I added a small piece of metadata to every post on this blog this week. Each one now carries an authorship tag: human-written, ai-assisted, or ai-generated. They show up as a distinct chip at the top of the post, alongside the topic tags, in a different style so they’re hard to miss.

The implementation was an afternoon. The categorization was a few hours of staring at posts and asking myself an uncomfortable question: who actually wrote this one?

This post is about why I think that question matters, why the answer is rarely as clean as the labels suggest, and why I decided to publish my best honest answer anyway.

What the labels are trying to do

Three labels, three rough zones:

• human-written: I wrote the post. The argument, the sentences, the rhythm. AI may have helped with a small thing — a typo check, a Portuguese translation that I then rewrote — but the prose is mine.
• ai-assisted: I directed and edited, but a meaningful amount of the prose came from the model. I rewrote, cut, restructured, and pushed back. The voice is mine because I made it mine. The first draft wasn’t.
• ai-generated: the model produced most of the text and I lightly edited it. I’m not pretending I crafted every sentence. I picked the topic, set the constraints, signed off on the result.

These aren’t policy categories from a journal. They’re rough self-assessments. Two writers using the same workflow could legitimately tag their posts differently and both be telling the truth.

The retroactive part is harder than it sounds

Going through eight years of posts and assigning a tag was strange.

The 2018 posts were easy. They were written before the modern wave of generative tools. I sat in a chair, opened an editor, and typed them. human-written is unambiguous and there are twenty-one of those.

The recent ones were trickier. Take the post where I migrated this whole site from React to Astro. The migration was done in a single conversation with an AI coding agent — I described the goal, the agent wrote most of the code, I reviewed and pushed back where needed. The post about that migration was also ai-assisted, but in a different way: I directed the argument and structure, the agent helped me draft the prose, and I rewrote until it sounded like something I’d actually say.

That distinction — what kind of collaboration was it? — is the part the labels are forcing me to think clearly about. The code work was AI-assisted in a “the agent wrote most of it” sense. The writing was AI-assisted in a “the agent drafted, I shaped” sense. Same label, different texture. The tag captures direction, not detail.

I went the other way for some of the early 2026 posts. The internationalization post and the search-and-archive post were tagged ai-generated originally. Looking back, that’s not quite right either. The argument and structure were mine. What the AI did was help me say it more cleanly than I would have on a first pass. That’s ai-assisted. So I changed those tags. Two days of self-debate to move a five-character string in a YAML file.

Why I’m bothering at all

The simple version: I think readers deserve to know.

Not because AI-assisted writing is bad — I don’t think it is, and I’d lose any argument for that pretty quickly given that the post you’re reading is itself ai-assisted. I think readers deserve to know because the calibration matters. When you read something a person wrote, you adjust your trust based on what you know about that person. When you read something a model produced, you should adjust your trust based on knowing that. Without the label, you can’t do either correctly.

It also matters for me, not just for readers. Looking at the archive and seeing “this post was mostly mine, this one was a collaboration, this one was substantially the model’s work” gives me a kind of mirror. I can see the shape of how my writing has been changing. I can notice when I’m leaning on the tool more than I want to be. I can also notice when I’m under-claiming AI involvement to feel more authentic, which is its own form of dishonesty.

There’s a third reason that’s harder to phrase. The whole conversation about AI in writing has gotten a little tribal — pure-human writers on one side, full-automation enthusiasts on the other, and a quiet middle that doesn’t say much because it’s hard to defend a moving target. The tags are my small contribution to that middle. They say: most things sit somewhere on a spectrum, and pretending otherwise gets in the way of an actual conversation.

The temptation to over-claim — in both directions

There are two easy ways to lie about this. I could mark everything human-written because that sounds more impressive and hope nobody notices the AI involvement. Or I could mark too much as ai-generated because it makes the automation sound more capable than it really was.

Neither ages well. The first hides the tool; the second erases the editing, rejection, structure, and judgment that made the post work. The ethical move is the one that doesn’t flatter me in either direction. Some posts are mine, some are collaborations, some are mostly the model’s. The reader gets to know which.

What ai-assisted actually looks like

Since the middle category is doing most of the work, let me describe what it actually is in practice, for me.

A typical ai-assisted post starts with a few hours of thinking — what’s the argument, who’s the audience, what’s the one thing I want the reader to take away. That’s still my job entirely. Then I open a session with an AI coding agent that has access to the rule files I wrote about voice and structure. I describe the post, sometimes with a rough outline, sometimes just a paragraph of intent. The agent produces a first draft.

Then the work begins. I rewrite the opening because it’s too generic. I cut a section the model thought was important and that I think is filler. I add a personal anecdote the model couldn’t have known. I push back: “this paragraph is too neat, the actual situation was messier, rewrite it with more ambiguity.” Sometimes I throw away the entire draft and ask for it again with different framing. Sometimes I keep most of a paragraph because the model phrased something better than I would have.

By the end, the post has my structure, my opinions, my anecdotes, my edits, and a fair amount of the model’s phrasing. That’s ai-assisted. It’s not “AI wrote this and I rubber-stamped it” — that would be ai-generated. It’s not “I wrote every word” — that would be human-written. It’s a real collaboration, and the label tells the reader that’s what they’re getting.

What I’m not claiming

I want to be careful here: the labels are not precise, universal, or a quality judgment. They’re a fuzzy process disclosure for this blog. I’d rather publish an honest, imperfect standard than wait for a clean industry-wide one that may or may not arrive.

A small ritual

There’s a moment now, before I publish a post, where I pause and ask: honestly, which one is this? That pause is small and free and slightly uncomfortable, which is the right combination for a useful ritual.

It also keeps me from drifting. If most of my recent posts are tagged ai-assisted, that’s a fact about how I’m working right now. If they all start drifting toward ai-generated, that’s a signal I should probably pay attention to. Tags don’t just describe the past. They quietly point at the trend.

That’s the part I want. Tags don’t solve the larger problem, but they leave a small honest disclosure on each post. Not perfection. Just a record.