Gemini Packed Their Bags and Mount Shasta Sent the Bill

Picture it. You’re somewhere on Mount Shasta, the light is gone, and the trail you thought you were on has stopped being a trail. Your water ran out a while back. The food situation is worse. Somewhere in your phone is a chat log that told you what to bring, and it did not mention this part.

That’s roughly where three hikers ended up this week before a rescue crew pulled them out. All three made it. According to the Siskiyou County Sheriff’s Office, the men told a deputy on scene that they had relied heavily on Google’s Gemini for information about the route and what to pack. The food and water advice was not adequate. They descended in the dark and got lost. The sheriff’s office posted a public warning afterward, telling hikers to check with local authorities and the Forest Service instead.

I review AI tools for a living. I use them daily. And I want to be careful here, because the easy version of this story is “AI bad, people dumb,” and that’s not what happened.

What actually failed

The failure wasn’t that Gemini lied. It’s that a general-purpose chatbot was asked to do a job it has no way of doing well, and it answered anyway with total composure.

Ask a model how much water to bring on a mountain and it will produce a number. That number comes from patterns in text, not from knowing your pace, your fitness, the temperature that day, whether there’s snow, whether the water sources on the route are running, or how long the descent takes once the sun drops. A ranger knows some of that. A recent trip report knows some of that. A model trained on the internet knows what a plausible answer looks like.

Plausible is the dangerous part. If Gemini had said “I don’t know, ask the Forest Service,” these three men would have had a boring trip. Instead they got a confident packing list. Confidence is the product feature that makes these tools pleasant to use, and it’s the same feature that makes them unsafe in situations where being wrong has a physical cost.

The category error nobody warns you about

Here’s the pattern I keep running into when I test these things. AI assistants are excellent at tasks where you can check the output yourself, and unreliable at tasks where you can’t check it until it’s too late.

• Draft an email? You read it. You’d notice if it were wrong.
• Explain a Python error? You run the code. Feedback in seconds.
• Summarize a document you already have? You can spot-check it.
• Tell you how much water to carry up a 14,000-foot mountain? You find out when you’re out of water.

That last category is where people get hurt, and it’s the category the marketing never distinguishes from the others. The interface is identical. The tone is identical. Nothing in the chat window signals that you’ve crossed from “verifiable in ten seconds” into “verifiable only through consequences.”

What I’d want from the tool

Some of this is a product problem, not a user problem. Trip planning for backcountry routes is a known high-stakes query type. It would not be hard for an assistant to recognize it and behave differently: refuse to give specific quantities, point to the managing agency, name the actual authority for the area. Google has the ranger district data. The gap between “here’s a packing list” and “here’s the Forest Service page for this exact mountain, call them” is a design decision, not a technical limit.

Until that changes, the burden sits with you.

How I’d use AI for a trip like this

Not never. Just differently. A model is fine as an orientation layer, the thing that tells you what questions exist. It’s not the answer layer.

• Use it to learn vocabulary and general concepts, then take those terms to a primary source.
• Get the name of the agency that manages the land. Then contact that agency yourself.
• Treat every specific number, quantity, distance, or time estimate as unverified until a human with local knowledge confirms it.
• Assume the model has no idea what conditions are like right now, because it doesn’t.
• Plan for the descent taking longer than anything told you, and carry a light either way.

The three hikers on Shasta got lucky in the way that matters most. They came home. The tool they trusted did what it was built to do, which was produce a fluent answer, and that turned out not to be the same thing as producing a correct one.

That distinction is the whole job of reviewing these tools honestly. A chatbot that sounds sure of itself is not a source. It’s a starting point, and the mountain doesn’t grade on tone.

🕒 Published: September 7, 2026