How Laterstack Uses AI
You probably want to know whether a person wrote what you’re reading. Fair enough. So here’s the whole thing, including the parts that don’t flatter me.
Laterstack is small. I use AI to do work a bigger newsroom would hand to a staffer, and I’d rather lay that out than let you guess at it.
What it does
Transcription first. Every interview and recorded call runs through it.
Document work, which is honestly most of it. A public records production can land as four hundred pages of permits and monthly reports, and AI is how one person actually reads all of that instead of skimming it and hoping. It also does arithmetic on numbers pulled out of government files, and then I redo the arithmetic myself, because it has been wrong before.
Drafting and structure, which I then rewrite. And finding the holes in my own reporting before it publishes, which is the use I’d defend hardest of any of them. It is very good at explaining why your best finding is probably wrong.
What it never does
Four lines here. These aren’t preferences, they’re code that runs before anything can publish, and a draft that trips one of them doesn’t go out.
No factual claim without a source. If a sentence asserts a fact, it carries a citation to a primary document or to a named person. A draft that asserts facts and cites nothing gets blocked.
No AI-written quotes. Nobody gets words put in their mouth. If something is in quotation marks, that person said it, on a recording or in writing I still have.
No generated or altered images. No AI photos, no AI video, nothing touched up in editorial. If I ever run an AI-made illustration it will say so on the image itself.
Coverage is never a favor. If a story would name someone I’m trying to build a relationship with, the story stops. Every draft gets checked against my own outreach list, and that rule exists because I came close to getting it wrong.
Who checks it
Brianna Rice and I edit Laterstack together. I do most of the reporting, and every story goes through her before it publishes. She checks claims against their sources and looks at how AI was used getting there. Nothing here runs on one person’s judgment.
Your material stays out of it
Confidential source material does not go into third-party AI tools. Off-record conversations, documents somebody handed over in confidence, unpublished records, none of it. This is the rule that costs me the most time and it is not up for discussion.
The system
All of this runs through something I built called Cortex. It handles research, drafting and workflow, and it carries the rules above as code rather than as a promise on a page.
Worth saying plainly: I built every one of those checks after getting something wrong, not before. The sourcing rule exists because I wrote a piece carrying three load-bearing claims with nothing under them. The outreach rule exists because that same piece was pointed at opening a door. A human caught both, the system caught neither, and that is exactly why the system has them now.
What about everything before this page
Laterstack has used AI since it launched. This page writes down practices I was already following, but I had never published them anywhere, and nothing before August 23, 2026 carries a story-level AI disclosure. I’d rather tell you that here than have you turn it up yourself.
From now on, when AI played a real part in how a story got reported, the story says so.
This will change
These tools move faster than a policy page can. I’ll revise this one and date it every time I do.
If you think something on here is wrong, or missing, or that I’m not actually doing it, tell me: lamar@laterstack.com. I answer those.
Related
Corrections Policy
Editorial Standards and Ethics
About Laterstack, including ownership, funding and how we handle disclosure