Who Did the Thinking?

Martin Fowler published a post last week called LLM Voice . It describes the uneasy feeling of reading prose an LLM probably wrote, and it advises writers not to let a model write their drafts. The Hacker News crowd  mostly agrees.

I’ve read some version of this post many times now, and the evidence is always the same: the author can tell. AI prose is bland and uncanny, and they can spot it. I think it misses the point, and I don’t want to write another argument about whether AI writing is bad. There are plenty of those. I want to show what the good version looks like and how I use AI to enhance my writing.

Every toupee looks fake

Every toupee looks fake when you only notice the bad ones. The good ones walk past you on the street, and you never notice them.

Fowler’s post makes this point for me. In a footnote he describes one piece where he knows the first draft was written entirely by an LLM, and he couldn’t detect any LLM voice in the final version. It “did, however, need a professional editor at a major periodical publication to clean it up.” He only knows about that one because someone told him. He also writes that he’s “suspicious of my own ability to accurately recognize LLM voice,” since some phrasing that grates on him is just as likely to be generational, which is to say, a matter of taste.

So the one carefully edited case he knows of passed. Everything else he’s judging is the stuff that got caught, which is the careless stuff. That tells you a lot about careless writing and nothing about the tool.

Careless writing isn’t new, either. Drop any of these critics into a high school English class. The essays open with a dictionary definition and close with “In conclusion.” Nobody calls that good writing. But nobody tells those students to stop writing, either, because they’re beginners, and being bad at it is how everyone starts. A lot of the AI writing people mock is the same thing: people early in learning a new way to write. Telling them to stop because the early work is bad just means they will never have the opportunity to improve.

Two arguments

Fowler makes two arguments, and the footnote only answers the first. The first is a reader’s argument: this prose puts me off. The second is a writer’s argument, and on its face it’s the better one.

“Often it’s only by writing something down,” he writes, “that I can really get my thoughts on a topic into coherence.” Writing is how he figures out what he thinks. If you hand the writing to a model, you hand off the thinking, and you never come to understand your own subject. That’s why, imagining a day when LLM prose is indistinguishable from ours, he writes: “In many ways, that’s more worrying.” For now the voice works as a warning label. If it disappears, you’d be reading something that looks like a person talking to you with no person’s thinking behind it.

It sounds right. I think it’s wrong, for a few reasons.

Drafting is thinking out loud

Fowler says he speaks his words in his head before he writes them down. I do the same thing, but out loud on my morning walks. Most mornings I walk my dogs and talk into my phone for half an hour. Sometimes I’m working through an idea, and sometimes I’m just rambling. It’s my inner monologue, dictated. Then I hand the transcript to a model and we argue about it. It asks questions, finds the counterargument, and points at the source I haven’t read. I push back and redirect. My drafts are generated from these types of conversations.

So when Fowler tells writers not to let a model write the draft, in my process that means: don’t think out loud with a partner. For me, LLMs are the best thinking tool I’ve ever had. The struggle he values doesn’t disappear. It moves into the conversation, and then into the edit. I’ve written about the workflow before, in AI-Native Writing.

The edit is another piece critics picture wrong. I don’t hand-type the prose. I read the draft, say what’s wrong with it, and the agent makes the change. Often I reject the fix, and the next one, until one lands. Sometimes the model proposes a better sentence than mine and I take it. Every decision about what gets written is mine. The judgment never lived in my keystrokes. It lives in reading a sentence, knowing it’s wrong, and knowing why. That’s an editor’s job, and it’s the job Fowler accepts in his own footnote.

The model doesn’t do any of this on its own. Left alone, it sits there and produces nothing. The conversation is what makes something new, the same way a DJ makes something new out of records that already exist. Nobody confuses a good mashup with Yo-Yo Ma, and nobody should. They’re different crafts, and both have real value.

The exception is the rule

Fowler makes exceptions too. His colleague Jim Highsmith writes with an LLM because, Fowler says, it’s essential to “overcome the bane of Parkinson’s disease,” and Fowler accepts the voice to keep reading him. He also mentions colleagues who use LLMs to translate into English.

I think those exceptions are right, and I don’t think they stay small. Once you grant them, the rule is no longer “authors should fill the prose themselves.” It’s a cost-benefit judgment, and everyone’s costs are different. Someone with dyslexia has one set of costs. So does someone who has good ideas and writes badly. So does someone who writes fine but slowly, and has a job. I’m in that last group. Writing by hand took me so long that I never had a blog. I only started this one once LLMs made writing time-efficient for me.

I’ve watched longtime programmers react to AI coding the way some longtime writers are reacting now. When something gets easier, the people who learned it the hard way tend to defend the old way of doing it. I think Fowler is doing a version of that. He’s clearly working it out in public, and I respect that, but his advice comes down to writing the way he learned to write.

Judge the result

My bolder answer is that it doesn’t matter who did the thinking. It matters whether the thinking is good.

Fowler’s argument puts the value in the author’s struggle. That’s a writer’s view of writing. A reader gets nothing from my struggle. They get the idea, and an idea is just as true or useful wherever it started. Deng Xiaoping’s line applies: “It doesn’t matter whether a cat is black or white, as long as it catches mice.” You don’t read a newspaper to admire the reporter’s process. You read it to find out what happened.

We already accept this almost everywhere else. Nobody asks whether a CEO wrote her own shareholder letter, or whether the professor whose name is on a paper ran every experiment. Speeches have speechwriters. Memoirs have ghostwriters. Code has always had bugs, written by humans, and nobody concluded that humans should stop writing code.

My name on a post means I’ve read every sentence, I believe it, and I’ll defend it. That’s the test I’d give anyone writing with AI: can you defend every sentence without the model in the room? If you can’t, Fowler is right about you.

Where Fowler is right

He’s right that unedited output doesn’t belong in front of readers. I’m in a group text with two friends. One of them is very anti-AI. I think his opinions come from a lack of firsthand experience and a lot of secondhand information. The other loves AI and uses it haphazardly, pasting unedited chat logs straight into the thread. To the first one, that’s AI psychosis.

I don’t mind the chat logs. My other friend is open about using AI, and I want to know what he’s trying to say, so I skim them and ask what he means. Among friends, that mostly works, though even there it can cost you half your audience. For a large audience it doesn’t, and it never did. The fix hasn’t changed either. Care about the result. As long as the care is there, it shouldn’t matter what tools you use.

What the good version looks like

Here’s how I do it.

Talk it out. Record yourself thinking through the idea for as long as it takes. Ramble, follow tangents, contradict yourself. Nobody will read this.

Argue with it. Give the transcript to a model and ask it what you’re actually saying, where it’s weak, and what the other side would say. Let it push back. Some of the best moves in this post came from that conversation, and some came from me rejecting what it offered.

Edit in conversation. Read the draft as a reader would. Say what’s wrong in short directives: cut this, this claim is too strong, this isn’t what I meant. Keep rejecting fixes until one is right. Edit so the idea comes across clearly, in your own voice.

Write the tells down. This is the step I wish critics talked about. Every time you notice a habit in the drafts that you hate, write a rule for it, and give the rule to an automated critic, a second model that reads every draft before you do.

From tell to rule

Here’s a recent example. My Polymath Profiles are long portraits of interesting people, researched and drafted by an agent. I make them for an audience of one: me. I pick people I’m curious about, and since nobody is writing the profile I want to read, I generate it.

The drafts had picked up a habit of narrating every small disagreement between sources. The Stafford Beer profile said he married his second wife “in 1968 or 1969 (Wikipedia and Leonard’s memorial differ)” and that he gave up “the London (or Surrey) house.” Every aside was accurate and honest, and together they made the piece unbearable to read.

I want these profiles to be about how someone lived, not an audit of the record. So I wrote a rule: on a minor detail, go with the best-supported version or round it, and never narrate the disagreement. The critic now flags those asides as something to fix. The profile says “in the late 1960s” and “the big house,” and it reads better.

Another one came from translating this blog into Traditional Chinese. A native speaker kept rejecting translations that were correct, calling them cringe. Each rejection became a specific rule about what made it cringe, and those rules now guide every translation before a person sees it.

None of these rules came from the model. I had to notice each tell first. That also answers the best objection to working this way: that choosing from a model’s suggestions slowly pulls your taste toward its taste. My rules are my taste, written down where the model can’t wear it away.

How this post was made

This post started as a half-hour voice memo on a dog walk, the morning I read Fowler’s post. Then came a long chat with Claude, where I made the case and it pushed back, and then an editing session like the one I described above. The raw transcript is rough. At one point I say, “Maybe I’m not saying this super clearly. But that’s also my point.” The transcription renders Deng Xiaoping as “dung Shopping.” Getting from that to this is the work I’m talking about.

Many posts on this blog had far less of me in them. The Polymath Profiles and the Dead Ends series are mostly generated, with light editing. They’re experiments: I’m building pipelines that can write with very little input from me, and I’m letting them be bad first so I can make them good. This is a blog. Nobody pays me to write it, and I’m lucky if anyone reads it, which makes it a good place to experiment. Some early entries are worse than I’d like. The Stafford Beer rule exists because I published a draft that had the problem, read it, and fixed the pipeline instead of just the post. That’s how getting good at anything works. You’re bad at it first, you notice, and you correct.

Fowler asks writers on his site to disclose how they used AI, partly so readers can learn which uses are worth copying. That’s a good rule, and this is my disclosure.

Be additive

A lot of the people I talk to about AI focus on minimizing the downside. That’s worth doing, but it’s not the only thing worth doing. You can avoid the downside and still leave room for the upside, and the upside here is large. Most people aren’t good communicators. It’s a skill that takes years, and most people don’t have the time to spend on it. Now there are tools that help us work out what we think, find what we’re missing, and say it clearly. That’s a good thing.

I’m glad Fowler wrote his post. It made me work out why I disagree with it so strongly. So here’s my ask to anyone who’s sure they can spot AI writing: if you can see the tells that clearly, write them down. Turn each one into a rule someone else can use. Spotting it is the easy part.