How to use AI without losing your voice

1 The bottom line

Use AI to attack, check and offer options on a draft you wrote yourself, and keep every sentence and every conviction your own. The thinking is the point, and a machine that writes for you takes it away.

2 Key takeaways

• Write the draft yourself before AI sees it. Writing is how you find out what you think.
  • Morgan Housel9:16 says that if an LLM does even the broad structure, you skip the thinking and lose what is good about writing.
• Ask AI to attack your draft, then fix the draft yourself. It shows you blind spots you cannot see.
  • Jimmy Soni34:01 says nine times out of 10 the counterarguments force him to stop and think.
• Use AI for options, then judge them yourself. Reacting to something is easier than starting from nothing.
  • Sam Altman37:48 calls ChatGPT a super thesaurus, but says it will not replace coming up with the ideas.
• Verify every quote and fact it gives you. It makes things up with full confidence.
  • David Perell1:08:10 published a hallucinated Steinbeck quote to about 30,000 people before readers caught it.
• Protect the oddness that makes your voice yours. AI and grammar tools pull writing toward the average.
  • Robert Macfarlane1:10:40 treats Grammarly's disapproval of his deliberately ungrammatical lines as a badge of success.

once I see how it's criticized the piece I can then go back to my piece and strengthen it against those criticisms

Jimmy Soni, The Founders

2.1 Draft on your own first

Do the thinking and the first pass without AI, so your intuition and your sentences stay yours.

• Write the first draft yourself, because writing is what makes you think. Morgan Housel9:16.
  • Housel says most authors start with one brave sentence. That sentence teaches them something and leads to the next paragraph.
  • Dan Wang1:06:00 may declare it verboten to let AI into his writing, because he wants every sentence to come from himself.
  • Wang still uses AI as a sparring partner. After rereading The Red and the Black he asked what critics had said about it.
• Get your own idea down before you ask AI anything. Alain de Botton1:23:47.
  • He says AI only summarizes what has already been thought. If he gave it an essay on nostalgia, it would crush his nascent intuition.
  • He writes the essay first and only then asks what he has missed.
• Put in what only you would notice. Ezra Klein1:43.
  • Klein says AI typically sees what everybody else would see. A better prompt does not fix that.
  • Perell points to a line in Tyler Cowen42:06's piece on Indian food. He and his wife ate lobster on a Kerala houseboat, cooked from what he bought from a local fisherman. No AI could have supplied it.
  • Maria Popova48:28 says only the writer knows the feeling out of which a piece was created.
• Ask for background on a new topic, then write without pasting it in. Tyler Cowen8:15.
  • Cowen asks o1 Pro for the legal background before writing on a subject he does not know. He does not use what it gives him, but it stops him sounding ignorant.
  • He did this for a column on declassifying classified documents.

2.2 Ask AI to attack the draft

A finished draft is the right moment to ask a machine where you are wrong, and to rebuild the piece yourself.

• Ask for a takedown and tell it not to hold back. Jimmy Soni57:34.
  • Soni finishes an op-ed, then asks Claude for a 700 word criticism. He sometimes adds a personality.
  • He reads the objections and goes back to strengthen his own piece against each one.
  • He took the idea from Tyler Cowen's point that you must be able to argue the opposite position.
• Tell it to attack, and then do not add hedges. Packy McCormick41:41.
  • Packy finds Claude and ChatGPT weak editors that mostly flatter. ChatGPT does well when told: "Your job is to cynically attack this. Where did I get this wrong?"
  • He uses it to catch factual blind spots.
  • The risk comes after. Piling on qualifiers weakens a piece that needs a point of view.
• Ask what readers will find obnoxious. Tyler Cowen10:16.
  • Agnes Callard suggested asking the AI what in your writing some people will find obnoxious, and to explain in detail. It flagged a passage of Cowen's as supercilious and condescending.
  • Cowen had considered keeping the passage. The explanation convinced him to change it.
• Ask for the weakest and most boring parts, and answer its questions. David Perell44:30.
  • Perell asks for the weakest points of his argument, the most boring parts, the best parts to double down on, the unclear transitions and what a story still needs.
  • The feedback is instant and free. He says it is not the best editor in the world, but it is always available.
  • On a story about his mother being mugged, the AI asked what the night was like. He answered that it was foggy, and the fog became a useful motif.
• Brief it with your own context, in a long first message and short follow-ups. Tyler Cowen13:59.
  • Cowen dictates his opening message by voice for a minute and a half to three minutes. After that his follow-ups stay short.
  • He says people find AI mid because they will not put in enough of their own time12:30 generating context.
  • Steven Johnson2:20 says the main skill in prompting is clear, persuasive prose. Plain feedback works too. A model once turned Hawaii's climate into a symphony of Mother Nature, and the reply "dude that is a little over the top" got a better version.

2.3 Cut, and ask for options

Let AI show you the fat and the alternatives, then do the rewriting and the choosing yourself.

• Compress your draft and see what survives. Michael Dean1:17:37.
  • Dean gives ChatGPT a 3,000 word draft and asks for versions at 10,000 words, 100 words and one sentence.
  • When the short version carries the idea, he sees he was rambling. He then rewrites it himself.
• Ask several models to cut the piece, and compare. Atul Gawande1:02:30.
  • A 4,000 word talk needed to be 2,000. Gawande asked three platforms to cut it and got six versions. None was the right length, but all helped him find the fat.
  • Every version cut his opening anecdote. He admitted it was cute but unnecessary. One model kept returning 500 and 1,000 word versions when he asked for 2,000.
• Use it as an editor who is always awake. Jimmy Soni57:06.
  • At four in the morning Soni pastes in a paragraph that feels flabby, or a word that does not sit right. He asks what is wrong with it, or for other options.
  • He compares it to moving from a typewriter to a word processor. The research, interviews and words are still his own.
• Generate options, then say yes or no. Nat Eliason1:38:45.
  • When stuck on a description or a plot turn, Eliason asks for ideas, rejects most and combines the pieces he likes. Even a bad suggestion shows him what should happen instead.
  • For a whole paragraph, Michael Dean1:23:45 asks for ten versions. He pastes none. He weaves the good parts of the second, seventh and eighth into his own words.
  • Eliason compares himself to Miranda Priestly in The Devil Wears Prada, who never designs anything but says yes or no. He says the writer's taste is the competitive advantage.
• When one word will not come, paste the sentence with a blank. Packy McCormick42:27.
  • Packy puts the sentence in quotation marks, leaves the word blank and describes the emotional tone he wants. The answer is sometimes wrong, but it gets close enough to jog his own memory.
  • Steven Johnson59:00 has a bigger version. Ask a model to rewrite a boring paragraph with the same meaning but different metaphors, to get out of your routines.
  • Johnson's wife pointed out that the same tool could be used for plagiarism, and he took that seriously.

2.4 Check facts and sources

AI is good at leads and orientation and bad at exact facts, so verify everything that reaches the page.

• Never trust it for exact quotes or facts. David Perell1:08:10.
  • Perell cites Benedict Evans: LLMs are good at things with no wrong answer, like ten ideas to fix a sentence, and bad at precise retrieval.
  • He spent two days building an article and video around a Steinbeck food quote that GPT-4 supplied. Readers wrote to say Steinbeck never wrote it in East of Eden.
• Use it for leads, then go to the real source. Atul Gawande1:00:37.
  • Gawande ran several models to find 19th century surgeons' first hand accounts of anesthesia. This cut three months of library work to two weeks.
  • The models made up the quotes, so he used the real accounts they pointed to.
  • He ended up with ten genuinely good surgical accounts.
• Keep humans at the start and the end. Jimmy Soni35:05.
  • Soni cites Balaji Srinivasan: AI is middle to middle, not end to end. Humans set the vision at the beginning and fact-check at the end.
  • Editing a grief memoir, he gave Claude a paragraph on Jewish burial customs and asked for sources, quotes and word meanings. One query replaced about ten Google searches32:45, and he verified the results.
  • John Lennox53:50 says it may hallucinate because it wants to please you. He checks its claims as one would a Google search, so it stays a servant and not a master.
• Use it for rough orientation only when you can spot the errors. David Perell1:06:30.
  • In Buenos Aires, Perell asked ChatGPT for tables of immigration by decade and origin. He did not publish the numbers as gospel. He knew enough to see that a claim of large immigration from China would be wrong.
  • Michael Pollan57:00 finds it useful for defining a term and finding where things came from. He asked how long we have known about electromagnetism, learned it is only about 200 years old, and got an analogy for panpsychism.
  • Fareed Zakaria9:35 finds it thin for deep dives. It cannot read most of the 30 best books on the French Revolution, only reviews and excerpts.
• Do your own reading instead of asking for a summary. Ezra Klein44:50.
  • Klein says AI does not know what he wants to know and would not make his connections. Spending hours with the text is the point.
  • He read a RAND report on housing construction costs for over an hour, marking it up. A summary would have given useful factoids, but the marking up is what integrates it into his mind.

2.5 Keep your voice in the final pass

The last edits are where tools smooth you into the average, so decide each change yourself.

• Keep the wrongness that carries emotion. Lulu Cheng Meservey36:24.
  • Lulu ran founders' manifestos through AI. Its suggestions often made the writing better at the expense of emotion and conviction. Emotion makes people write ungrammatical, rule-breaking lines, and that is the point.
  • Decide on each piece whether the trade is worth it.
  • Ocean Vuong20:52 says the sentence was being homogenized long before AI. Tools like Microsoft Word put squiggly lines under Shakespeare. He wants idiosyncrasy and strangeness, not the style of an invisible butler.
• Teach the tool your bad habits, not your voice. Henrik Karlsson48:23.
  • Karlsson keeps a prompt listing words he overuses and other things to watch for. He asks the model to flag them in his essay.
  • He also asks for ten versions of a phrase, just to get out of his own head.
  • He uses LLMs for about 1% of his words.
• Keep every sentence your own. Michael Jamin1:09:20.
  • Jamin says AI is not a tool, it is in lieu of writing. A joke pulled from it also will not match the voice of the rest of your piece.
  • He says the satisfaction of writing comes from turning something painful into a vase, a way to understand yourself.
  • Steven Pinker39:00 finds LLM prose plain, orderly and free of jargon, but so generic you can almost recognize it. Jack and Ryan of AJR49:01 expect it to blend artists into a nebulous middle with no point of view.
• Choose the argument and put your name on it. Fareed Zakaria12:13.
  • Zakaria says AI can write six columns for six positions, maybe better than he can. The scarce skill is deciding which one is right to present now, and staking your credibility and courage on it.
  • He notes that when asked to write, AI tends to give "on the one hand, on the other" answers because it avoids controversy.

2.6 Where writers disagree

Great writers split on whether to touch AI at all, and on how much it can judge.

• Some use no AI, others use it in the middle of the work.
  • Dean Koontz24:42 says he has no desire to use it. He adds that using it beyond fact-checking makes the work uncopyrightable.
  • Jon Yaged56:00 says US copyright law does not protect work created by AI. He calls the test for a human author something like "substantially created by a human", but says it is not a bright line.
  • Yann Martel38:14 says readers want connection and imperfection from another person, and that text is devalued once they know it is AI.
  • Robert Greene1:21:20 calls it a tool like a pen, used only to find an obscure fact, and never to compose paragraphs.
• Whether AI can critique your work honestly is contested.
  • Ezra Klein31:40 says AI will help you tweak the thing you think you should be doing, but will not tell you the whole approach is wrong.
  • Michael Dean1:21:57 says LLMs lack a consistent sense of quality. They also ignore most rules when given 25 at once.
  • Dean's own prompt tests scored about 50 percent accuracy at first. Someone had AI rewrite Paradise Lost and judge the two versions, and it preferred the slop.
• Whether to train it on your voice is unsettled.
  • Nat Eliason1:33:20 sends a 30,000 word sample of his best chapter and a ban list of phrases. He says the output is close to his style, but he still deletes and rewrites most of it.
  • Scott Galloway6:45 asked for a 1,500 word post in his own voice and got something anodyne. He explains the argument to a person and adds the voice himself in editing.
  • David Perell48:58 builds a style guide and a folder of paragraphs he finds interesting. He uses it to get feedback that pushes him to write like that.
• Brainstorming with AI is a matter of taste.
  • Andy Weir1:05:16 says AI fiction has an accent of repetition and misplaced detail, and he does not use it for brainstorming for now.
  • Sam Altman6:05 describes a student who had an idea she could not articulate. She asked for more ideas to get unblocked and produced something better than most people would alone.