Note on currency. This page was compiled from material recorded in late 2024 / early 2025. LLMs and AI have changed significantly in the two years since — model capability, tool integration, and the range of legitimate use cases have all shifted. Some specific claims (about hallucination rates, source-finding, Kolbs feedback quality) may no longer hold to the same degree. The underlying principle — that AI should facilitate thinking rather than replace it — is likely timeless. Treat the specifics as a starting point for your own testing, not as settled rules.
The dominant failure mode when using AI for learning is not using it too much — it is using it at the wrong level. When AI chunks information, identifies connections, or evaluates what matters on your behalf, it produces a finished cognitive artifact without your brain having done the work that creates learning. The output looks correct. The encoding never happened.
The clearest framing: AI for learning is 95% the same as a Google search. The value depends entirely on the quality of the question going in and what you do with the answer coming out. The remaining 5% is the ability to handle more complex conversational queries — not a fundamental change in the relationship between tool and learner.
The Processing That Gets Skipped
Asking AI to do your cognitive work is the same shortcut problem that appears in every other technique:
- The weight doesn’t lift you. Paying someone else to lift a weight means the weight moves but you don’t get stronger. AI that produces your chunk structure, your connections, your importance rankings produces the artifact without the adaptation.
- It lowers the order, not raises it. When AI synthesizes and structures for you, the remaining task is memorizing what it gave you. It converts a higher-order problem into a lower-order one and hands you the downgraded version.
- The output looks like learning. A mind map AI built looks like a mind map you built. The difference only shows up when you try to use the knowledge.
The Meta-Level Rule
Never ask for the answer. Ask for the information that helps you figure out the answer.
The sequence: feeling → thought → question. Notice you don’t understand something → identify precisely what you don’t understand → ask for the specific missing piece that would let you resolve it yourself. The schema stays yours throughout.
The mental test: imagine a strict mentor who judges you on the quality of your questions. Low-level questions — “why is this important?”, “how does this connect?” — are questions that ask someone else to build your schema. A good question exposes a specific gap in your existing model and requests only what you need to resolve it.
Three Legitimate Uses
These keep the heavy lifting in your hands:
- Keyword seeding. Ask AI to generate the key terms and concepts for a topic before you start. Collecting keywords is necessary but cognitively worthless — the act of collecting them contributes almost nothing to learning. Outsourcing it is legitimate because the output is raw material for your own chunking, not a substitute for it. Do not ask AI to rank by importance — deciding what matters is the processing you need to do.
- Hypothesis validation. Once you’ve built a tentative model, test it: “I think A influences B in this specific way — is that right?” You constructed the hypothesis. AI validates or corrects it. The schema stays yours; AI fills only the specific gap that blocked your judgment.
- Gap-checking after retrieval. Do a full brain dump first, then ask AI to find gaps or missed perspectives in what you produced. The value is in the brain dump — active reconstruction, confronting what you can’t reproduce. AI flags the blind spots you’d circle past on your own, which is a legitimate but small contribution.
What Remains Off-Limits
- Asking AI to chunk, group, or find similarities in new material. This is the processing that creates the schema.
- Asking AI why something is important. Importance is relational — you’re asking it to build the relational structure for you.
- AI-generated analogies for things you’re learning for the first time. You have no way to evaluate whether the analogy is accurate. If it’s wrong, you encode the wrong model and may never know.
- Asking AI to organize keywords by importance. Introduces framing bias and removes the decision that should be yours.
- Using AI for Kolbs-style reflective feedback. Too nuanced for reliable AI output. The skill of giving yourself useful self-feedback is more valuable to develop than getting fast AI feedback.
Source-Finding
For research contexts, AI over-indexes on high-citation, popular sources. It hallucinates citations for niche queries and cannot surface specific, recent, low-citation work that may be the most relevant. Domain-specific tools (such as Research Rabbit) consistently outperform AI for source discovery because they catalog your existing papers and surface adjacent work by similar authors through their own algorithm.
Open Questions
- How do you ensure your brain doesn’t fall back to passive learning modes when AI is available? The shortcut is always easier and always available. What habits, constraints, or structural rules make the higher-order mode the path of least resistance?
- How do you offset the lower-order pull and keep thinking at higher levels? If AI’s default effect is to downgrade the cognitive order, what techniques or disciplines counteract that drift during a session?
- How can AI actively enhance higher-order thinking rather than just avoid degrading it? The risks are well-defined. The more interesting question is whether AI can be used to push thinking upward — generating counter-arguments, stress-testing models, forcing territory the learner would not have explored alone.
- Where is the line between cognitive tool and cognitive crutch? Short-form video and social media show how tools optimized for ease can structurally degrade attention and depth over time — not in any single session, but across habitual use. The risk with AI is similar: normalizing lower cognitive demand until it becomes the baseline. How do you make sure it is always pulling toward higher levels rather than making lower ones feel sufficient?