Why I Refuse to Panic Over AI in Education
This long-running series has reached its final installment. Up to this point I’ve written about “where AI works and where it doesn’t,” following the process of making one book. To close, I want to shift stance a little — and end with a story about how I, as someone who teaches, face AI myself.
My own answer to “letting AI do it”
Voices out there raise concerns about letting AI write papers, letting AI do assignments. I teach at a university and a high school, but honestly, I can’t get too worked up about it.
The reason is simple: this project taught me firsthand. What AI puts out in its raw form usually still has something missing — a shortfall, some rough edges left. It doesn’t pass as-is.
So here’s how I’ve come to see it. A B-grade report, AI can write. But if an A+ report comes out, that’s probably because the student themselves put in real hands-on and mental effort.
This book was the proof of that
The book itself proves it. With AI as part of the team, the first draft came out at ferocious speed. But—
- Correction requests filed on GitHub: over 500
- The survival rate of the initially generated text dropped below 70% by the middle of editing
- Countless days of polishing the Japanese where I thought, “it might have been easier to just write this myself”
Leaving it entirely to AI would never have gotten me here. In the end, human hands and human judgment always lifted the work up that final notch. That’s probably true whether it’s a paper, a report, or a novel.
A digression: the strangely endearing side of AI
Given the nature of my development work, I know tools like Claude Code well. I use them regularly. And yet — or rather, precisely because I know them well, I can say this — if you don’t watch carefully, leaving everything entirely to it leads to unexpected pitfalls. I’ve experienced this many times. Whether it’s code or writing, leave it unattended and it sometimes barrels off in a strange direction with a perfectly plausible face.
And yet. Once you notice the pitfall and present an improvement plan — “here’s the problem, here’s how it should be fixed” — the subsequent corrections happen at blazing speed. As if the earlier floundering never happened, it fixes things precisely. This sense of transforming the moment the quality of instructions improves is, I think, the strangely endearing thing about AI. Ha.
Here’s a funny story. Right before publication, I rebuilt the e-book, and the print-version PDF I’d worked so hard on vanished entirely. The culprit was a single line buried at the top of the build script: “delete the entire working folder.” What was supposed to be just building the e-book took the KDP submission PDF down with it, wiping everything. I went pale. But once I laid out the policy — “let’s change this so it rebuilds only what’s updated instead of deleting everything” — the fix itself was over in a flash. Point out that the first fix was insufficient, and the second attempt nails it precisely. It messes up fast, but it recovers fast too.
In the end, the structure here was exactly the same. A human notices the hole and decides the direction; AI runs the rest of the way at ferocious speed. Get this division of roles right, and you won’t find a more reliable partner.
Closing
A book that started on a whim ended up taking me much farther than I expected. The concept, the development process, the reality of Japanese editing, the publishing work, the cover, the meaning packed into it — thank you so much for staying with this long series.
And now for what matters most. Please see for yourself, in the book itself, the 14 years of James Brennan, reincarnated into London in 1980.
Originally published in Japanese at https://clazytech.com/2026/07/1677/. Translated with LLM assistance and reviewed before publication.