I Let AI Write the Last Two Posts. Here's What I Actually Did.

Before I get to the next stop on this list, a confession: the last two posts on this blog — the one about the civil engineer and the cement, and the one about Turnitin and the classroom — were both written by AI. I approved them. Total hands-on time was a few minutes each.

I'm telling you this not to undercut them, but because it's the cleanest piece of evidence I have for the argument I've been making the whole series.

What I actually did

Here's the honest breakdown. I had Claude Code's "Marketing Author" skill loaded — a tool tuned to write in a particular voice. But before any of that, I had the idea already solid in my head: the cement anecdote, the orchestration thesis, the argument about what a Bachelor's program should and shouldn't ban. I wrote a prompt that told my own story, roughly, in my own words — closer to a voice memo than a manuscript.

What I didn't have to worry about was syntax, spelling, prose rhythm, section structure, or where to find a decent stock photo of a construction site. That used to be the craft. That used to be the two hours between "I know what I want to say" and "this is publishable." AI did that part.

My actual editorial work was small and specific: I picked a title and an excerpt from a short list of options. I asked for one structural change — turn a bulleted teaser section into something that reads as a continuing story instead of a list. And I caught two word choices that wouldn't land for this audience — "dorm room" isn't how Greek students talk about where they live, and "freshman" is a very American way to say "first-year student" — and had them fixed. Then I approved it, and it went live.

That's the whole process. It's not a special case. It's how a large and fast-growing share of the internet's writing gets made right now — a person with a real idea, an AI handling the craft, a human doing the final read and the final call. The reader on the other end has no way to tell, and — this is the part worth sitting with — they don't care. They're not checking whether the piece was typed by hand. They're checking whether it said something true and worth their two minutes.

The number that didn't matter

Out of curiosity, I ran both of those posts through a free AI detector, the same category of tool that powers a lot of university plagiarism checks. The cement post came back 0% — "human written." The classroom post came back 6.4% — also, per the tool, "human written." Two posts I know for a fact were AI-drafted, both cleared as human by the exact kind of software a professor might use to fail a student.

I'm not telling you this to make fun of the tool. Detectors are guessing at a statistical fingerprint, and that fingerprint is noisy — plenty of very human, very formal writing sets them off, and plenty of AI writing slips past them, especially once a person has actually edited it, which is exactly what happened here. The point isn't that the tool is broken. The point is that the number it produces was never the thing that mattered about either post. What mattered was whether the cement anecdote landed, whether the argument about orchestration held up, whether a reader walked away thinking something they hadn't thought before.

So why do we still grade the wrong number?

Here's the question this raises, aimed squarely back at the classroom conversation from the last post: if a percentage from a detector isn't the thing that decides whether a blog post is worth reading, why is it treated as the thing that decides whether a piece of coursework is worth crediting?

I don't think the honest answer is "it shouldn't be checked at all." A university isn't shipping content to readers who'll judge it on its own merits — it's trying to certify that a specific person learned a specific skill, and that's a genuinely different job than mine on this blog. Nobody needs a certificate proving I personally typed every sentence here. A student's degree is supposed to mean something closer to that.

But it's worth pressing on whether the current form of that certification — a single percentage, guessed at by a tool that can't tell my two AI-drafted posts from something a person wrote alone in a library at 2 a.m. — is actually measuring the skill anyone should care about. If the argument from the last post holds, the skill worth certifying isn't "produced zero percent on a detector." It's judgment: did this person ask a good question, catch what came back wrong, make it better, and know why. That's harder to grade than a number. It's also the only version of grading that would have caught what actually happened on this blog this week.

Next stop, for real this time: why so many companies are chasing one AI-fluent senior instead of training three juniors, what's worth handing a curious teenager long before any of this shows up in a syllabus, and a line item a lot of finance departments would rather not say out loud — what a company is spending on tokens and subscriptions next to what it used to spend on people.