AI Newsrooms: One Article Is Easy, the Operation Is Work

Last week a quote came back that didn't sound right. It was attributed to a real person, on a topic we'd covered before, and it was close enough to pass. It wasn't exact. The words had drifted. The AI had been working through one long session, moving from story to story, and by the time it reached for that quote it was remembering the source instead of reading it. You played this as a kid. It's called telephone.

We caught it, and not by luck. The draft was checked against the verified quote in our research file, the two didn't match, and we traced the problem to the long session. So we changed the process: every new piece now starts by reloading its sources, even in the middle of a session. Problem found, process fixed, move on.

That small repair says more about what we've built than any description of the technology could.

Anybody can get one article

Anybody can get an AI to write an article. That stopped being impressive about five minutes after it became possible. The hard part is the operation: consistent, verified local coverage across several publications, every day, without the quality sliding when nobody is looking.

That doesn't come from the model. It comes from everything around it: what we've already published, which facts have been verified, what each publication sounds like, and what we learned the last time something broke.

Rebuilding the ship at sea

The philosopher Otto Neurath had an image for this. We are like sailors, he wrote, who have to rebuild their ship on the open sea, without ever being able to take it into dry dock and start over from the best parts.

That's the job. You don't design the system and ship it. You build it while it runs. Every session turns up a gap: a piece that goes flat when it has to explain a legal point, a research file that should have existed before the draft started, a quote that drifted because the session ran long. You fix the gap, update the instructions and run it again. The next session is better because the last one broke something.

Brian Alvey, the chief technology officer of WordPress VIP, has been building publishing systems since he made the first TV Guide website in 1995. He made a version of this argument on a podcast this week. His point: if you can write down the decisions your business makes, the voice, the quality checks, the production logic, you've built something an AI can work inside. Not a replacement for people. A structure people designed.

He's right, and I'd add the part you only learn by doing it: the structure is never finished. Neurath's sailors don't get to stop replacing planks.

What the machine can't do

The question most people still ask is whether AI can write. That one's settled. The question that matters is whether you can build an operation around it that catches its own mistakes, remembers what it learned, and gets better because it runs and breaks and gets fixed.

What the machine can't supply is taste: knowing what to cover, which angle matters, and when a quote doesn't sound right even though it looks right. You can't write that down. But you can build an operation that carries it further than one person could reach alone, to every town that lost its paper and every county commission meeting nobody shows up to anymore.

The old economics aren't coming back. The coverage gap is real. The operation is how you fill it anyway.

Related coverage

• The Newsroom Is a System. I Built One.

• The Soul of the Company Is a Markdown File