Key Takeaways — Chapter 29: Writing with AI
The summary card. If you remember nothing else, remember this page.
The one idea
AI can draft, but it can't think for you. A large language model is a fluent next-word predictor that optimizes for plausible, not true — so it produces the writing without doing the thinking the writing was for (Chapter 1). Use it to revise thinking you've already done; never to replace the thinking itself. The judgment about whether the text is true, right, contextual, and yours stays entirely with you, no matter how good the model gets.
🚪 Threshold concept: The question is not "can AI write this for me?" but "have I done the thinking, so this tool can help me express it?" Cross that line and an empty, fluent draft stops reading as "a good start to polish" and starts reading as a warning — a sign no thinking has happened yet.
The load-bearing fact (everything follows from this)
The model produces plausible text, not true text. Most plausible text is true (the tool is useful); some is false and looks identical (the tool is dangerous). It isn't lying when it errs — lying needs a known truth to betray, and the model has no store of facts. It's doing its only job, and a plausible-looking falsehood is just that job producing a wrong answer.
Good at / bad at — and why
| Genuinely good at | Genuinely bad at |
|---|---|
| Brainstorming, outlining | Accuracy (hallucination) |
| Rephrasing your own sentences | Calibrated nuance (hedging vs. proving) |
| Summarizing (then verify) | Institutional/contextual knowledge (your situation) |
| Low-stakes first drafts | Original thinking (it gives the average take) |
| Format/mechanical conversion | Knowing your specific audience |
The "good" column shares one feature: you stay the judge, the task is mechanical, or the stakes are low. The "bad" column shares one root: plausible text without a held meaning — the five failures are one fact seen five ways.
The reframe: revision, not replacement
Do the thinking → write the draft in your own words → then bring the model in to react, argue against you, pressure-test structure, and rephrase. The test: did the idea originate with you? If yes, augmentation (good). If the model supplied the substance, outsourcing — you handed away the thinking that was your job.
Prompting = writing a good brief
The four levers — role, audience (highest-leverage), constraints, examples (few-shot beats description) — are mostly Chapters 2, 4, and 7 aimed at a machine. A good prompt requires you to have thought (you can't specify an audience you haven't considered), so it sneaks the thinking back in. Vague prompt → generic mush.
Integrity: three rules
- Disclose per your context (academic: usually required; workplace: usually not; journals: increasingly required). When unsure, disclose.
- Verify everything — accuracy is non-delegable. "The AI told me" is no defense (lawyers have been sanctioned for unchecked hallucinated citations).
- Don't present as yours reasoning you can't evaluate.
AI-assisted (you thought, it helped revise) is usually fine; AI-generated (it supplied the substance, you passed it off) is the problem. Academic writing is strictest because its purpose is to demonstrate your thinking.
The governing rule (the whole chapter in one line): If you cannot evaluate whether the output is correct, you should not use AI for that task. The tool is safe exactly as far as your judgment reaches — and most dangerous in the very situation where it's most tempting: when you don't know the subject and want the model to know it for you.