Gawande uses AI tools across his current book. Asked what works, he names three specific uses and one explicit failure mode.
Research compression: Google on steroids. For the anesthesia history section: Three months of library research to two weeks to find, you know, getting down to, oh, here are 10 surgical accounts from the 19th century that are actually gold.1 AI accelerated source-discovery dramatically. Not perfect — it found great stuff but made up the quotes2 — but the source-discovery itself was real.
Fat-finding: Copy editing. I find it incredibly valuable helping me find where is the fat.3 AI is good at identifying redundancy, wordiness, places where the prose could be tighter. Acceptance of the suggestions requires writer judgment, but the surfacing is valuable.
Darling-killing: Cut-this-down requests. Gawande had a 4,000-word talk that needed to be 2,000 words. He used AI to produce multiple shorter versions. Everyone killed my opening anecdote and you know and I was like and I had to confess. I had to concede all right it's a it was a cute story but it wasn't really working or necessary for it.4 The AI's external view killed a darling the writer couldn't kill alone.
What AI cannot do: drafting. Getting it to draft stuff or um not useful or fruitful I I found at this point.5 The drafting itself remains the writer's work.
And one absolute prohibition: AI quotes are always wrong. Never trust AI quotes.6 Verify every quoted material against the actual source.
AI-as-research-fat-finder-darling-killer is the operational AI-writing-practice principle that names which AI uses help writer-craft and which don't. The taxonomy:
Help (acceptable use):
Don't help (limited use):
The discipline is using AI for its actual strengths (information-retrieval, pattern-recognition, external-perspective suggestion) while keeping the writer's actual work (drafting, judgment, argument) intact.
The need fires for writers either over-using AI (using it to draft, then publishing the AI-flavor as their own work) or under-using AI (refusing it entirely on principle, missing the legitimate research-and-editing speedups). Both extremes lose. The principle's value is operational specificity.
Three mechanisms:
AI excels at pattern-recognition over text. Source-discovery, fat-detection, redundancy-identification are pattern-recognition tasks. AI does them well.
AI fails at original judgment. Drafting and argument-construction require judgment AI doesn't have. The output reads as generic because it's pattern-matching to existing similar work, not generating from first principles.
Quote verification is non-negotiable. AI-generated quotes are unreliable. The verification protocol must be installed; trust-and-publish leads to disasters.
The principle pairs with writing-as-grappling-with-confusion: grappling requires original cognition AI can't provide. The AI is a research and editing tool, not a thinking-substitute. It pairs with the book-as-math-problem framework: AI compresses research time, allowing more of the math to be spent on writing rather than information-gathering.
Gawande needed material on how anesthesia spread in mid-19th century. Pre-AI workflow: months of library research, primary-source hunting, citation-chasing. AI workflow: prompt three different models for 19th-century firsthand accounts of surgeons discovering anesthesia. Get back results. Verify the results (quotes always wrong, but pointers usually real). Compress to 2 weeks what would have been 3 months.
The compression isn't free. The verification step is essential. But the time-compression is real and substantial. AI didn't write Gawande's anesthesia chapter; AI made the research input that the writer's drafting required possible in a shorter timeframe.
You're considering AI tools in your writing practice. Apply the operational taxonomy.
Acceptable uses to install:
Uses to refuse:
The discipline is using AI as accelerant for the work you're doing, not substitute for the work itself.
You'll know AI-over-use has trapped you when your work reads as generic. The voice has been homogenized. The fix is reducing AI use in drafting.
You'll know AI-under-use has trapped you when you're spending hours on research AI could compress. The fix is installing the acceptable uses.
Karlsson's llm-as-standardization-pressure-grammar-wildness-defense (vault concept) names AI's homogenizing effect on prose. Gawande agrees implicitly — he refuses AI drafting precisely because of this. Both writers converge on the use AI for input/edit, not output principle.
The compound insight: AI-as-research-not-drafter is tool-skill complementarity + cognitive-offloading risk applied to writer practice.
The Sharpest Implication
If you're considering AI in your writing practice, the diagnosis isn't AI good or bad. The diagnosis is which uses are acceptable and which atrophy your craft. Install the acceptable uses; refuse the atrophy-producing ones.
Generative Questions