Google Gemini gave three hikers bad advice, and they nearly paid for it with their lives. According to TechCrunch, the group used Gemini to plan a Mount Shasta expedition, and the chatbot reportedly told them to pack far less food and water than they actually needed. What was supposed to be an 8-hour climb turned into a multiday ordeal that ended with a Forest Service rescue operation.
The basic facts, as reported by the Siskiyou County sheriff’s office, are worth sitting with. The three men started their hike at 3am. Standard guidance on Mount Shasta tells hikers to turn around if they haven’t reached the summit by noon. This group reached the top at 7pm. They then attempted to descend in the dark, called the sheriff’s office asking for directions, spent the night stranded in Mud Creek Canyon, and were rescued the following morning by rangers and volunteers.
The sheriff’s office was direct about Gemini’s role: the hikers were advised by the AI to bring supplies well short of what their group needed, a problem that became critical once a planned day hike stretched into something much longer. That said, the chatbot didn’t make the call to keep climbing past noon, or to attempt a night descent. These were human decisions, and bad ones.
But the broader point matters. Gemini, like ChatGPT, Claude, and every other general-purpose AI assistant, is not built for high-stakes, real-time outdoor planning. These models draw on general training data. They have no live weather feeds, no knowledge of current trail conditions, and no way to account for a specific group’s fitness level or experience. When someone asks Gemini how much water to bring on a technical mountain climb, the model will produce an answer, but that answer carries none of the liability or local expertise that a ranger station call would.
The sheriff’s office put it plainly: always contact the local USFS Mount Shasta ranger station before a trip, and never rely solely on AI for planning. That’s good advice, though it’s the kind that tends to land only after something goes wrong.
For developers and product teams building on top of these models, this incident is a useful data point. AI assistants are being used for decisions their architecture was never designed to support. The gap between what users trust these tools to do and what they can actually do reliably is still wide. This rescue is a concrete example of what that gap looks like in practice.
