The Mountain It Matched You To.
A chatbot asked about his fitness and personality, then recommended a steep 1,500-metre peak to a man who had never really hiked. It never mentioned boots.
What happened on Tonggongjian
A man from Suzhou argued with his wife and wanted to get out of the city. He asked an AI chatbot where to go. It questioned him about his state of mind, his personality, his height and his fitness — the shape of a careful assessment — and then recommended Tonggongjian, one of the Three Peaks of Western Zhejiang, in the Lin'an district of Hangzhou.
Tonggongjian rises above 1,500 metres. Its paths are steep and undeveloped, the wilderness around them dense, and the temperature at night far below what the city gets. He had almost no outdoor experience. He went straight there.
He arrived in flat-soled skate shoes, carrying nothing but a small folding chair. On the night of 12 August 2026 he slipped and could not get himself out. Twelve people came for him — seven emergency rescue personnel, two police officers and three firefighters — in an operation that took about five hours.
What the rescuers made of it
Speaking to The Standard, the rescue team said that in many years of emergency operations, this was the first time they had dealt with someone stranded purely because they had followed an AI-generated itinerary.
Their read on the failure is the sharp part. They suspected the chatbot had simply scraped and compiled the online trip reports of experienced backpackers, without accounting for the fact that the man asking had no outdoor experience at all — and that it never prompted him to bring the equipment that terrain requires.
The first person they had found stranded solely for following an AI-generated itinerary.
What a chatbot can't know about matching you to a mountain
The questionnaire is the tell. Being asked a lot of questions feels like being assessed, but nothing in the exchange could measure the thing that mattered.
What You Can Actually Do
A questionnaire about height, fitness and personality feels like assessment. It isn't. Whether you can descend loose ground in the dark after six hours on your feet is a fact about experience, and no self-report captures it.
That the Guides Were Written by Experts
Trip reports are written by people who already own the boots, already know the descent, and leave all of that unsaid because to them it's obvious. Summarising those into advice for a beginner silently strips out the assumptions that made them safe.
What the Ground Is Like Underfoot
Steep, undeveloped, unmaintained paths are a different activity from a paved tourist route up a famous mountain. That distinction is local knowledge, rarely stated in the sources a model learned from, and decisive once you're on it.
How Cold It Gets Up There After Dark
Temperature falls with elevation and falls again at night, and the gap between a city evening and a ridge at 1,500m is larger than most first-timers expect. It's a number you can look up for tonight — if something goes and looks it up.
The inherited-expertise problem
This is the most useful idea to take from this one, and it applies far beyond one mountain in Zhejiang.
Almost everything written about a route online was written by someone who could already do it. They don't mention footwear because footwear is assumed. They don't flag the descent because they descend things like that routinely. Their competence is baked invisibly into every sentence.
Summarise a hundred of those into a recommendation and you inherit the route information while quietly discarding the expertise that made it safe. The output reads like advice for anyone. It was never advice for anyone.
A tool built for the outdoors has to close that gap deliberately — state the gear, state the terrain, and be willing to say the unsatisfying thing: this is not the right first mountain.
Already have a plan a chatbot gave you?
Paste it into Roamze's plan check and we'll check what we can against the real forecast and conditions for your route and dates, including whether the gear it told you to bring matches what those conditions call for — and we'll say plainly what we couldn't check.
It doesn't replace your own judgment, and it can't tell you whether a mountain is the right one for your experience level. For that, ask people who have been up it.
If it's your first real hike
- Pick a maintained, well-travelled trail before an undeveloped one, whatever a chatbot suggests
- Footwear with real tread is not optional on steep ground — it is the single cheapest safety item you own
- Check the overnight low at elevation, not the forecast for the nearest city
- Tell someone your route and when you expect to be back
- Go with somebody who has done it before, the first few times
Sources
Other AI planning failures
Same pattern, different mountains — a confident plan from a tool with no way to check it.
Mount Shasta, California
Gemini planned the climb and under-provisioned food and water. Three hikers spent the night stranded in Mud Creek Canyon.
Read guideTatra Mountains, Poland
A chatbot's shortcut to the Valley of the Five Lakes put two hikers on a traverse they couldn't finish. Helicopter evacuation.
Read guideUnnecessary Mountain, BC
ChatGPT and Google Maps picked the route. Nothing mentioned spring snow at elevation, and the pair went up in sneakers.
Read guideWhat Is an AI Trip Planner?
How these tools work, and what separates one that helps from one that just sounds confident.
Read guideHave a Chatbot Plan? Get a Second Opinion.
Paste the itinerary an AI gave you. Roamze checks what it can against real forecast, avalanche and snowpack data for your dates — and tells you what it couldn't check.