These posts were drafted by AI. Here’s why they’re still mine.

Every post on this blog was drafted by AI. Specifically, by Claude, working from briefs I wrote: the story arc I wanted, the receipts I wanted anchored, my voice guide, and a list of things to avoid. Every draft came back to me, and I read it, corrected it, cut it, and decided whether it shipped. The sentences you’re reading were generated by a machine. I’m Julie Kaiser – a science writer for regulated biotech and pharma since 2011 – and I want that on the record before you keep reading.

This page isn’t an apology, because I’m not sorry. It’s a disclosure and an argument: that these posts are still mine, and that I can name exactly what would make them not mine. Both halves matter, so here are both.

What the printing press actually changed

Before the printing press, publishing knowledge required wealth. If you couldn’t afford scribes, your knowledge stayed in your head, or in one letter, or in one book you copied out over years. The press meant that people with something to say could reach readers without being rich. What changed wasn’t the knowledge or the thinking. It was who could share it.

The word processor did a smaller version of the same thing: it made long, heavily revised writing possible for people who couldn’t retype a manuscript five times. Nobody now argues that a typewritten essay is more authentic than a word-processed one. Both have the same relationship to the thought underneath.

I think about AI drafting the same way. These tools didn’t make new thinkers. They lowered the friction between having something to say and being able to say it at scale. I have more tested, useful knowledge than I have drafting hours, and that gap used to decide what got shared. It doesn’t anymore. AI is my press.

So the question I’d ask you to hold while you read this blog: if the information is useful, and it’s mine, and I stand behind it – does it matter that AI drafted the sentences? I think the honest answer depends entirely on what went in before the drafting started. Which brings me to the distinction this whole page exists for.

The two classes of AI-drafted content

Not all AI-drafted content is the same thing, and the difference is the whole argument.

Class A: someone with no particular expertise, no lived stories, and no strong opinions asks an AI to write about a topic. The output is fluent, because fluency is what these models sell. It’s also a compilation of the most plausible things anyone could say on that topic. It’s generic because the input was generic. No amount of polishing rescues it – the problem isn’t the AI, it’s that there was no person underneath the request.

Class B: someone with real expertise, lived receipts, captured stories, and specific opinions writes a detailed brief. They tell the AI what the story is, what to anchor, what to avoid, whose voice to echo. The AI drafts. The person reads, corrects, cuts, and sends it back until it’s right. The output is a compression of that person’s actual thinking into prose they didn’t have time to write by hand.

These two classes look identical to a detector. They’re not the same thing to a reader who wants something true.

What actually goes into these posts

To make Class B concrete, here’s what was inside one post before the AI wrote a sentence of it. The post about AI hallucinations contains my first documented AI hallucination, from June 2023 – a real incident, with the actual quote from my chat archive. It contains the 88 fact-checking challenges I counted across three years of my ChatGPT conversations – ChatGPT alone, counted by searching my chat archive for a handful of the phrases I use when I push back on an answer. Perplexity and Gemini aren’t in that number. 88 is a floor, not a total.

Every specific in that post existed in my head (or my notes) before it existed in a draft. The AI wrote sentences that put my specifics into an order you can follow. That’s the division of labor, on the record.

What I hand off, and what I don’t

Expression work goes to the AI. Judgment work stays with me. Sentence-level choices, structural options, first-draft compression – AI. What’s true, what matters to you, what to leave out, whether the analogy holds – me.

And review isn’t a rubber stamp, so here’s what it looks like in practice. When I catch a sentence that lands wrong, I rewrite it. When I catch a fact that’s slightly off, I correct it. When I catch a framing that isn’t mine, I cut it. When a draft needs something only I know, it stops until I’ve supplied it. The editorial gate this runs through is part of my working setup, and it’s the reason this arrangement doesn’t scare me. If I ever hand off the judgment, I’m not using AI badly – I’m not doing my job.

What an AI detector would say about this blog

It would flag every post, and it would be right – in the narrow sense the tool measures. It would be wrong about everything that matters: whether the posts are true, whether the receipts are mine, whether I stand behind each claim when you push back. Detectors measure surface. They don’t measure honesty.

I’ve written about what detectors get wrong about human writers. This page is the other half of that honesty: what they’d get right about me, told to you before any tool does.

What would make a post not mine

Here’s the test I run, and you’re welcome to hold me to it.

If I read a paragraph and don’t recognize the story, the frame, or the receipts as mine – that’s not my post, and I fix it or cut it. If a post contains a claim I couldn’t defend when a reader emails me pushing back – not my post. If I let a draft ship because it sounded fine, without checking whether the fine-sounding thing was true – that’s not my post. That’s slop, and I refuse to put slop out under my name.

Whether working this way is cheating is a real question, and I took it seriously in Is using AI cheating?. The short version: hand off expression and you’ve saved time. Hand off judgment and you’ve cheated – mostly yourself.

The rule underneath

A byline is a promise about judgment, not about typing. Every specific here is mine, every claim survived my checking, and every sentence passed one test before shipping: would I say this out loud? A machine drafted them. I signed them.

Will readers still care about this distinction in five years? I don’t know. None of us do. But you should get to care or not care with the facts in hand – and now you have them.