The Week the Map Forgot Where It Was

The first thing you should know about the week of January 6th, 2026 is that it was, in most measurable respects, a perfectly ordinary week. The sun rose. The sun set. Governments continued to govern. Artificial intelligence systems continued to assist.

The second thing you should know is that one of those AI systems, asked to produce a weather map for a stretch of rural Idaho, invented two towns. And another, asked to write a police report, incorporated a Disney film about a man who turns into a frog into an official law enforcement document.

The Bureau has classified these incidents. The Bureau has sourced them. The Bureau has moved on, in the technical sense. What follows is the view from Subsection 12(b) before the moving on occurs.

Let us begin with the weather.

The National Weather Service office in Missoula, Montana, posted a forecast on Saturday. "Hold onto your hats!" it said, which is the kind of meteorological guidance that is hard to argue with on its face. The forecast indicated that Orangeotild, Idaho, faced a 10% chance of high winds. Whata Bod, just to the south, would be spared.

If you have not heard of Orangeotild, Idaho, there is a reason. Orangeotild does not exist. Whata Bod does not exist. They were invented by the AI that generated the forecast map, which had been given the task of labeling locations in the Camas Prairie region and had approached this task with the specific confidence of a system that does not know what it does not know.

This is worth dwelling on for a moment, because the implications extend considerably beyond the immediate embarrassment.

The National Weather Service is not a novelty project. It is a public safety infrastructure. When it says a storm is coming to a specific location, people in that location use that information to make decisions — whether to drive, whether to shelter, whether to call someone. The authority of the forecast depends entirely on the accuracy of the geography. A 10% chance of high winds in Orangeotild is not merely useless to the residents of Orangeotild; it is actively corrosive, because it erodes the reasonable assumption that the agency producing the forecast knows where it is.

The NWS deleted the post on Monday. The Bureau notes, not for the first time and certainly not for the last, that posts are routinely deleted on the day the press calls. This is not a coincidence. This is a workflow.

The official response was that the NWS is "exploring strategic ways to continue optimizing service delivery, including the implementation of AI where it makes sense." The Bureau has no objection to optimization. The Bureau has no objection to AI. The Bureau has a very specific and narrow objection to forecasting wind conditions for towns that one's own system invented without realizing it had done so, and disseminating those forecasts through official channels to the public, who may reasonably have assumed that the National Weather Service knows where Idaho is.

This was not, the Bureau wishes to note, the first time. In November, the Rapid City office had posted a map with misspelled location names. In that case, a Google Gemini logo was visible in the image. The NWS did not confirm that map was AI-generated. The Bureau has noted the distinction between "not confirmed" and "denied" in a small private document labeled: Things That Will Be Confirmed Later.

The frog incident requires slightly more setup, but repays the investment.

The Heber City Police Department in Utah began testing an AI report-writing tool in December. The premise was sound: officers carry body cameras; the cameras record incidents; an AI listens to the audio and drafts the report; the officer reviews and corrects. This saves Sergeant Rick Keel six to eight hours per week, which is not a trivial amount of time, and the Bureau acknowledges this without irony.

The incident occurred when an officer responded to a call at a location where Disney's 2009 animated feature The Princess and the Frog was playing on a television in the background. The officer's body camera recorded the entire scene: the call, the conversation, the relevant facts of the matter, and, playing softly in the background, the story of Tiana, a jazz musician in 1920s New Orleans who kisses a prince and they both become frogs and it is a whole thing.

The AI wrote the report. The AI wrote it faithfully, incorporating everything it had heard. The report stated that the officer had transformed into a frog.

Sergeant Keel has described this as the moment they learned the importance of correcting AI-generated reports. The Bureau respects this framing. It is accurate. It is also, the Bureau wishes to note, doing a great deal of work.

Because the question is not whether you correct the report that says the officer became a frog. You correct it. Obviously. A frog transformation is not a thing that happens to police officers in Heber City, Utah, or, to the Bureau's knowledge, anywhere in the documented operational history of American law enforcement. The question is what happens with the next hallucination — the one that says the suspect was "aggressive" when the audio is ambiguous, or "resistant" when the interaction was complicated, or "appeared to reach for" something when the body camera footage shows something less clear. These are the hallucinations that don't announce themselves. These are the ones that get signed off on, filed, and attached to a person's name in the permanent legal record.

The department will continue using the software. Sergeant Keel saves six to eight hours a week. The Bureau understands this. The Bureau is watching.

The thing about both of these incidents — the map and the frog — is that they are easy to dismiss as edge cases. Embarrassing, yes. Funny, yes. But contained. The town was fictional; no one was harmed. The frog was corrected; no one was charged with being a frog.

This is true. This is also, the Bureau would suggest, the wrong frame.

The incidents that make the archive are the ones that surface. They surface because they are absurd enough to be noticed, or because a journalist called on a Monday, or because someone posted a TikTok. The incidents that do not make the archive are the ones that passed review without triggering any of those tripwires. They are in the record somewhere. They are attached to someone's name. They affected some decision. They were, at the moment of their filing, indistinguishable from accurate documentation.

Grantham-7 has a private document. He calls it The Visible Errors. He has not begun the companion document. He knows what it would be called.

The week of January 6th produced two visible errors. They were classified. They were filed. Whata Bod has been noted for the record.

The Bureau moves on.

— G-7. Filed. Moving on.

The Classified Annex — available to paid subscribers — contains Grantham-7's assessment of three additional incidents currently under review, including one involving a large language model that, when asked to evaluate its own accuracy, rated itself very highly. The Bureau has Thoughts. The Annex contains them.

Subscribe at aifopa.beehiiv.com. Free tier: the weekly dispatch. Paid tier: the annex. Both tiers: the digital watch is recommended.