Writers now use AI more than ever, and trust it less than ever. A recent survey of American writers found 71% use AI in their work, yet 60% distrust AI-supported writing, including their own. Both things can be true at once, and together they make a strong case for thinking twice before using AI to write your next draft.
The essay that hit a nerve
Software engineer Paul Bakker published a short essay in August called Don’t use AI to write, and it climbed to the front page of Hacker News within hours. His argument fits in one line: writing requires thinking, and letting a chatbot do the writing means letting it do the thinking too.
What makes the essay interesting is that Bakker isn’t an AI holdout. He uses Claude, ChatGPT and Gemini all day, for coding, research, and digging through data. His complaint is narrower. When you feed bullet points into a model and get a polished document out, you skip the part where ideas get forced into shape. The draft looks good. You accept it with small tweaks. And you never find out what you would have thought if you’d had to write it yourself.
That’s the trap. AI is genuinely good at writing, which is exactly why outsourcing to it feels harmless. The output is fluent, so nobody argues with it. But fluency isn’t the product. The product is the idea underneath, and the idea is what got skipped.
What the numbers say
The discomfort isn’t just one engineer’s opinion. A Harris Poll survey of 1,175 American writers, run for Medium in August 2026, put numbers on it. Seventy-one percent now use AI in some part of their writing. Yet 60% of respondents distrust AI-assisted writing even when they produced it themselves, and 69% distrust it when someone else did.
Read that again. Most writers don’t trust work they know they used AI on, work they presumably reviewed. That’s not technophobia. That’s people telling you the review didn’t make them confident.
Readers outside the writing world are catching on too. A public test by SEO research firm DEJAN found that people correctly identify AI-written answers 64% of the time, and their accuracy climbs the longer they read. A separate consumer survey by Fractl found the share of people who say heavy AI use in a brand’s content lowers their trust has nearly doubled in a year, from 20% to 39%. Sounding human isn’t a stylistic preference anymore. It’s becoming a trust requirement.
For anyone using AI to write anything that carries their name, the direction is clear. Usage keeps climbing, and so does the penalty for sounding generated.
Writing is the thinking
Here’s the core of Bakker’s case, and it deserves a fair hearing. Ask yourself what actually happens when you write something hard. You start with a vague sense of what you want to say. Then sentence by sentence, the vagueness collapses. Words refuse to fit. The argument has gaps you hadn’t noticed. Finding the right phrasing is how you find out your idea was half-formed.
Skip that process and you skip the discovery. A generated document can’t reveal what you think, because you didn’t think it. You’ll notice this the first time someone asks you a question about your own AI-written proposal and you can’t answer without rereading it.
It also explains why using AI to write leaves so many people vaguely unsatisfied with the result. The document exists, yet its author has no deeper grasp of the subject than before starting. Fluent output hides an absent process.
Productivity obsessed teams get this backwards, Bakker argues. Output volume measures nothing here. A three-page strategy document you actually thought through beats a sixty-page generated one that hedges every point, because only one of them contains a decision. Programming has the same trap, he points out: nobody counts lines of code as progress, but people count pages.
Where AI genuinely helps writers
Nobody’s saying banish the tools. The useful distinction is between thinking work and support work.
Before writing, AI shines. Sifting through research, structuring messy notes, finding patterns in data, summarizing what competitors have said. All of that speeds up preparation without touching your argument. During this phase the AI is a research assistant, and the ideas are still yours.
After writing, AI works as a reader. Bakker’s own method: don’t ask it to fix your draft, ask it questions about your draft. What questions would a skeptical reader have? What’s the single most important takeaway here? Where does the argument get shaky? The answers send you back into thinking mode, which is exactly where you want to be. Compare that with our guide to ChatGPT custom instructions, where the goal is making the tool respond the way you need rather than think for you.
The middle part, the actual writing, is where he draws the line. That’s the part that’s slow because it’s supposed to be slow.
A workflow that keeps you in charge
If you want to try this without going full purist, here’s the shape of it.
- Research with AI freely. Gather sources, compress notes, map the territory.
- Close the chat. Write the first draft yourself, even if it’s clumsy. Especially if it’s clumsy.
- Reopen the chat as a critic, not a co-author. Ask what’s unclear, what’s missing, what a reader would push back on.
- Decide what to change yourself. Take the feedback into your own edits.
Think of it as using AI to write around your draft instead of the draft itself. This mirrors the rule we keep coming back to in pieces like our guide to using AI at work safely: the tool assists, but the judgment and the accountability stay with you. It also matters more as detection gets easier. The AI job application doom loop showed what happens when everyone automates the same step; sameness gets punished, and the people who sound like themselves cut through.
The honest tradeoff
Let’s not pretend this is free. Writing your own drafts takes longer. Sometimes a generated first draft is fine, like routine meeting notes or a boilerplate update where the ideas were never the point. Bakker’s rule targets work where the thinking matters: strategy docs, essays, important emails, anything with your name and your judgment attached.
For that work, the math has flipped. Using AI to write it saves you an hour and quietly costs you the one thing the document was supposed to produce, which is a thought you actually own. The survey data suggests people sense this even when they keep using the tools. Sixty percent distrust their own output. Your readers probably do too.
So try the split for a week. AI for research, AI for critique, and the drafting keeps your hands on the keyboard. Worst case, you lose a little speed. Best case, the next thing you publish says something only you could have said.
