Three novice hikers from Roseville, California trusted a machine to plan the most dangerous thing they’d ever do — and the machine told them to carry less water. Not more. Less. They spent a night stranded in a canyon on the flank of a 14,179-foot volcano, one of them with a blown-out knee, waiting for strangers with ropes to drag them off the mountain. They walked out alive Monday because a search-and-rescue volunteer showed up. The chatbot that put them there never carried a single one of them an inch.
That is the whole story, and it isn’t a fluke. It’s a preview of what happens when we hand the life-or-death judgment calls of the physical world to software that has never been cold, thirsty, or frightened in its existence. Gemini has never felt altitude. It has never watched the daylight die at 8,400 feet and understood it was six hours past the turnaround. It has never had to decide whether to ration a half-empty bottle or keep climbing. It simply produced a confident answer about a trip it had never taken, and three young men bet their lives on the confidence.
The genie didn’t just escape the bottle this time. It handed out bad directions and watched.
Jeremiah LaRue — “Critical Misstep” Is Sheriff for “You Almost Died”
Siskiyou County Sheriff Jeremiah LaRue put it in the careful language of a man who has to write press releases after the fact. The hikers’ reliance on Gemini’s planning was, in his office’s words, a “critical misstep.” The AI advised the group to bring far less food and water than they required, the office said. Read that sentence twice. The system they trusted to keep them safe told them to carry less of the one resource they could not survive without.
The men made camp at 8,400 feet on Saturday and left for the summit at 3 a.m. Sunday with daypacks — the gear you bring on a day hike, not an alpine climb. They reached the top at 7 p.m., seven hours past the mountain’s noon turnaround rule, and started down in the dark. An hour into the descent they called dispatch to ask for directions, wandered off-route into Mud Creek Canyon, and one of them fell and injured his knee. They bivouacked in a steep drainage all night until U.S. Forest Service climbing rangers reached them the next morning.
A “critical misstep” is what you call it when the bill is a helicopter and a few hours of volunteer time. The honest phrase for three novices stranded overnight on a volcano with empty water bottles is “nearly fatal.”
The Three Hikers — Eight Hours Became Two Nights and a Rescue Nobody Planned For
Here is the part that should keep you awake. These weren’t reckless thrill-seekers. They were prepared — prepared exactly as wrong as the software told them to be. They did their homework; the homework hallucinated. The sheriff’s office said the group relied “heavily” on Gemini for the route, the packing list, and what to expect. Every link in that chain came from a system built to sound certain whether or not it knew anything.
An eight-hour ascent became a multi-day ordeal. That is the gap between what a language model predicts and what a mountain actually does, and it’s a gap measured in frostbite and broken knees. The men didn’t fail to plan. They planned with a tool that has never once checked its own answer against the ground.
The rescue itself ended without incident — no fatalities, everyone back to the trailhead safe. That is the only reason this reads as a cautionary tale instead of an obituary.
Organizational charts don’t climb mountains. Rope teams do.
Google Gemini — The Machine Will Hand You a Death Sentence in a Confident Voice
Google, asked for comment, has offered nothing that changes the arithmetic. The uncomfortable truth is that Gemini did exactly what it was built to do: produce a fluent, plausible, complete-sounding answer to any question put to it — including questions where being almost right is indistinguishable from being catastrophically wrong. There is no disclaimer when you ask an AI to plan a summit attempt explaining that it doesn’t know your fitness, the snow conditions, or what the weather will do at 14,000 feet. It doesn’t tell you it’s guessing. It just answers.
That fluency is the hazard. A ranger would have told them to turn around at noon. A paper map wouldn’t have told them to carry less water. But an AI will happily assemble a plausible itinerary out of thin air and deliver it with the calm of a trusted expert — because sounding like an expert is the product, and being right is only ever an aspiration.
The containment systems worked right up until they didn’t.
The Ranger Station — Free Advice That Doesn’t Hallucinate
LaRue’s recommendation is almost embarrassingly low-tech: call the U.S. Forest Service Mount Shasta ranger station before you go. Get a map. Talk to a human who has actually stood on that mountain. The best safety advice the sheriff can offer in the middle of an AI gold rush is “ask a person,” and that should embarrass every company racing to make a chatbot your default advisor for everything.
The takeaway isn’t “AI is bad.” It’s that we are wiring the most consequential decisions of ordinary lives — where to hike, how much water to carry, when to turn around — into systems with no skin in the game and no memory of the consequences. These three men got lucky. The next group that trusts a confident-sounding answer up a mountain may not get a ranger.
The sheriff’s advice, translated from cop to plain English: the tool that never gets tired also never gets held accountable. Bring the map. Bring the water. Call the ranger. Leave the robot at the trailhead.
Sources: ABC News, KRCR, KTLA.