Why AI Told Hikers to Pack Too Little Food

The Rescue That Sparked a Reckoning

<cite index="1-1">Three hikers were rescued from California's Mount Shasta this week after using Google's AI chatbot Gemini to plan their expedition, according to the Chicago Tribune.</cite> What should have been a straightforward climb turned into a multi-day survival situation, all because a widely-used AI model gave dangerously wrong advice about basic supplies.

This wasn't just bad luck or user error—it's a case study in why you shouldn't outsource critical decisions to general-purpose AI, no matter how capable those systems claim to be.

What Actually Happened

<cite index="5-6">The hikers, described as "novice" climbers from Roseville, California, admitted to relying heavily on Gemini while preparing for their journey.</cite> <cite index="2-4">After the rescue, the men told a sheriff's deputy that they relied heavily on Google's Gemini AI for information on the route as well as what to pack for their trip.</cite>

The core problem: <cite index="3-5">they were advised by Gemini to bring far less food and water than their group required, especially when their planned 8 hour ascent became a multi-day ordeal.</cite> The hikers weren't casual day-trippers—they were attempting a serious alpine climb at 14,179 feet.

<cite index="1-2">A report from the Siskiyou County sheriff's office said that three young men set off on their hike at 3am and — although hikers are told to turn around if they haven't reached the summit by noon — they made it to the top at 7pm.</cite> That's a red flag. <cite index="6-3,6-4">They called the Siskiyou County sheriff's office for directions, but opted to veer off course, ending up in an adjacent canyon. That's where they decided to sleep overnight, with one of them having suffered an injured knee.</cite>

Why Gemini Failed Here

This isn't about Gemini being uniquely bad. It's about how all general-purpose LLMs struggle with specialized knowledge. <cite index="19-1">A general-purpose foundation model might perform well on everyday language tasks but hallucinate in specialized domains where its training data is thin or noisy.</cite>

Outdoor mountaineering falls into that gap. Gemini was trained on internet-scale data, which includes plenty of blog posts about hiking—but not necessarily vetted, expert-reviewed information specific to Mount Shasta's terrain, seasonality, and actual packing requirements for alpine climbing. <cite index="22-2">LLMs, primarily trained on broad, publicly accessible datasets, often lack detailed knowledge in specialized domains, leading to inaccuracies in domain-specific queries.</cite>

When Gemini didn't have reliable training data on "how much food for an 8-hour Mount Shasta climb," it did what LLMs do: <cite index="18-10">the model lacks the information needed to answer accurately and generates a plausible-sounding response anyway.</cite> The hikers got a confident-sounding answer that felt reasonable—until reality disagreed.

The Real Warning Signal

<cite index="3-2">Using AI to plan was a "critical misstep," according to the sheriff's office.</cite> But here's the uncomfortable part: for non-critical tasks (brainstorming, drafting, coding basics), Gemini performs well. Most users won't notice the difference between a good answer and a plausible-sounding wrong one—and neither will the AI.

That's why <cite index="2-7">the sheriff's office advised hikers to call the local USFS Mount Shasta Ranger station ahead of your trip to ensure you have the most accurate information, and to never rely solely on AI for your trip planning.</cite>

The cost of failure matters. <cite index="18-15">A 5 percent hallucination rate is fine for creative writing and unacceptable for medical advice.</cite> The same principle applies here: a 5 percent error rate on supply estimates for an alpine climb is unacceptable. You don't know which response is the false one until you're shivering in Mud Creek Canyon.

What This Means for Developers (and Users)

If you're building tools that use Gemini or other LLMs for high-stakes decisions, understand the domain-specificity problem. <cite index="19-2">A domain-specialized model or a foundation model adapted to a specific field can often follow domain norms more closely and is less likely to fabricate details in high-stakes areas.</cite>

For outdoor apps, fitness trackers, medical tools, or anything where wrong answers carry real consequences, you need either:

1. Domain-specific fine-tuning: Train or adapt the model on curated expert data

2. Human verification: Route outputs through domain experts before they reach users

3. Explicit refusal: Program the AI to say "I don't know" rather than guess

4. Cross-referencing: Always recommend users validate AI suggestions with official sources

If you're using NeonCodex AI or other platforms for prototyping AI features, test edge cases where the model lacks training data. Ask it to plan a mountaineering trip, diagnose a rare disease, or give legal advice. See where confidence breaks down.

The Bottom Line

This rescue is a useful reminder that AI can sound authoritative while being dangerously wrong—especially in domains where training data is sparse or noisy. <cite index="2-7">It is always advisable to call the local USFS Mount Shasta Ranger station ahead of your trip to ensure you have the most accurate information, and to never rely solely on AI for your trip planning.</cite>

That advice scales beyond hiking. For any specialized, high-stakes decision—medical, legal, financial, or otherwise—treat LLM output as a starting point, not a substitute for expert consultation.

Source: [TechCrunch](https://techcrunch.com/2026/09/05/hikers-rescued-after-using-google-gemini-for-planning/)