Mount Shasta Doesn’t Care What Your Chatbot Told You

Gemini didn’t fail these hikers, and neither did AI in general — the planning process failed, and a chatbot was the only thing in the room pretending to be one.

Here’s what we know. Three hikers were rescued from California’s Mount Shasta after using Google’s Gemini to plan their expedition. According to the Siskiyou County Sheriff’s Office, the men told a deputy at the scene that they had relied heavily on Gemini for information about the route as well as what to pack. The advice on food and water turned out to be inadequate. They ended up descending in the dark, got lost, and needed rescuing. All three made it out. The sheriff’s office then posted a public warning urging hikers not to repeat the mistake, and to consult local authorities and the Forest Service instead.

That’s the whole story. It’s small, it’s specific, and it’s going to get flattened into whatever headline serves someone’s argument. So let me be blunabout what I actually think is going on here.

Chatbots are confident about things they cannot know

The core problem isn’t that Gemini gave wrong numbers for food and water. Any guidebook can be wrong. The problem is that a general-purpose language model will answer a question about a specific mountain on a specific week with the same steady tone it uses to explain photosynthesis. There is no wobble in the voice. No “I don’t have current conditions.” No “call the ranger district, they’ll know if the spring is running.”

Trip planning is one of those tasks that looks like an information retrieval problem and is actually a risk assessment problem. How much water you carry depends on your pace, the temperature that day, whether snowmelt sources are flowing, how much you sweat, and how badly things go if you’re wrong. A model producing a plausible average answer is doing exactly what it was built to do. It’s just that the average answer is useless on a mountain where being off by two liters is the difference between a long day and a rescue call.

Same goes for the descent in the dark. Somewhere in the planning, the time budget didn’t survive contact with reality. That’s not a hallucination problem. That’s the absence of anyone in the loop asking the deeply unglamorous question: what happens if this takes three hours longer than we think?

The part where I stop blaming the tool

I review AI tools for a living, and I’m tired of the reflex where every bad outcome becomes proof that the technology is fundamentally broken. It isn’t. These hikers made a choice that a lot of people are quietly making right now — they treated a chatbot as an authority rather than a starting point. The sheriff’s office named the fix in the same breath as the warning: talk to local authorities and the Forest Service. Those are the sources with current conditions, closure notices, and actual knowledge of the terrain.

A model can help you build a packing list. It can explain what altitude does to your hydration needs. It can draft questions to ask a ranger. What it cannot do is take responsibility, and responsibility is the entire product in backcountry travel.

What this says about where AI assistants are heading

The interesting failure here is a design failure, not a model failure. If you build an assistant that people will obviously use for consequential physical-world planning, you need refusal and escalation behavior baked in. Not a legal disclaimer buried under the answer. Something closer to: I can sketch a rough plan, but conditions on Shasta change weekly and I have no live data. Here’s the ranger district number. Call before you go.

That’s a harder product to ship because it feels less capable. Users like answers. They don’t like being told to make a phone call. But the gap between “sounds authoritative” and “is reliable” is exactly where incidents like this live, and no amount of model scaling closes it, because the missing piece is current, local, physical information that nobody has fed into the system.

How to use these tools without needing a helicopter

• Use a chatbot to generate questions, not answers, for anything with physical consequences.
• Verify route details, water sources, and conditions with the agency that manages the land. On federal land, that’s the Forest Service or the Park Service.
• Assume any packing recommendation is a floor, not a target, and pad it.
• Build a turnaround time and honor it, regardless of what your plan said.
• If the model sounds certain about something it can’t possibly know, that certainty is a bug you’re supposed to notice.

Three people got off that mountain, which makes this a cheap lesson. The sheriff’s office is right to publicize it. My read is simpler than the headlines suggest: an AI assistant gave average advice to a situation that needed specific advice, and nobody in the group was assigned the job of checking. That job still belongs to a human. It’s going to for a long while.

🕒 Published: September 6, 2026