I was halfway through a structured reading program — three sessions a week, extract one principle per article, write it in a daily note — when I realised I’d learned more about AI governance in one conversation about my son’s school refusal than in two weeks of reading.
The conversation went like this: my two-year-old won’t go to school. I mined an LLM for developmental psychology frameworks. That led to asking what humans are still good at when AI can retrieve any framework instantly. That led to asking whether any human advantage is permanent. And somewhere in that thread, I needed to articulate what AI governance actually means for a bank — not the textbook version, but the version that would help a real client make real decisions.
What came out was six questions: Should we build this? Is it working? Is it fair? Can we explain it? What happens when it breaks? Who’s accountable?
I didn’t get these from a reading list. I got them by thinking about what a bank actually needs to know about its AI systems. The questions reconstruct themselves when you understand the problem deeply enough.
This is the distinction that matters: retrieval vs reconstruction.
Retrieval is memorising “AI governance has six dimensions.” You can recall it in a meeting. But when someone asks “why six and not four?” or “which three matter for us?” — you’re stuck. You have the answer but not the understanding that produced it.
Reconstruction is understanding the problem well enough to derive the framework. You might produce five questions one day and seven the next, depending on the context. That’s not inconsistency — it’s adaptation. You’re not recalling a fixed answer; you’re rebuilding from the problem each time, which means the answer fits the situation.
Spaced repetition, flashcards, structured reading extractions — these are retrieval tools. They work for facts you need to recall but don’t naturally encounter: exam definitions, regulatory names, foreign vocabulary. For those, memorisation is the right strategy.
But for frameworks, judgment, and positioning — the things that actually drive decisions — retrieval is the wrong tool. You internalise these by thinking through real problems, not by reviewing bullet points. The consulting readiness program I was following had me extracting principles from articles three times a week. The HSBC pre-work, which forced me to think about what a real bank actually needs, taught me more in one session.
This connects to something uncomfortable about AI. Large language models have perfect retrieval. Ask for the six dimensions of AI governance and you’ll get a polished answer instantly. What they don’t have — yet — is the ability to reconstruct in context. To know that for this bank, with this regulator breathing down their neck, only three of the six questions are urgent, and here’s why.
That reconstruction ability is what humans still bring. It comes from having thought through the problem in specific situations with real stakes — not from having memorised the general framework.
The practical implication for learning in an AI world: stop optimising for recall. The model already recalls better than you. Optimise for understanding deep enough that you can rebuild the answer from the problem it solves. When someone asks “why?”, you shouldn’t need to look it up. You should be able to derive it.
The reading program was training retrieval. The real work was training reconstruction. I cancelled the reading program.