Writing & Expression: Writing in the Age of AIFramework · Killing the AI Smell · Collaboration · Prompt as Writing

Treat AI as a "draft engine" and yourself as the editor-in-chief: AI owns volume, you own judgment. It's best at the step you dread most — facing the blank page; it's worst at the step you care about most — choosing, taste, deciding what to cut.

The whole value of AI is that it hands you that "terrible first draft" instantly — so you can spend your energy where it counts: revising.

Writing isn't one act; it's an assembly line: intent → gather material → draft → restructure → tune voice. AI is astonishingly efficient in the middle three; at the two ends — what you actually mean, and whose voice it's said in — it can barely help, because those require your judgment and experience, while it has only statistical averages. The most common mistake is outsourcing the ends too: the result is correct, fluent, and indistinguishable from anyone else's.

• ✓ Use for: routine email, first drafts, turning scattered notes into structure, finding counter-examples
• ✗ Beware: decision memos you haven't thought through — AI hides a hollow argument behind fluency
• Trap: pasting a whole AI draft — readers spot the "no person, no judgment, no risk" flavor at once
• Trap: asking AI for the "intent" — you forfeit writing's biggest dividend: being forced to think clearly

The "AI smell" is a scent of being correct yet hollow: every sentence flows, yet not one names a real person, time, or cost. You remove it with three moves — get concrete, cut words, break the symmetry.

Orwell described this flaw seventy years before LLMs — and it is precisely how they work: a language model literally "gums together" word-strips by statistical likelihood.

A model predicts "the most likely next word," so it slides toward the most average, least offensive, most common phrasing — the root of the AI smell: statistical mediocrity. The classic tells: the all-purpose opener ("In today's…"), an addiction to tricolons, words like "delve / leverage / crucial / it's worth noting," uniform sentence length, and a mandatory uplifting close. Each is a "strip set in order by someone else." The antidote is the old advice Orwell and Zinsser kept repeating: replace the abstract with the concrete, cut what you can, and deliberately break the rhythm — because a real person's sentences are uneven.

• ✓ Use for: anything carrying your name and building trust — newsletters, reviews, public statements
• ✗ Skip for: purely functional text (auto-receipts, template notices) where flavor is irrelevant
• Trap: only swapping synonyms — "essential" for "crucial" is still AI smell; replace the abstraction itself
• Trap: keeping tricolons and all-purpose openers — these are the strongest AI fingerprints

Don't aim for a draft in one shot. Make writing a multi-turn dialogue with AI: you supply judgment, it supplies options; you draw conclusions, it plays critic. The core claim must be yours — otherwise you lose the whole point of writing.

Hand the drafting to AI and you give up this chance to be "exposed" — the moment writing reveals the gaps in your own understanding.

Writing is thinking, not merely its record. Outsourcing the thinking is "cognitive offloading": easier now, but over time it erodes your ability to turn chaos into clarity. So the rule of division is: outsource search and critique, never the core judgment. The most powerful use is AI as sparring partner, not ghostwriter — you rough out your claim first, then ask it to play the skeptical reader: find the weakest argument, surface counter-examples you missed, propose three different structures. You think it through by answering it.

• ✓ Use for: decision memos, design docs, contested proposals — anything that must survive rebuttal
• ✓ Universal critic prompts: "Where would a reader not believe this?" "Find the weakest link."
• Trap: letting AI just "fix it" — you never see the holes in your reasoning, only a smoothed surface
• Trap: accepting the first answer — good thinking comes from the second and third follow-up

Writing a good prompt is writing. A prompt is a brief: whatever you leave unsaid, the model fills with its "average." A vague instruction must yield a vague draft.

The same law holds for prompts: if you can't think it clearly, you can't ask for it clearly.

A good prompt is the same craft as a good brief to a junior colleague or a clear design doc. It needs five parts: Role + Context + Task + Limits + Example. "Limits" and "Example" are the most underused — limits tell the model what not to do (length, tone, taboos), and an example hands over the standard of "good" directly. Every part you omit gets filled by the default: the most common, safest, most mediocre option. Learning to prompt is really practicing "stating requirements clearly" — exactly what a leader does to people every day.

• ✓ Use for: any repeatable high-quality output — one good prompt beats ten edits of a bad draft
• ✗ Beware: treating an ever-longer prompt as a cure-all — past five parts, switch to multi-turn dialogue
• Trap: giving only the task, no limits or example — the most common form of a "vague instruction"
• Trap: an AI-smelly prompt ("please go deep and be comprehensive") — vague modifiers buy no precision