AI Learns for You. That's the Problem.

A while back I took the family to Kurashiki, and we visited the Ohara Museum of Art, the first private museum of Western art in Japan. Inside, tucked away, was a wing of Asian art, and hanging on one wall was a chart comparing the histories of Asia: China, Japan, Korea, timelines running side by side.

I know Chinese history fairly well, so that line of the chart didn't tell me much. But I kept scanning, one grid at a time, and at some point a thought showed up: I've lived in Japan for over three years, and I barely know its history.

For three years, everything on my mind was immediate. Visas, business, getting the kids into school. The last thousand years never felt like something I needed to know. Standing in front of that chart was a strange feeling. I live inside this country every day, and I know nothing about its past.

So I did what comes naturally now. I pulled out my phone and asked ChatGPT to recommend a few books on Japanese history and culture. It gave me a list. I picked two, downloaded them to my Kindle, fully intending to read them.

Then came the part that stuck with me.

The books sat on the Kindle for a few days. When I finally opened one, my first instinct was not to read it. It was to have AI summarize the key points first, so I wouldn't have to read slowly.

I paused for a second.

Because when I write code, untangle an architecture, hunt down a bug, or plan a workflow, I hand it to AI and things go fine. My efficiency genuinely doubles. But the book hadn't even been opened, and my reflex was to "run a summary pass" on it. I couldn't argue with myself on the spot, either. Summarizing saves time, right? That's what it's for. It's how I've always worked.

Then I remembered what Andrew Ng said recently. He came out with something he admitted he hadn't said publicly before:

"Frankly, AI models are terrible for learning. Students score higher on homework when they use AI — yeah, higher homework scores — but retention, their long-term performance, is much worse, because the AI did the work for them."

This is a man who has spent his career teaching AI to millions of people. Him saying AI is terrible for learning carries a different weight.

And I know exactly what he means, because he described my own habit better than I could. "I've asked AI models over the last six months... 'Building some project, how does this front-end, back-end component work... give me the answer, get the job done' — it was fantastic," he said. "But six months later, I don't remember the answer when I need to redo that front-end, back-end component. I ask the AI again."

That's me. That has been me for years.

My work and study have mostly lived on the technical side: writing code, rebuilding workflows, automating anything repetitive. These things share a property. The knowledge is a network, and what matters is how the dots connect. That part AI is genuinely great at. Ask it to walk through a tech stack and it hands you a clean map of the relationships in minutes. Reading the docs yourself takes ten times longer. Once the map is clear, the rest moves on its own.

So for years I "felt" that AI massively boosts learning efficiency, and in technical subjects that feeling never lied to me. New framework, new skill: I sped up, saved real effort, got real results. The hard part of a framework is the relationships. Get those and most of the understanding is done. On top of that, plenty of hands-on technical work that used to belong to people now belongs to AI. I state requirements and check the output. That's the arrangement, and it works.

History is different.

History has structure too, and AI can draw you a timeline in seconds. It could generate that museum chart instantly. But the chart isn't the core of history. The core is massive amounts of reading and remembering. Dynasty after dynasty, upheaval after upheaval, and concrete people making concrete choices in concrete situations. AI can lay all of that out beautifully, and none of it matters unless I read it into my own head, word by word, and keep it there.

Ask AI for a historical fact and it answers in two seconds. But the answer lands in the chat window, not in my brain. Close the phone and I still know nothing about that period. It never went through my own understanding, memory, digestion, thinking. It stays "something I looked up," never "something I know." Next time, I look it up again.

Language is the same. I can have AI translate an entire article, fast and flawless. The translated piece still isn't my Japanese. If I don't memorize the vocabulary myself, internalize the grammar myself, open my mouth and get it wrong and fix it myself, I'll never use the language freely. AI finished the article for me. It can't do the part where a language gets installed in your head. That part has no shortcut, and it takes the years it takes. No practice, no learning.

Somewhere in there, I started to see the pattern. AI's effect on learning isn't just uneven between technical and humanities subjects. The difference is categorical.

In technical fields, AI helps you sort the network. It accelerates the untangling of relationships, and once the relationships are clear, most of the understanding is done. In humanities, AI can't do the installing. It can compress an entire book into one page of key points, but that one page isn't memory. If anything, it leaves out too much. The speed at which a human brain absorbs and retains, as of today, AI cannot touch. Your eyes are yours. Your memory is yours. That part cannot be outsourced.

Ng's "cognitive offloading" fits neatly into this frame. In technical work, what AI takes over is execution, and execution is a load you're allowed to hand off. In humanities, the memorizing and thinking are the learning. Hand them to AI and what you've offloaded isn't the burden. It's the learning itself.

So the real question isn't "does AI help learning?" It's this: within each subject, which parts belong to AI, and which parts stop being learning the moment AI touches them?

That boundary differs by subject, and probably by person. There may be no standard answer. There's just the slow work of figuring out, subject by subject, how to live with this thing. When to hand it the wheel, and when to keep your hands on it. You only find out by doing.

Which is why I finally opened that Japanese history book on my Kindle. No summary first. Page one.

It's slow. But it feels steadier. When I hit something I don't understand, something worth talking through, I bring ChatGPT back out. It sits quietly next to me while I read.

In an era where everything keeps getting faster, this slow way of reading feels pretty good.