Three. That’s how many hikers had to be pulled off California’s Mount Shasta this week after planning their expedition with Google’s Gemini. Not three hundred, not a statistic buried in a safety report — just three guys who asked an AI what to pack and how to get up a mountain, then found out in the dark that the answer wasn’t good enough.
According to the Siskiyou County Sheriff’s Office, the men told a deputy at the end of the rescue that they had relied heavily on Gemini for information about the route and what to bring. The AI’s food and water advice fell short. They ended up descending in darkness, got lost, and needed help getting out. Everyone made it home. The sheriff’s office turned the incident into a public warning: don’t lean on an AI chatbot alone for trip planning. Ask local authorities and the Forest Service instead.
I write about AI agents for people who don’t build them, and this story is one of the clearest teaching moments I’ve seen all year. Not because the technology is bad, but because of the specific gap between what these tools feel like and what they actually are.
A chatbot sounds like an expert because it was trained on experts
When you ask Gemini or any similar assistant about a hike, you get back a tidy, confident answer. Bullet points. A packing list. Estimated timings. It reads like advice from someone who has been there.
But the model hasn’t been there. It’s producing the most plausible-sounding response based on patterns in enormous amounts of text. The tone of authority is a byproduct of how these systems learned to write, not a signal that the underlying information has been checked against reality. There’s no little voice inside saying “I’m 60% sure about the water estimate.” The uncertainty exists — it just doesn’t make it into the output.
That’s the mismatch. Confidence in the presentation, variability in the substance. On most tasks you’d never notice, because being slightly off about a recipe or a movie plot costs you nothing. On a mountain at nightfall, being slightly off about water costs you a lot.
What the AI didn’t know it didn’t know
Trip planning is exactly the kind of task where general knowledge breaks down. The details that keep you alive are local and current:
- How much daylight you actually have on that specific date
- Whether a water source is flowing or dry this season
- Current trail conditions, closures, or snow on the upper route
- How the terrain changes your pace compared to a flat-ground estimate
- What the weather is doing this week, not on average
A chatbot working from training data has a fuzzy, averaged-out picture of all of this. The Forest Service and a county sheriff’s office have the specific one. Same question, wildly different quality of answer, and no obvious way to tell them apart from the way the text looks on your screen.
The pattern worth remembering
I’d put it this way. AI assistants are genuinely useful for the shape of a plan and unreliable for the numbers inside it. Ask one to explain what altitude sickness feels like, or to help you think through what categories of gear a mountain trip needs, and you’ll get something helpful. Ask it exactly how many liters to carry on a specific route on a specific weekend, and you’ve quietly handed a safety decision to a text generator.
The useful habit is to treat AI output as a first draft that needs a second source. For low-stakes questions, skip the second source and move on. For anything where being wrong hurts — health, money, legal matters, remote wilderness — go find the authority who actually maintains the ground truth. That’s a ranger station, a doctor, an accountant, a person whose job is to know.
Not a story about a bad chatbot
It would be easy to read this as “Gemini failed.” I don’t think that’s the lesson, and I don’t think it’s fair. The tool did what these tools do: produced a fluent, plausible answer to a question that needed a verified one. The failure was in the handoff — in how naturally we slide from “helpful suggestion” to “trusted plan” when the writing is smooth enough.
Three hikers got a rough night and a rescue out of that slide. It could have gone worse, and the sheriff’s office clearly knows it, which is why they bothered to post about it at all.
So use these tools. They’re good company for thinking out loud. Just remember that a mountain has never read a single word of anyone’s training data, and it isn’t going to adjust its plans based on what a chatbot told you.
🕒 Published: