The Siskiyou County Sheriff's Office in California reported a search-and-rescue operation: three young men from Roseville got lost while descending Mount Shasta and were rescued after being stranded overnight. They told a sheriff's deputy that they had relied mainly on Google Gemini to plan their route and gear list. TechCrunch, ABC News and CBS picked up the story around September 5.
According to the sheriff's account, the three first made camp at roughly 8,400 feet, then set out for the summit push at 3 a.m. the next day carrying day packs. The climb was originally planned as an 8-hour ascent, but they didn't reach the summit until 7 p.m. — well past the noon turnaround time recommended for the route. The descent ran into darkness; about an hour after starting down, they called the sheriff's office dispatcher for directions, then wandered off route into the Mud Creek drainage on the mountain's south side. One of them injured a knee, and the group had to bivouac in the steep drainage overnight until they were reached by a U.S. Forest Service climbing ranger team the next morning, with the sheriff's SAR team joining shortly after.
Investigators Zeroed In on the Packing List
The report's conclusion is specific: Gemini recommended far less food and water than the group actually needed, especially once the planned 8-hour ascent turned into a multi-day ordeal. The sheriff's office followed up with advice not to rely solely on AI for trip planning, and to check in with the local ranger station before setting out.
The report leaves out more than it includes. What exactly the three asked Gemini, what the list actually said, whether they followed it item by item, which model version they used, and whether it was connected to real-time weather or route data — none of that is in the public record. Google has not commented. Pinning the blame entirely on the model isn't supported by what's available; pinning it entirely on the three men doesn't erase the line in the report about food and water.
A Similar Gap in China
The same pattern shows up in China, where hikers increasingly ask large language models for route advice and packing lists, especially on trails without established guide services. The difference is the safety net: in this case, the sheriff's office is the lead search-and-rescue authority for the area, and the report pointed to a concrete fallback — calling the local ranger station. In China, the equivalent entry points are scenic-area management offices, local rescue teams and official route-registration channels, but they're scattered and rarely show up on the first page of search results.
The list a model produces can look complete and sound confident. A guide who actually knows the mountain would first ask how many people are in the group, their fitness level, and where the snow line sits that season before deciding how much water to carry.
There isn't yet enough of a public record on incidents like this to judge how risky AI-planned trips are overall. What's clear is a gap in the process: the model doesn't know a user's fitness, experience or the mountain's conditions that day unless the user volunteers it — and most people don't. The variables that matter most for trip planning are exactly the ones the model can't ask about.
Sources: Siskiyou County Sheriff's Office statement, TechCrunch, CocoLoop, ABC News, CBS Sacramento; the 8,400-foot camp elevation, the 3 a.m. departure and 7 p.m. summit, the recommended noon turnaround, the Mud Creek drainage location, and the agencies involved in the rescue were cross-checked against the sheriff's account.