Nearly everything we publish now has AI in it somewhere. Here is where, what it is not allowed to do, and who answers for the result.
The short version
AI never adds a fact here. It reads records we already hold: public dispatch logs, agency reports, and scanner traffic that our own equipment records and transcribes. Every line it reads carries a source and a trust rating. A person decides what gets called reporting.
We were building this before “AI slop” had a name. The rules below are why our work does not read like it.
Four kinds of content, four labels
Every piece we publish carries one of these:
- Reported. A person wrote it from on-scene work, records and sources. AI may or may not have helped with grammar, spelling, formatting. A person published it.
- AI-assisted, Editor-reviewed. AI drafted it from our incident records. A person checked it against the evidence, edited it and published it. Most of our stories are this, and we have gone back and labeled the earlier ones too.
- AI-researched, Editor-reviewed. Not about a single incident: an explainer or a researched editorial piece. An editor defines the article and outlines, AI researches and drafts, and the editor makes the calls, checks the sources and edits before publishing.
- Automated. Built and posted with no human review, such as SlackerBot alerts and the daily summary. It always carries the line “may contain errors.”
How a story gets built
- Records and radio come in. Dispatch logs from county fire and police agencies, plus scanner audio we record and transcribe ourselves.
- They are matched into one incident. A crash logged by CHP, the fire dispatch, and the radio traffic that follows land in one file, with a record of why they were joined.
- The file becomes an evidence pack. Each line is tagged with where it came from and how far it can be trusted.
- AI drafts. It may only use lines from the pack. It hands back the draft and a list of what it left out and why.
- A person edits and labels it. For automated posts, the disclaimer goes on instead.
The rules the AI works under
- Trust order. Agency records first, then the dispatch broadcast, then working-channel radio, then anything AI itself pulled out of a transcript.
- No claim built on a single radio line that nothing else supports.
- Never turn a summary into a quote from a responder.
- When sources disagree, say so. Two size-ups that conflict get reported as conflicting.
- Sensitive detail is filtered before any text leaves our machines, and checked again after. Patients, minors, protected locations, exact addresses. We keep the specifics of those filters to ourselves.
What that looks like on a real story
What AI never does here
- No fake photorealism. Our photos and video are captured by people. When we use an AI illustration or a composite header, it looks like one.
- No invented quotes, names, causes or outcomes. Gaps stay gaps.
- Nothing called reporting without a person. Automated posts exist, and they are always labeled as automated.
- No voices, no avatars.
To journalists and skeptics
You are right to be suspicious of AI in a newsroom. Working journalists and researchers can ask us for the evidence behind a specific story: the records, the transcript lines, and the model’s own draft with its notes on what it left out. Corrections are marked on the story, not buried.
Evidence requests: Contact Form
