A quote came back last week that didn't sound right. It was attributed to someone real, on a topic we'd covered before, and the phrasing was close enough to pass. But it wasn't exact. The words had drifted — the AI had been in a long production thread, context had shifted between beats, and by the time it pulled that quote it was working from a degraded version of what the source actually said.
We caught it. Not because someone happened to notice, but because the system caught it. The production log flagged a mismatch between the quote in the draft and the verified version in the beat research file. We traced it back to thread drift — the AI's context had decayed over the course of a long session. So we added a step: reload source material at the start of every new piece, even mid-session. That's what the system is supposed to do. Surface the problem, fix the process, move on.
That small correction is more representative of what we've built at Mercury Local than any description of the technology could be.
What We Built
Mercury Local publishes across The Charlotte Mercury, The Farmington Mercury, and Strolling Ballantyne. Each publication has its own authors and its own editorial voice. Jack Beckett covers Charlotte government — dry, precise, the kind of reporter who actually read the procurement document before writing about it. John Speedway does Charlotte sports like the last great local anchor who showed up to the game and won't stop talking about it. Nell Thomas covers Ballantyne like a neighbor who's been to every restaurant on the strip and will tell you which ones are worth it and why.
None of them sound alike. None of them should. And none of them exist without a system underneath that tells the AI who they are, what they've already published, and what facts have been verified — before it writes a word.
That system is the difference between writing an article and running an editorial operation. Anybody can get AI to draft something. The harder question — the one we spend our time on — is whether you can sustain an operation that produces consistent, verified, voice-accurate local coverage across multiple publications, day after day.
You Build It by Running It
This is the part that doesn't show up in the conference talks about AI and media. The system is not a thing you design and ship. It's a thing you build by running it.
Every production session surfaces a gap. A voice that drifts flat when explaining legal context. A research file that should have existed before the draft started. A quote that decayed because the thread ran too long. You fix the gap, update the documents, run it again. The next session is better because the last one broke something.
Brian Alvey, the CTO of WordPress VIP, made a version of this argument on a recent podcast. He's spent thirty years building publishing systems for media companies the size of News Corp. His framing: if you can document the decisions your business makes — the voice rules, the quality checks, the production logic — you've built something an AI collaborator can operate inside of. Not replacing humans. Working within a structure humans designed.
He's right. But the part he's describing from the theory side, we can tell you from the production side: the system is never done. It improves because it runs, and it runs because someone decided this is what the operation looks like today, knowing it will look different tomorrow.
We've been building in public since Mercury Local started. The core conviction hasn't changed: the newsroom is a system. The economics that supported the old version are gone. But the system itself — the editorial judgment, the voice discipline, the fact verification, the beat knowledge — can be rebuilt. It has to be rebuilt differently. It's being rebuilt now.
The Question That Matters
The question most people ask about AI and media is whether AI can write. We're well past that. The question that actually matters is whether you can build an operation around it — one that self-corrects, that accumulates institutional knowledge, that gets better because it runs and breaks and gets fixed.
That's what we wake up doing. Not chasing a breakthrough. Doing the accumulated work of running the operation every day and tightening what needs tightening.
Taste is the thing AI can't do — knowing what to cover, which angle matters, when a quote doesn't sound right even though it looks right. You can't encode that. But you can build an operation that lets it reach places it couldn't before. Every city that lost its paper. Every community that stopped being covered. Every county commission meeting nobody shows up to anymore.
The coverage gap is real. The old economics aren't coming back. The operation is how you fill the gap anyway.