It’s getting dark on Mount Shasta. The temperature is dropping the way it does on a big mountain, fast and without asking permission. Three hikers are in a steep canyon, off the route they meant to be on, and the water bottles are lighter than they should be. Somewhere in a phone, there’s a tidy little itinerary that told them this would be enough.
They were rescued. That’s the good news, and it’s worth sitting with for a second, because this story could have ended much worse. Three novice hikers used Google’s Gemini to help plan their climb, the AI advised them to bring less food and water than they actually needed, and they ended up stranded overnight until help arrived. The Siskiyou County Sheriff’s Office followed up with a public warning: don’t plan your trip by AI alone. Talk to local authorities. Talk to the Forest Service.
I write about AI agents for people who don’t build them, and this is one of those stories I want to slow down and unpack, because the lesson isn’t “AI bad.” The lesson is more useful than that, and it applies to a lot more than hiking.
What the chatbot was actually doing
When you ask an AI assistant how much water to bring on a mountain, it doesn’t check the mountain. It doesn’t know the weather that week, the snow conditions, whether the route is icy, or how fast you personally walk uphill with a pack on. It doesn’t know that you’re three people who have never done this before.
What it does is generate an answer that looks like the answers it has seen. Fluent, organized, confident. A bulleted packing list with numbers next to each item. That formatting is the trap, honestly. A list of numbers reads like a measurement, but it’s a prediction of what a plausible list looks like.
Here’s the mismatch that got these hikers into trouble:
- They needed a number grounded in current conditions on a specific mountain.
- They got a number grounded in patterns from text.
- Both look identical on a phone screen.
The AI has never been thirsty. It has never had cramping legs at 9,000 feet or watched the light go while it was still on the wrong side of a canyon. It has read about those things. That’s a genuinely different kind of knowing, and the difference doesn’t show up in the output.
The stakes test
I’ve started thinking about AI help in terms of one question: what happens if this answer is wrong?
If I ask a chatbot to suggest dinner ideas and it gives me a mediocre one, I eat a mediocre dinner. If I ask it to rewrite an email and the tone is off, I edit it. Low stakes, easy to reverse, fine to just go with.
Water on a mountain sits at the other end of that scale. Being wrong isn’t inconvenient, it’s dangerous, and you can’t undo it once you’re up there. Same category as medication doses, legal deadlines, electrical wiring, tax filings, anything involving a rope. When being wrong is expensive and hard to reverse, the AI answer is a starting point for your research, not the end of it.
The sheriff’s office pointed at exactly the right fix, and it’s worth naming why it works. Local authorities and the Forest Service have something Gemini structurally cannot have: eyes on that mountain, this week. Current trail conditions. Recent incidents. Actual snowpack. A ranger can tell you the thing that isn’t in any training data yet, because it happened on Tuesday.
How to use AI for a trip without ending up in a canyon
None of this means you should close the app. AI assistants are genuinely useful for planning, as long as you use them for the parts they’re good at.
- Ask it what questions to ask. “What should I check before climbing Mount Shasta as a beginner?” is a great prompt. It’ll surface things you hadn’t considered, and then you go verify each one.
- Use it to understand jargon. Have it explain what a permit requirement or a trail rating means in plain language.
- Don’t accept its numbers. Quantities, distances, times, weights: check those against an official source every time.
- Make one real human call. A single conversation with a ranger station would likely have caught the shortfall here.
- Treat confidence as style, not evidence. These systems sound equally certain when they’re right and when they’re guessing.
What I keep coming back to is that these three hikers weren’t careless. They planned. They did the responsible-seeming thing, which was to research before going. They just used a tool that wasn’t built to answer the question they were asking, and nothing about the tool signaled that.
That’s the part worth carrying with you. The skill we all need now isn’t avoiding AI. It’s knowing which of your questions the AI can genuinely answer, and which ones need someone who has actually stood on the mountain.
🕒 Published: