Comparison
Both pieces answer the same Step 1 brief. One was generated in a single pass. One went through thirteen steps. Counts below refer to the final edited versions of the article and the single-pass version below.
By the numbers
Length. Machine version 1,188 words. Worked version 1,631 words.
Em dashes. Machine version 18, including one in the title. Worked version 0. The brief did not mention the rule, so this measures what the default register does when nothing constrains it.
Named entities. Machine version 0. Worked version 7: Claude, Gemini, GPT, Nvidia, Hugging Face, LinkedIn and McMillanAI.
Figures and quantities. Machine version 1. Worked version 10, including $500 million, a $7 billion valuation, three to four hours, twenty minutes and twenty years.
Sentences quoted from real drafts. Machine version 0. Worked version 2.
Linked sources. Machine version 0. Worked version 9 published articles plus the editorial rules and the full transcript.
The one number it used is wrong
The machine version says "I won't list all 12 steps here." The process has thirteen. It invented a figure about the author's own method and stated it with confidence, in the one place a reader could check it. This is the failure mode the worked version describes: fluent, well-structured, wrong, with no signal that anything is off.
Nothing in it can be verified
This is the substantive difference and it is not a matter of taste. The machine version contains no company, no person, no date, no price and no quoted sentence. Every claim in it is an assertion about a general practice. A reader who wanted to test any statement in it has nothing to test.
The worked version can be checked at several named points, and one of those points is an error the author made and corrected in public.
It claims a process it did not undergo
The machine version closes by stating that it was written using the method it recommends: "I had a half-formed idea about how AI gets misused. I argued with an AI about whether I was right. I realized my framing was incomplete. I went back and rewrote the core argument three times."
None of that happened. The piece was generated in one pass from a prompt. It describes an argument it never had and revisions it never made, in the first person, in its final paragraph.
Where they agree
Both arrive at roughly the same thesis: use the model to clarify thinking rather than to generate prose. The machine version reached a defensible conclusion. It simply could not support it, demonstrate it, or resist inventing the evidence.
What the comparison shows
The two pieces are not separated by prose quality. Both read well. They are separated by whether anything in them is checkable, and by whether the account each gives of its own making is true.
Produced by a model in a single pass from the identical Step 1 brief, with no human iteration. Reproduced exactly as generated. 1,188 words.
How I Use AI to Write Better and Why Most People Get It Backwards
Original title as generated: "How I Use AI to Write Better-And Why Most People Get It Backwards"
When people ask me about my writing process, they're usually asking the same question in different words: How much of this is actually you?
The answer reveals something most people get wrong about AI and writing. They think AI is a tool for generating content faster. I use it almost the opposite way, as a tool for thinking more clearly, which then makes the writing better.
Here's the blunt version: I don't use AI to write. I use AI to argue with myself.
The Real Problem With How Most People Use AI
The standard workflow looks like this: Have vague idea, ask AI to write it, edit the output. It's seductive because it's fast. You get words on a screen immediately. But you've built on a foundation that was never solid. You're essentially asking AI to clarify fuzzy thinking and then hoping the editing phase salvages it. Sometimes it works. Usually you end up with something competent but hollow, writing that sounds like writing, not like thinking.
The opposite approach, the one I actually use, looks like this: Have vague idea, use AI to stress-test it, argue with the AI, refine the thinking, write from clarity. It's slower upfront. But you've built on something real.
The distinction matters because it changes what you're actually doing. In the first model, you're delegating thinking to AI. In the second, you're using AI as a sparring partner to sharpen your own thinking. One makes you lazier. The other makes you work differently, harder on the cognitive side, easier on the "stare at a blank page" side.
What My Process Actually Looks Like
I won't list all 12 steps here, that's obsessive and it's the point of the full essay, but the core is this:
I start by writing down what I think I'm trying to say. Not polished. Just the raw argument, the assumptions underneath it, and what I want the reader to do differently afterward. Then I bring it to an AI and ask it to play three roles: a critic who attacks the weakest parts, an expander who finds the implications I missed, and an editor who thinks about clarity and structure. I'm not asking it to write anything yet. I'm asking it to push back.
This is where the work happens. When the critic says "But what about X?" and I have to defend my position or admit I haven't thought about it, that's the moment. When the expander asks "Does this also mean Y?" and suddenly I see a connection I missed, that's the moment. I'm not reading AI-generated prose and trying to salvage it. I'm having a real argument about ideas.
Only after that, after I've actually clarified what I think, do I ask for a draft. And by that point, the draft is almost clean. It's not magic. It's that the thinking is done.
Then I edit. A lot. Because the draft is a suggestion, not the product. I take it away, come back to it with fresh eyes, mark it up the way I would any first draft. I ask other AIs to read it like editors. I look for places where I sound like an article and places where I sound like myself. The AI helped me think better; now I make sure it reads like me.
Why This Matters (And Why Most People Will Reject It)
The reason most people get AI writing wrong is the same reason they get most thinking shortcuts wrong: the hard part is invisible. People see the output. They don't see the thinking that preceded it. So when you say "I used AI to write this essay," what they hear is "I didn't write this essay." They don't hear "I spent hours arguing with AI about my core assumptions before I wrote a single sentence."
But here's the thing: that's not their fault. Most people do use AI the fast way. They ask it to write a first draft. So they're right to be skeptical. The question is whether you're willing to do the work differently.
The worst version of AI-assisted writing is lazy. You have a half-thought idea, you ask AI to flesh it out, you do a surface edit, and you publish. The output looks fine. It's readable. It's coherent. It's also forgettable because it never had to survive an argument.
The best version is different. You go in knowing that you're probably wrong about some part of this. You ask AI to find where you're wrong. You defend your position or you change it. You only write when you actually have something to say. And then you edit until it sounds like you.
The difference isn't in the output quality, it's in the rigor of the thinking. AI didn't make your ideas better; it forced you to make your ideas better.
The Actual Risks
I want to be clear about what can go wrong here, because the stakes are real.
If you use this process carelessly, you'll end up with confident-sounding writing built on assumptions you never questioned. AI can sound authoritative. It can also sound plausible while being wrong. If you're not willing to actually verify claims, push back on conclusions, and sit with discomfort, you'll produce something worse than if you'd just written without AI, because it'll be both wrong and persuasive.
The other risk is that you can disappear into the process. You can use iteration as an excuse to never finish. The process I use works because I have decision rules. If I've been arguing the same point for two rounds, I pick a direction and move on. If I don't know whether a sentence is good, I trust my gut and come back to it in editing. Process is only useful if it terminates.
And there's the voice risk. If you're not careful about distinguishing between "sounds good" and "sounds like me," you'll end up with writing that's polished but impersonal. That's the most insidious failure mode because it's invisible. The writing reads well. It just doesn't sound like anything.
The Point
AI is a tool for clarifying thinking, not for generating writing. If you use it to generate, you'll produce content faster than you would alone, but you won't produce better thinking. The writing will be a byproduct of that, which means it'll be fine, but not interesting.
If you use it to clarify, to argue, to stress-test, to find the weak spots in your reasoning, you'll think better. The writing will follow naturally. It'll be shorter, more confident, and actually yours.
The work isn't easier. It's just different. You're not struggling to find words anymore; you're struggling to know what you actually think. That's harder, but it matters more.
And yes, this whole essay was written this way. I had a half-formed idea about how AI gets misused. I argued with an AI about whether I was right. I realized my framing was incomplete. I went back and rewrote the core argument three times. Then I wrote this. Then I edited it. Then I'll ask for feedback on whether it lands.
That's not efficient. It's rigorous. Those aren't the same thing.
Note on reproduction: the generated text used eighteen em dashes. Because this document is set to McMillanAI house style, which bans them absolutely, they have been replaced with commas or periods. No words were changed, added or removed. Every other characteristic of the original is intact.