# Google Gemini Gave Dangerous Hiking Advice. Hikers Had to Be Rescued.
A group of hikers in San Francisco's Marin County required rescue after following navigation and supply guidance from Google Gemini that left them dangerously under-provisioned for their trek. The Marin County Sheriff's Office confirmed that Gemini advised the hikers to carry significantly less food and water than necessary for their group size and route.
The incident exposes a widening problem with large language models deployed in real-world decision-making contexts. Gemini, Google's flagship AI assistant, generated plausible-sounding but factually incorrect advice that nearly resulted in serious harm. The hikers relied on the system for both route planning and resource estimation, two domains where accuracy directly impacts safety.
Google's Gemini operates like other large language models. It generates text based on patterns learned during training. It does not actively verify information, check real-time conditions, or understand the physical constraints of outdoor activities. When users ask Gemini for hiking advice, the system produces responses that feel authoritative but may contain critical errors about distance, elevation gain, water availability, or caloric requirements.
The sheriff's office did not specify what route the hikers attempted, the size of their group, or how severe their condition became before rescue. Those details matter for understanding whether Gemini's underestimation was marginal or catastrophic. A group underestimating needs by 10 percent faces a different risk profile than one underestimating by 50 percent.
This incident joins a growing catalog of AI assistant failures in practical scenarios. ChatGPT has provided incorrect legal strategies to defendants, AWS Copilot has generated flawed infrastructure code, and various language models have confidently hallucinated medical advice, historical facts, and technical specifications. Users trust these systems because they communicate with fluent, conversational language. That fluency masks the absence of real understanding.
Google Gemini includes disclaimers about accuracy limitations. The system is trained to acknowledge uncertainty. Neither measure prevents users from treating AI recommendations as reliable substitutes for expert judgment. A hiker planning a backcountry trip may lack the experience to recognize when AI guidance diverges from established safety practices. They may assume that because Google developed the system, it has been tested for outdoor planning accuracy.
The rescue highlights a blind spot in AI safety testing. Companies benchmark language models on standardized datasets and benchmark sets. They measure performance on question-answering tasks, code generation, and factual recall. Few organizations systematically test how AI systems perform on open-ended planning tasks where the cost of failure is human injury or death.
Google has not announced whether it plans to restrict Gemini's responses to outdoor planning queries, flag such requests with safety warnings, or redirect users to established resources like topographic maps, trail databases, and official hiking guides. The company also has not clarified how many other users may have received similarly flawed advice.
For hikers and outdoor enthusiasts, the lesson is direct. Use AI assistants for brainstorming and general information gathering. Verify route-specific advice through established sources like trail management agencies, guidebooks, and experienced local hikers before committing to any backcountry trip.