Gemini told hikers to skimp on food and water on Shasta

Three men climbing Mount Shasta got stuck overnight in Mud Creek Canyon after Google's Gemini planned their trip, according to the Siskiyou County sheriff's office. The group started at 3 a.m. and didn't summit until 7 p.m., seven hours past the noon turnaround time guides recommend for the mountain. Rangers reached them after they spent the night exposed.

The sheriff's office said the men had been "advised by Gemini to bring far less food and water than their group required," per TechCrunch's report. That's a specific, named failure mode: not a hallucinated fact in a chat window, but a provisioning error that changed what three people carried onto a mountain that kills people who get its weather wrong.

What's missing from the account is the prompt. We don't know if they asked Gemini for a packing list, a turnaround estimate, or both, and we don't know how specific the mountain, route, or season details were in their request. A vague question gets a generic answer, and Shasta's summit routes vary enough in difficulty and exposure that a general-purpose model without local trail data is guessing at load-out just like an inexperienced hiker would.

This is the first search-and-rescue case I've seen where a sheriff's office named a specific assistant's advice as a contributing factor, and that's the part worth sitting with. People are treating chatbot output as trip-planning authority for activities where the failure mode is exposure and dehydration, not a wrong answer on a quiz. Mountain rescue teams have spent decades building turnaround-time and provisioning rules into training precisely because people underestimate both under normal circumstances. An AI tool that skips that training doesn't remove the risk, it just moves the wrong guess earlier in the process, before anyone's stress response kicks in to catch it.