Picture it. You’re somewhere on Mount Shasta, the light is gone, and the trail you thought you were on has quietly stopped being a trail. Your water ran out a while back. So did the snacks. The packing list you followed came from a chat window, and that chat window is now a glowing rectangle in your pocket that cannot carry you down the mountain.
Three hikers lived that exact evening this week. They were rescued from Mount Shasta after descending in the dark and getting lost. When deputies asked what happened, the men said they had relied heavily on Google’s Gemini for information about the route and for what to pack. The food and water advice was not enough for the trip they actually took. The Siskiyou County Sheriff’s Office posted about the incident and asked other hikers to check with local authorities and the Forest Service instead.
I review AI tools for a living. I use them daily. And this story is the cleanest example I’ve seen of the failure mode I keep trying to explain to people who email me asking which model is “the most accurate.”
Accuracy is not the same as accountability
Ask a chatbot how much water to bring on a mountain and you will get a number. The number will sound reasonable. It will be formatted nicely, probably in a bulleted list, possibly with a friendly note about staying safe out there. What you will not get is anyone who is on the hook if the number is wrong.
That gap is the whole story. A ranger who tells you the creek at mile four is dry has skin in the game and current information. A model has neither. It has patterns from text, and text about hiking tends to describe average conditions on average days for average people. You are not average conditions. You are three specific people on one specific mountain on one specific afternoon with one specific amount of daylight left.
The tool did not malfunction. It did exactly what it does. It produced fluent, plausible output on a topic where fluent and plausible are not enough.
Where I draw the line in my own workflow
I’ve developed a rough test for whether I’ll trust AI output without verification, and it comes down to one question: what happens if it’s wrong?
- Wrong is annoying. Bad variable names, a clunky first draft, a recipe that needs more salt. Ship it, fix it later. Chatbots are great here.
- Wrong is expensive. Tax categories, contract language, a database migration. Use the AI for a first pass, then have a human who knows the domain check it.
- Wrong is dangerous. Dosages, electrical work, structural loads, and yes, how much water to carry above the treeline. The AI is not the source. It can help you form questions to ask the actual source.
Trip planning slides into that third bucket faster than most people expect. It feels like the first bucket. It looks like the first bucket. You’re just asking about a hike. But the consequences of a bad answer scale with elevation and subtract with daylight, and neither of those variables is visible in the chat window.
The interface is doing something to us
Part of what makes this hard is that these tools answer everything with the same confidence and the same tone. A question about a Python list comprehension and a question about summiting a 14,000-foot volcano get the same calm, organized, mildly encouraging reply. There is no visual signal that one answer is well-grounded and the other is a statistical guess dressed as advice.
Search engines were messier, and that mess was accidentally useful. Ten blue links with conflicting information forced you to notice you were making a judgment call. A single tidy response removes that friction, and the friction was doing real work.
I’d love to see AI assistants get better at routing. Ask about backcountry conditions and the most useful reply might be a Forest Service phone number and a link to a current permit page, not a packing list. Some products are starting to move that direction for medical and legal questions. Outdoor safety deserves the same treatment.
What I’d actually tell a friend
Use the chatbot to learn the vocabulary. Ask it what questions to ask a ranger. Ask it to explain what a Class 3 scramble means, or what glissading is, or why people bring an ice axe up Shasta in July. That’s genuinely useful. It compresses the beginner learning curve in a way nothing else does.
Then close the tab and call someone who was on that mountain this week.
The three hikers made it home, which is the part that matters most. They also handed the rest of us a clear lesson at their own expense: these tools are excellent at sounding prepared, and being prepared is a different skill entirely. Your chatbot has never run out of water. It cannot imagine what that feels like. Plan accordingly.
đź•’ Published: