AI Didn't Write This. We All Did.

Every article on this site carries a small disclosure: AI-assisted content, created under human editorial oversight. It deserves more than a footnote.

"AI-assisted" is currently one of the most abused phrases in publishing. It covers everything from a writer who used ChatGPT to fix a comma, to a content farm that hit publish on raw model output. None of those things are the same. So let's open the hood — using a real post from this site as the example.

The Post We're Talking About

A few hours ago, this site published Opinions About Opinions: Why We Argue Framing, Not Facts — a piece about how both sides of the 2024 crime debate cited real data and still told completely opposite stories. That post was AI-assisted. Here's what that actually meant.

Step One: The Brief

The process started with a brief written by a human — not a single sentence, but a genuine editorial directive: what question does this post answer, who is it for, what angle hasn't been taken, what's the core argument.

For the crime data piece, the brief named the two datasets at the center of the dispute (the FBI UCR and the NCVS) and set the argument's direction: this is a data access problem, not a media bias problem. That framing — why the problem exists — was entirely human. The AI didn't generate that thesis. It was handed one.

This is the part that gets skipped in conversations about AI writing. The brief is the intellectual work. A vague prompt produces a vague post. The model is a capable executor. The director's chair is still occupied by a person.

What the AI Did Well

It found and synthesized primary sources fast. The crime data post links to the FBI's official Reported Crimes release, the BJS NCVS report, the Council on Criminal Justice's divergence analysis, and a CBS News fact-check from the 2024 campaign. Locating those sources, extracting the key figures, and weaving them into a coherent narrative took minutes. A human researcher would have spent hours on the same task.

It built a clean structure. The post's progression — hook, datasets explained side by side, framing exploit named and dissected, structural fix proposed — followed logically from the brief. The AI didn't just dump information. It sequenced it.

It wrote with appropriate confidence. One of Claude's known failure modes is excessive hedging on factual claims. The brief explicitly directed confident, declarative prose — and the model followed that instruction.

What Needed Human Hands

Truthfully, it just needed them to click submit. That won't always be the case. It's even possible I missed something and the post needs editing. But after the first read through, I felt comfortable pressing publish.

The Limits Are Structural

The model has no stake in the argument. It can execute a thesis it's given, but it can't feel the difference between a claim that's technically defensible and one that's actually worth making. It can write a conclusion, but it can't decide whether that conclusion earns the evidence that preceded it.

These aren't bugs waiting for a patch — they're structural properties of what a language model is: a sophisticated pattern-matcher trained to produce text that looks like good text. The judgment about whether it is good text lives outside the model.

As freelancers in Nieman Lab's 2026 AI journalism survey put it: the writers who find AI most useful have offloaded research synthesis and first-draft structure — freeing themselves to focus on analysis, editorial judgment, and narrative decisions. That's not AI replacing writers. That's a tool changing what writers spend their time on.

The Two Bad Takes

"AI is going to replace writers." This treats language generation as the core of writing. It isn't. Generation is the mechanical part. The actual work is deciding what to say, to whom, why it matters, and whether you got it right. Models can assist with all of that. They can't do any of it independently, because doing it requires having something at stake.

"AI output is garbage." Raw model output given a vague prompt is often generic. Raw model output given a precise brief and edited by someone who knows what they're doing can be genuinely good — faster and more thoroughly sourced than an equivalent solo draft. Dismissing the whole category because the worst examples are bad is like dismissing word processors because you've seen bad writing.

The reality: AI is a capable junior collaborator that works at inhuman speed and has no ego about changes. It also has no opinions, no judgment, and no ability to know what it doesn't know. You get leverage. You don't get a replacement.

What This Means for Reading This Site

Every post here follows the same process: a human writes the brief, the model drafts and researches, a human edits and rewrites what needs rewriting, and publishes what they're willing to put their name on. The brief is displayed publicly alongside each article.

The goal isn't to produce AI content. It's to produce good content — and to be honest about what that looks like. Because the alternative — AI-generated text dressed up as fully human writing, or disclosure that doesn't mean anything — is worse for everyone trying to figure out what to trust.

You deserve to know how this was made. Now you do.