments AI cannot make — it is not in your situation, and it does not have your specific feeling.
Five: Bearing the consequences of wrong judgments
This is the most fundamental one: if the judgment you make is wrong, you bear the consequences. AI does not.
Bearing consequences has a huge cognitive effect — it keeps your judgments from being a joke. When you know that every judgment you make will bring real results (good or bad), your judgment process automatically becomes more cautious, more careful, and more honest.
AI has no such mechanism. It can "look cautious," but that is the caution of language, not the caution forced out by consequences. The quality of the two is worlds apart.
This is why AI cannot learn. It can generate very elaborate reasoning chains, but its reasoning will not self-correct because of "consequences." You can. Every time you bear the consequences of a wrong judgment, your cognitive system receives the most genuine calibration.
▍Other boundaries of AI
Beyond this fundamental 20% boundary, there are a few technical boundaries of AI worth reminding you of.
One: the training data cutoff. The model has a cutoff date for its training data. It has no direct knowledge of things after that date. Although most mainstream models now have real-time search and can look up new things, what search returns and the deep understanding internalized during training are not the same thing — search gives you "what it found," training gives you "what it understands." Keep this in mind when discussing the latest events, the latest research, and the latest developments.
Two: bias. The model's training data carries bias — usually Western-centric, English-centric, and mainstream-media-centric. Ask it a question involving China's specific context, and its answer may carry default assumptions from an American perspective. Ask it a question involving a minority viewpoint, and it may lean toward the mainstream view.
Three: hallucination. Models will confidently fabricate facts: false citations, nonexistent research, invented names and institutions. The less familiar the domain, the higher the hallucination rate. Keep a dose of skepticism toward the specific factual claims AI makes.
Four: context limits. There is an upper bound on how much information you can give AI in a single conversation. Beyond a certain amount, it starts to "forget" earlier information. Consistency in long conversations is something you must actively maintain.
These boundaries are not meant to keep you from using AI — they are to let you know where AI can go wrong, and where you need to compensate yourself.
▍AI is a mirror, not an answer
Put everything above together.
AI is not a tool that gives you answers. It is a mirror that lets you see yourself clearly.
Use it to review your judgments — it reflects the flaws in your judgments. Use it to mine your experience — it reflects the structure you already have. Use it to interrogate new fields — it lets you see the relationship between a new field and the fields you're familiar with.
In every use, you are the protagonist. Your judgments, your experience, your interrogations. AI merely lets these things be seen more clearly by yourself.
Those who treat AI as an "answer provider" (asking AI "what should I do," "which option is better," "what should I do in this situation") — the answers they get are all averaged, standard, and devoid of personal specificity. These answers apply loosely to anyone, so they do not truly apply to the specific you.
Those who treat AI as a mirror (having AI reflect what is already in their heads, helping them organize, helping them interrogate, helping them calibrate) — what they get are insights about themselves. These insights apply only to themselves, but they truly apply.
This is the fundamental difference between two ways of learning in the AI era. The former is consumption; the latter is construction. The former increases your "knowledge" but leaves your capability unchanged; the latter makes your judgment system truly grow.
To be continued...
Chapter 1 · How Most People Use AI Damages Cognitive Decoupling
Chapter 2 · What Questions Are Worth Asking
Chapter 3 · Modeling with AI (Part 1): Thinking Tools as Reviewers
Chapter 4 · Modeling with AI (Part 2): Cross-Domain Isomorphism, Starting from the Mirror You Already Have
Chapter 5 · What AI Can and Cannot Do
Chapter 6 · A Starting Point: AI Programming
Chapter 7 · Finally
