Notes from my part of the opening keynote panel at the LION Independent News Sustainability Summit — “Humans all the way down: The risks, ethics and opportunities of AI in local news.” San Diego, September 9, 2026.
I publish the Moab Sun News, a weekly print paper and daily website in Grand County, Utah — about ten thousand people, and a few million visitors a year on their way to Arches. I’m the newsroom, with some freelance help. I came out of a humanities background and the ACLU of Michigan. I’m not a developer.
Almost everything I said on that stage came out of one project: automating the record of local government meetings. Here’s the short version, and then the part I actually want you to take home.
1. Add your local context
Transcription is a commodity. You can buy it anywhere, and on its own it’s of marginal use — there are commercial tools that will happily transcribe your meetings and do none of the rest of this.
What makes the output worth publishing is everything I hand the model before it writes a word:
- A roster of names and local entities — about 370 corrections I built by hand, so the transcript is fixed before the model ever sees it. There are six recorded spellings of one commissioner’s name in that file. A generic tool doesn’t know those are the same person, which means the transcript it hands you isn’t searchable by the one thing you’d search for.
- The agenda and the full staff packet — so item titles, addresses, applicants and dollar figures come from the official document rather than from audio.
- Three years of prior meetings — so it knows whether an item is new or the fourth time a rezone has come back.
That’s the product. The transcription is just raw material.
And it generalizes well past meetings: the best thing you can do to improve your work with any AI is to understand context. Write a specific document that describes your work in great detail — your bodies, your beats, your names, your rules, the things you routinely get wrong. Everything good in my system came out of writing that document, not out of clever prompting.
2. Think of how this can fail
These weren’t discovered, they were designed. Before any of it ran unattended, I sat down and asked how it could put a false record in front of readers, and wrote a rule against each answer.
- Machine captions are worst at exactly what we can’t be wrong about — one turned $1,157,895 into “one billion, one hundred fifty-seven million.” So dollar figures come from the staff packet.
- When the transcript and the agenda disagree, the agenda wins. The machine is never permitted to correct the public record.
- Sometimes the record genuinely doesn’t resolve — the minutes give an outcome but not how one member voted. So it publishes a question mark and I go back to the video. It never guesses.
- Automation fails silently, so every check has to prove it can fail. A check that only ever comes back clean is indistinguishable from a broken one.
You can write guardrails and housekeeping into your work. None of that was luck.
3. Most of this isn’t AI
Five stages, and exactly one of them is a language model writing anything. The rest is ordinary software — a shell script, a transcription service, 278 regular expressions, a database table, a hook that unpacks the votes and renders them everywhere. By line count, roughly 58% of the system is plain deterministic code.
What AI did was let me write that code. That’s the part I think gets missed. Using AI is not the same thing as prompting a chatbot and publishing what comes back.
And you can push it further than people expect: you can use AI to build a hard-coded system that does what you want, and then not use AI or LLM services in an ongoing way at all. The model was scaffolding. Deterministic code doesn’t drift, doesn’t lean on a data center, and gives you the same output every time — which is a more professional result, not a compromise.
Where to go next
Two things, and they’re the honest answer to “what do I actually do on Monday.”
Write a context document. That’s the first takeaway above, and it’s the one that generalizes past meetings to any kind of work. I wrote up how and why to write one for your organization — what goes in it, how to build it from things that went wrong rather than from imagination, and the traps that quietly ruin it.
Then go to technical sources, not the AI-and-journalism conversation. There are a lot of hot takes that won’t help your comprehension. The people I actually read and watch are on my resources page — it’s short on purpose. You don’t have to want to build anything to benefit from understanding how these systems work; it’s also what lets you tell a capable tool from a thin wrapper when someone is selling you one.
What’s running in Moab
179 meetings and 8,722 individual vote records going back to January 2023 — the Grand County Commission, Moab City Council, and both planning commissions, including two bodies nobody else covers. Full-text searchable transcripts timestamped into the video, a per-official vote database, the print edition, and a twice-weekly civics email.
Transcription runs about 38 cents per hour of audio.
You can look at the whole thing: the Grand County Commission hub and the Moab City Council hub.
Research I referenced
- Jia, H., Appelman, A., Wu, M., & Bien-Aimé, S. (2024). News bylines and perceived AI authorship: Effects on source and message credibility. Computers in Human Behavior: Artificial Humans, 2(2), 100093.
- Gallegos, I. O., Shani, C., Shi, W., Bianchi, F., Gainsburg, I., Jurafsky, D., & Willer, R. (2026). Labeling messages as AI-generated does not reduce their persuasive effects. PNAS Nexus, 5(2), pgag008.
- Reuters Institute for the Study of Journalism. Generative AI and News Report 2025.
Get in touch
I’d genuinely like to hear from you, especially if you’re trying this in your own community and something isn’t working. maggie@moabsunnews.com · Moab Sun News