The Case Against Writing with AI — Ezra Klein

David Perell

Abstract

Fifty-five minutes on opinion writing, editing and what a tool can and cannot do for either. The opening reframing sets the terms: opinion writing is not spewing your own views, it is integrating voices and thoughts that belong to other people — which makes the reporting the job and the prose the last step.

• What a writer needs from an editor is judgement, not corrections; taste is the thing that is hardest to find, measure and teach.
• Editing is selection, so what matters is not what gets said yes to but what gets said no to.
• Writing is a technology of rigor: conversation lets you glide past a hole, writing stops you at it.
• The tool is better at the parts of the work that do not need doing than at the parts that do.
• Compression loses both the information and the power to change a reader — it is not the wave, it is the erosion.
• A written argument chisels to one best answer and can over-persuade its author.

Everything below follows the conversation in order, from the nature of opinion writing through taste, editing and the economics of media, into what writing actually is as a process, and out through the case against handing it over, ending on the danger inside writing itself. For anyone building a personal brand the useful pairing is the last two sections: the same property that makes writing clarifying is what makes it quietly persuasive to the person doing it.

Chapter summaries

00:39 - Opinion writing is integrating voices that belong to other people

Opinion writing is not spewing your own views; the work is integrating voices and thoughts that are not your own.

"does that mean? Yeah. So, this is my first editor in journalism, Michael Tamasi, who's"

— 00:39

A panorama of other people's voices is richer than one person's opinion, and reporting is what supplies it. A first editor told a young writer his draft isn't even journalism, this is only you — pick up the phone. A columnist is licensed to spew and he mostly chooses not to: not that interested in what I think.

01:58 - A voice that carries weight is a responsibility rather than a hobby

A voice that carries weight is a responsibility, not a hobby, and as reach grows the same sentence changes meaning.

"amazed as I was prepping for the interview how much your voice has changed. like"

— 01:58

At scale the punches land, so a takedown becomes punching down, and the turn is toward construction over destruction. A writer stopped doing the takedowns he did as a young blogger once he got big — still honest and authentic, true about what comes into it, treated as weight rather than toned down.

04:27 - What a writer needs from an editor is judgement rather than grammar

What a writer most needs from an editor is not grammar or clarity fixes but judgement — trust that the editor knows what is good.

"there, is like an institution and a genius."

— 04:27

Before something hot goes into a big world, the value is a second person with real judgement rather than a technical pass. An accomplished writer says he barely cares whether the editor is checking grammar or argument clarity — that is not what the editor is for.

05:11 - Taste is the hardest thing to find, measure and teach

Taste — knowing what is good — is the hardest thing to find, measure and teach, and it is what separates writers and editors.

"The thing that is hard to find is taste. The thing that is hard to"

— 05:11

It is what separates writers from editors. Voice is not engineered by deciding what a piece of yours is; you chase the idea and the voice follows.

07:47 - A step up in craft marries the granular to the general

A step up in craft comes from marrying the extremely granular with the highly generalised — the thirty-thousand feet and the three inches.

"the craft is a level up, like what about it? I think the book has"

— 07:47

Concrete reporting is what makes a big abstract point land, and matching cadence to the piece is the same move at the level of sound. Several introductions got thrown out until the voice of a book was almost a march — a manifesto requires a different kind of writing than a descriptive book.

09:29 - The surest signal is watching an idea catch fire before publication

You can often feel when a piece will spark, and the surest signal is watching the idea catch fire before you publish.

"you used the word kind of hot when it came to thinking about writing a"

— 09:29

The signal arrives before publication: the idea catches among the people it is tried on, and the reaction to a draft predicts the reaction to the piece. Testing the idea in conversation is the cheap version of that.

10:33 - Nobody can describe their own taste

Nobody can describe their own taste; the people best at what they do are the least able to explain how.

"that sense of taste was honed by volume? And if that doesn't account for all"

— 10:33

Taste is intuitive, so introspection returns nothing usable and why did this spark has no available answer. A podcaster's observation carries it: his best guests are the worst at explaining their own craft.

13:00 - Editing is selection, and the noes matter more than the yeses

Editing is selection: what matters is not what you say yes to but what you say no to.

"built on an idea of what it would mean to do explanatory journalism. I definitely"

— 13:00

What matters is not what gets said yes to but what gets said no to. Editing is selection, which makes the pile of rejected things the actual output.

14:14 - An editorial innovation gets fast-followed almost immediately

An editorial innovation done well is fast-followed so quickly that the differentiation is hard to hold.

"that's always hard about having I think any kind of editorial innovation is that if"

— 14:14

The differentiation is therefore hard to hold. The advantage is in doing it first rather than in owning it. The window is the reward, and it closes whether or not the innovation was any good.

14:48 - The scale-into-money model of digital media failed

The scale-into-money model of digital media failed; more audience did not become more revenue.

"digital media in that era is we were yolked to a business model that didn't"

— 14:48

The platforms took the advertising money, so the assumption that scale converts to funding collapsed under it — an era that turned print dollars into digital dimes, where the car could not keep racing because it could not keep paying the talent it had poached. It did happen to some degree: the failure was the rate, not a lack of ideas.

16:58 - There is no single best way to build a model of reality

There is no single best way to build a model of reality; for one writer reading is the wellspring of everything else.

"your consumption changed over the years? Because I like I do just get the sense"

— 16:58

For one writer reading is the wellspring of everything else, and long undistracted reading is not input-gathering — the associational state it produces is the goal.

19:34 - Improvement comes from repetition rather than from working on voice

You do not improve by scheduling work on "voice"; you improve through repetition and constancy while chasing the idea.

"style? I don't tend to work on things in that defined way. To be honest"

— 19:34

Repetition and constancy while chasing the idea are what produce it, because voice is a residue of volume rather than a project.

20:17 - A perceived weakness can be part of a subtler strength

A perceived weakness can be part of a subtler strength, so "fixing" it can make the work worse.

"And sometimes I'll listen to an episode of my own podcast and think"

— 20:17

Fixing it can therefore make the work worse. High output means many attempts, which is what allows any single piece not to be treated preciously.

21:18 - The hit rate rises once the work narrows to what a writer is best at

Hit rate rises when you stop trying to cover everything and only do what you are best at.

"preciously. Do you feel like your hit rate needs to be higher at a big"

— 21:18

Letting pitches go by, because somebody else will cover them, concentrates the effort where the edge actually is. A writer who at a startup had to be on everything now only does what he thinks he will be best at — still led by the news cycle, but only where it connects to him emotionally.

26:11 - The move to independence can be about building rather than toppling

The paid-newsletter model did not create a journalistic middle class; only a limited number do well there.

"So there are bright spots. I'm not saying nobody makes a good living on Substack,"

— 26:11

The move to independence can be about building an institution rather than toppling one. It changes what the move is for, and therefore what counts as succeeding at it.

30:34 - The paid-newsletter model did not create a journalistic middle class

Deep work is nonlinear: the days that feel wasted are the structural and mental work that make the productive days possible.

"Yeah, 30 minutes, three hours that in three hours you're surprised at how much writing"

— 30:34

The paid-newsletter model did not create a journalistic middle class — only a limited number do well there. Journalism was funded by a bundle, with classified and local advertising cross-subsidising expensive reporting, and the survivors rebuilt a digital version of exactly that.

31:29 - Deep work is nonlinear, and the wasted days are structural

An AI writing assistant is dangerous because it will help you tweak what you are doing and can never tell you the whole thing is wrong.

"reason I am completely uh I think it is very dangerous to use chat GPT"

— 31:29

The days that feel wasted are the structural and mental work that make the productive days possible. Deep work is nonlinear, which means the output measure is lying on most individual days.

31:59 - A writing assistant can tweak but can never say the whole thing is wrong

The "not right… got it" signal in writing is almost always intellectual, not stylistic.

"earlier about good editing and like this is also true for a writer to themselves."

— 31:59

It will help you tweak what you are doing and can never tell you the whole thing is wrong. The danger is specific and structural rather than about quality.

32:39 - The not-right signal is almost always intellectual rather than stylistic

Writing works by having something at the core to pass the baton to — a report, an interview, a book — not by solving a problem through rhetoric.

"constantly finding something to pass the baton to that it's like I read this book,"

— 32:39

The not-right-yet feeling, and then the got-it, is almost always intellectual rather than stylistic. Treating it as a prose problem sends the revision to the wrong place.

34:00 - Writing works by having something at the core to pass the baton to

Writing is communication, and the only question is what you are communicating; hiding a thin core in embellishment is a terrible habit.

"to fall flat. Look, writing is communication. The question with communication is what are you"

— 34:00

A report, an interview, a book — something at the core to pass the baton to. Writing does not solve a problem through rhetoric, and the absence of a core is what rhetoric usually gets deployed to hide.

34:44 - Writing is communication, and the only question is what is being communicated

The writing is only a vehicle; the job is to go out, get the reader something, and bring it back.

"did uh Pete Holmes' podcast. Uh, You Made It Weird. He's a comedian. He's great."

— 34:44

The only question is what is being communicated. Hiding a thin core in embellishment is named as a terrible habit rather than a stylistic weakness.

36:11 - The writing is a vehicle, and the job is to bring something back

Writing is not transcribed speech; each surface is a different form, and the way you'd say it in a coffee shop does not work on the page.

"no, it's not I I don't agree with that at all actually. Cool. Because the"

— 36:11

Go out, get the reader something, and bring it back. The writing is a vehicle, which puts the effort where the material is rather than where the sentences are.

37:15 - Writing is not speech written down

Writing is a technology of rigor: conversation lets you glide past a hole without noticing, writing stops you at it.

"it's very easy for me, even without noticing I am doing it to rhetorically glide"

— 37:14

Each surface is a different form, and the way you would say something in a coffee shop does not work on a page. The two only look like the same activity.

38:29 - Writing is a technology of rigor

Writing is the endpoint of a process that is mostly research and reporting — the selling of a product you already made.

"writing is the end point of the process. having created the product, like researched the"

— 38:29

Conversation lets you glide past a hole without noticing and writing stops you at it. Writing is a technology of rigor rather than of expression, which is why it is worth doing even when nobody reads it.

39:03 - Writing is the endpoint of a process that is mostly reporting

You can beat competitors simply by outworking them — reading the primary sources everyone else skips.

"believe, right? Um I had written a commencement address for something that got cancelceled maybe"

— 39:03

Mostly research and reporting, with the writing as the selling of a product already made. That ordering explains why the writing stage feels hard when the earlier stages were skipped.

39:59 - Competitors can be beaten by outworking them

Eighty percent of a conversation happens before you walk in the room; revelation is in the preparation, not the performance.

"writing the the conversation. I always say to my team on the show that 80%"

— 39:59

Read the primary sources everybody else skips. Competitors can be beaten by outworking them, which is unfashionable and available to anybody.

41:09 - Most of a conversation happens before anyone walks in the room

The same job can be done in profoundly different ways — disappear into the background, or deliberately stand out and make people come to you.

"It's about Tom Wolf. So there's sort of two kinds of immersive reporters. And one,"

— 41:09

Eighty per cent of a conversation happens before anybody walks into the room — the revelation is in the preparation rather than in the performance. It is the same claim about reporting, applied to interviewing.

42:31 - The same job can be done in profoundly different ways

AI is better at the parts of the work you do not need to do than the parts you do, so it outsources exactly the wrong thing.

"I'm not completely against anything and I have not and not for lack of trying"

— 42:31

Disappear into the background, or deliberately stand out and make people come to you. The same job done in profoundly different ways — and a wholly distinctive voice can make its author the real subject, which is what makes that work vivid.

45:52 - The tool is better at the parts of the work that do not need doing

Knowledge is not downloaded; it is built by grappling with a text and making connections, which only happens through time spent.

"use. Yeah. Also, I wrote a piece about this a couple years back. Um it"

— 45:52

It is better at the parts of the work that do not need doing than at the parts that do, so it outsources exactly the wrong thing. And it sees what everybody else sees, where the value of a writer is what they see that others would not.

46:41 - Prompting cannot rescue a missing instruction to go and read more

Knowledge's job is to change you, and it changes you only through engagement; a summary you didn't grapple with cannot.

"you, part of what is happening when you spend seven hours reading a book is"

— 46:41

When the missing move is to go and read more books, no prompt will produce it. The tool cannot tell you that you have not gone far enough, because that is not a question about the text.

48:19 - Knowledge is built by grappling rather than downloaded

Compression loses both information and the power to change a reader; not everything should shrink to the shortest form.

"think shortcuts often operate off of a misguided metaphor of how our minds work, right?"

— 48:19

It is built by grappling with a text and making connections, which only happens through time spent. Knowledge's job is to change you, and a summary nobody grappled with cannot do that.

50:45 - Compression loses both the information and the power to change

Writing-to-think is real, but it has a trap: you can be persuaded by the formalism of your own assignment.

"You can have epiphany through writing. I mean, I know writers always love the Joan"

— 50:45

Not everything should shrink to the shortest form, because compression loses both the information and the power to change a reader. Repeated long-form contact does more than any single hit — it is not the wave, it is the erosion.

51:10 - Writing-to-think is real and it carries a trap

A written argument chisels to one best answer and can over-persuade its author, closing off doubt that should stay open.

"is. I notice that when I'm doing a podcast, I don't feel need to come"

— 51:10

Writing-to-think is real and it carries a trap: you can be persuaded by the formalism of your own assignment. A written argument chisels to one best answer and can over-persuade its author.

52:56 - Writing uncertainty without resolving it is among the hardest things

Writing uncertainty without falsely resolving it is one of the hardest things, and the opinion-column form pushes against it.

"resolved. Yeah. I think that being able I think if I was to describe things"

— 52:56

The opinion-column form pushes against it, applying pressure toward a confidence the writer may not have. A column that ends in genuine uncertainty reads as a failure of nerve, which is a cost of the format rather than of the thinking.

54:25 - The person being convinced is usually the writer

You usually aren't convincing the audience — you're convincing yourself — so guard against talking yourself prematurely out of uncertainty.

"talking yourself prematurely out of uncertainty is difficult. And recognizing as a writer, you always"

— 54:25

You are usually not convincing the audience — you are convincing yourself. Guarding against talking yourself prematurely out of uncertainty is the closing instruction, and it is a warning about the instrument rather than about the writer.

Personal Branding Lessons

A conversation about writing as reporting. The moves below are the practical half.

Ask an editor for judgement rather than for corrections

The thing a writer most needs is not grammar or clarity fixes but somebody whose sense of what is good can be trusted. That makes the useful editor a taste instrument rather than a correction service — and taste is the hardest thing to find, measure or teach. 04:27

Marry the thirty-thousand feet to the three inches

A step up in craft comes from putting the extremely granular and the highly generalised in the same piece. Either alone is ordinary; together they produce the effect people struggle to name in writing they admire. 07:47

Judge the editing by what gets said no to

Editing is selection, which means the rejected pile is the real output. What ends up on the page is decided almost entirely by what was refused, and a process with no refusals has not edited anything. 13:00

Read long and undistracted for the state it produces

Long reading is not input-gathering. The associational state of mind it produces is the actual goal, which is why the same material skimmed does not substitute — the point was never the facts on the way through. 18:03

Raise the hit rate by only doing what you are best at

Stop trying to cover everything. The hit rate is a function of selection rather than of effort, and narrowing what gets attempted raises the proportion that lands without anybody working harder. 21:18

Have something at the core to pass the baton to

A report, an interview, a book — something the writing hands off to. Writing does not solve a problem through rhetoric, and reaching for rhetoric is usually what happens when there is no core to hand off to in the first place. 32:39

Do the reading everybody else skips

Competitors can be beaten by simply outworking them — reading the primary sources most people skip. It is unglamorous and available to anybody, which is exactly why so few people treat it as a strategy. 39:03

Guard against talking yourself out of uncertainty

You are usually not convincing the audience, you are convincing yourself — and a written argument chisels to one best answer whether or not one exists. Writing genuine uncertainty without falsely resolving it is among the hardest things to do. 54:25

Questions

Each answer ends at the moment in the recording where it is given.

Why does my writing feel thin no matter how much I polish the sentences?

Because the problem is almost never stylistic. The persistent sense that a passage is "not right" is usually intellectual — a signal that the reporting or the reading behind it has not been done, not that the words need buffing. Writing works by having something solid at its core to hand off to, one source at a time; when a piece fails, it is because that core was missing, and no quality of prose rescues an empty centre. Stop editing and go get the missing thing. 31:59

What should I actually be looking for in an editor or second reader?

Not a grammar pass. A capable writer can already tell whether a sentence is clean; what cannot be self-supplied is a second person with real judgement about what is good before something exposed goes into a large, attentive world. The most important quality in a reader is that they know what is good and what is not — and the relationship that makes their notes usable is trust built over time. 04:27

Can I train my taste, or is it just something you have?

Both, in an uncomfortable mix. Taste — knowing what is good — is the hardest thing to find, measure and teach, and technique can be upskilled only by someone who already has the underlying intuition. Honing it is partly development through volume and partly learning to trust judgement you can already feel but cannot explain. Nobody can fully describe their own taste; the people best at what they do are consistently the worst at accounting for how. 05:11

How do writers actually get better over time?

Through repetition and constancy while chasing the idea — not by scheduling periods to work on "voice." A high-output writer reported no years, and no individual pieces, spent consciously sharpening his craft; the improvement was a by-product of shipping again and again. High volume also lowers the stakes of any single piece, so none gets treated too preciously. Publishing in the open on a steady cadence does more than any amount of private practice. 19:34

Should I use AI to help me write and research?

As a faster search and for some preparatory fetching, yes. For the core of the work, no — because it is better at the parts you do not need to do than the parts you do. Hand it a book to summarise and it has no idea what you were looking for and makes none of the connections you would have. It optimises the assignment you give it and can never tell you the assignment itself is wrong. 42:31

Why does reading a summary never feel like the real thing?

Because knowledge is not downloaded; it is built by grappling with a text and making connections, and that happens only through time spent. Spending seven hours with a book is seven hours with your mind on the topic, and that duration is the mechanism, not a cost. A summary you did not wrestle with cannot change you — and the whole point of knowledge is that it changes you. It is more dangerous to think you have read something than not to have read it at all. 45:52

Is it better to write short and punchy, or long?

Compression trades away both information and the power to change a reader. A shorter form does not carry the same content — fidelity is lost at every step down from book to article to post — and its capacity to move someone shrinks with it. Repeated, long-form contact does more than any single hit: the change comes from erosion, not one wave. Offered as a tendency for most people most of the time, a case for length, not a ban on brevity. 48:19

How did the money that funded serious journalism actually work?

Through a bundle. Expensive reporting almost never paid for itself; it was cross-subsidised by revenue that had nothing to do with journalism — the classified ads and local-advertising monopolies a paper happened to dominate. The money for investigations came from telling readers where the sales were. When those revenue streams unbundled, the work they quietly funded went into crisis, and the direct-subscription model that replaced them only ever answered high-frequency commentary. 27:20

How do I keep from talking myself into a conclusion I am not sure of?

Guard against being persuaded by your own draft. Writing pushes toward a single resolved conclusion, and the first person it convinces is always the writer — you can be talked into something by the formalism of your own assignment. The failure is not being wrong; it is leaving no room for the doubt that should remain. You usually are not convincing the audience of anything; you are convincing yourself, so hold the uncertainty open on purpose. 50:45

Sources

The Case Against Writing with AI — Ezra Klein

The sections above follow the episode's own order. Each timestamp is the point in the recording where the idea is discussed, so the post reads as a map of the conversation rather than a rearrangement of it.

Quotations come from the episode transcript with spoken filler ("okay", "right", "you know", "um", "actually", "kind of", "sort of") and false starts removed. No words are added: square brackets mark anything inserted for sense, and an ellipsis marks any cut. British spellings are restored, and speech-to-text errors were repaired against context only where the intended word is unambiguous. Where a line appears both in a cold-open teaser and in its place in the conversation, the timestamp cites the second — the point at which it is actually said.

Quotes are stamped with the moment they are said rather than with a speaker's name.