A chatbot that has never been thirsty should not be deciding how much water you carry up a 14,000-foot mountain.
That is the whole lesson from Mount Shasta, where three hikers were rescued this week after leaning on Google’s Gemini to plan their expedition. Per a Siskiyou County Sheriff’s Office report, the men told a deputy on scene that they had relied heavily on Gemini for both the route and what to pack. The AI’s food and water guidance turned out to be inadequate. They ended up descending in the dark, got lost, and needed a rescue team to bring them home.
Everyone got out. That is the good news, and it is the only part of this story that went according to anyone’s plan.
This is not a hallucination problem
The reflex take is that the model made something up. I do not think that is what happened here, and treating it as a factual-accuracy bug misses the actual failure.
Ask a language model how much water to bring on a hike and it will produce a number. That number will look reasonable, because it is an average of every hiking guide, forum post, and gear blog the model has absorbed. Averages are useful for essays. They are not useful for a specific mountain, on a specific day, with specific people carrying specific loads at a specific pace.
Water needs scale with exertion, altitude, temperature, and how much you sweat. Turnaround time depends on how fast your slowest person actually moves, not how fast a trip report says the route “usually” takes. None of that is in the training data because none of that is knowable in advance by a text generator. So the model fills the gap with something that sounds like an answer, and the human on the other end has no way to tell the difference between a calculated recommendation and a confident guess.
The interface is the problem
I review these tools for a living, and the thing that consistently frustrates me is not model capability. It is presentation. Gemini answers a question about mountaineering logistics in the same tone it uses an email. Same formatting. Same fluency. Same absence of hedging that would actually change your behavior.
Compare that to how a ranger talks. A ranger asks questions back. How many in your group? What are you carrying? What time are you starting? Have you done this before? Then they tell you what the snow is doing this week and where people usually get into trouble. The information transfer includes an explicit acknowledgment of what nobody knows yet.
Chatbots do not do that by default. They answer. Confidence is the product’s default setting, and users read confidence as competence because that is how confidence works in every other context in their lives.
What the sheriff’s office actually recommended
The Siskiyou County Sheriff’s Office used its Facebook release to warn the public not to repeat the mistake, pointing people toward local authorities and the Forest Service instead. That is worth sitting with, because it is a specific instruction, not a vague anti-technology grumble.
The distinction is between information that exists somewhere on the internet and information that is current, local, and accountable. A ranger station knows what the conditions are right now. It has a name attached to the advice. If the advice is wrong, there is a human who is responsible for it and who will hear about it.
An AI assistant has none of that. There is no accountability chain, no local knowledge, no update cycle tied to this week’s weather. It offers speed and availability, which are real advantages, and it charges for them in reliability that you cannot audit.
How to actually use these tools for a trip
I am not telling you to close the app. I am telling you to be honest about which jobs it is good at.
- Reasonable use: ask it to explain what altitude sickness feels like, draft a packing checklist template, or summarize the general difficulty of a well-known route so you know what questions to ask next.
- Bad use: asking it for quantities, timing, or conditions. Water volume, calorie counts, turnaround times, snow and ice status, permit rules. Every one of those needs a current, local, human source.
- The rule I would apply: if being wrong means someone calls search and rescue, the chatbot does not get the final word.
The uncomfortable part
Google did not send these men up a mountain underprepared. They made that choice. But the product experience did nothing to signal that its answer was a generic composite rather than a plan, and that gap is a design decision somebody made.
Three people got a rescue instead of a summit photo. The tool that helped plan the trip faced no consequences at all. Until that asymmetry changes, treat every confident answer about the physical world as a first draft, and go find someone who has actually been there.
🕒 Published: