What an AI Trip Planner Can't Know.
Three hikers were rescued off Mount Shasta after planning the climb with a general AI chatbot. The gap between a confident answer and a grounded one is the whole story.
What happened on Mount Shasta
On the last weekend of August 2026, three novice hikers from Roseville, California camped at 8,400 feet on Mount Shasta and set out for the summit at 3 a.m. They had used Google's Gemini to choose their route and decide what to bring, and they were expecting an eight-hour climb.
It took them sixteen hours. They reached the summit at 7 p.m., started down in fading light, and got lost. One of them injured his knee. They spent the night stranded in the steep drainage of Mud Creek Canyon, and were reached the next morning by U.S. Forest Service Climbing Rangers and the Siskiyou County Sheriff's Office search and rescue team.
The Sheriff's Office was specific about where the plan went wrong. The hikers, its account said, “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.”
The sheriff's read on it
Siskiyou County Sheriff Jeremiah Larue put the problem in terms worth sitting with. AI, he said, “tends to want to give you favorable information and also make you seem like you're kind of invincible, and I think that that causes problems, especially when something can be extremely dangerous.”
AI can be a good tool, but it shouldn't be used as the expert. Go talk to actual people.
He's right, and it's worth being blunt about what that means for a company that builds AI trip planning. No planning tool — ours included — replaces talking to the people who know the mountain. The Sheriff's Office said the same thing in its own advice: call the local USFS Mount Shasta Ranger Station before your trip, and never rely solely on AI for planning. That is good advice about Roamze too.
The Sheriff's Office also gave three concrete habits that would have changed this outcome: hold to a noon turnaround rule, carry more than one way to find your way (not just a phone or a watch), and stop often enough to confirm you're still on route.
What a general chatbot structurally can't know
This isn't about one model having a bad day. A general assistant answering from training data has no connection to the four things that decide whether a mountain day goes well.
Today's Avalanche and Snowpack
Avalanche danger is published daily by regional forecast centers, and snowpack shifts with every storm. A chatbot answering from training data has no idea what today's rating is, whether a persistent slab problem is being tracked, or how much water is sitting in the snowpack right now.
The Real Clock for Your Date
Sunset moves by more than an hour across a season, and the hour you have to turn around is a function of it. A plan that says "expect about eight hours" is not the same as a plan that names the time you must be heading down, calculated from the actual sunset on the actual day you're out.
Closures, Permits and Quotas
Roads close, permits sell out, and ranger districts issue restrictions mid-season. None of that is in a language model's weights. It's in an agency's feed, and it changes week to week.
Whether the Water Is Running
A creek that's reliable in June can be dry in September. How much water you carry depends on that answer — and on the forecast temperature, the vertical, and the pace of your slowest person. It is not a generic number.
Why “favorable information” is a design problem
The sheriff's phrase is the sharpest description of the failure mode we've seen. A general assistant is tuned to be helpful and agreeable. Asked “can I do Mount Shasta in a day?”, the shape of a satisfying answer is yes, with a tidy plan attached. There is nothing in that loop that rewards saying “I don't know what the snowpack is doing this week, and that's the thing that decides this.”
A tool built for the outdoors has to be willing to give the unsatisfying answer. That means two things in practice: fetching real data instead of recalling plausible data, and having an explicit rule about what to do when the data isn't there.
Roamze's safety briefing has that rule written into it — when data is missing for a critical factor, it leans toward CAUTION rather than GO. It is not a clever prompt. It is the difference between a system that can disappoint you and one that can't.
What grounded planning looks like
Roamze's trip safety briefing reads live forecast, avalanche and snowpack data for your coordinates and your dates, then tells you what it found and where it came from.
Sourced, Not Recalled
Every conditions factor Roamze shows carries a source and an observation time. If we don't have real data for a factor, we say so rather than filling the gap with a confident sentence.
Missing Data Means Caution
Roamze's safety briefing leans toward CAUTION when data is missing for a critical factor, instead of toward the answer you'd rather hear. That rule is written into the system, not left to the model's mood.
A Turnaround Time, Not a Vibe
The briefing computes a hard turnaround time in code from the real sunset for your trip date, and shows it as a clock time you can read at a glance. It isn't written by the model, so the model can't talk itself into a later one.
Gear Checked Against Conditions
Gear gaps are flagged against the forecast, the avalanche picture and the snowpack for your dates — not against a generic packing list for a generic mountain.
Already have a plan a chatbot gave you?
You don't have to throw it out. Paste it into Roamze's plan checkand we'll run the parts of it we can check — the date, the location, the timing, the gear it told you to bring — against the real forecast, avalanche and snowpack data for that trip, and show you the differences we found.
It will not catch everything, and it is not a substitute for your own judgment or for a call to the ranger district. It is a second opinion from something that actually looked at the conditions.
The habits that matter more than any tool
- Set a turnaround time before you start, and hold to it whether or not you've reached the summit
- Carry more than one way to navigate, and at least one that isn't a phone
- Stop often enough to confirm you're still on the route you meant to be on
- Call the ranger district for the mountain you're climbing — they answer the phone, and they know this week
- Pack for the day going sideways, not for the day going to plan
Related reading
- What is an AI trip planner? — how these tools work, and what separates a useful one from a confident one
- Common trip planning mistakes — the failure modes that show up long before the summit
Sources
Other AI planning failures
Same pattern, different mountains — a confident plan from a tool with no way to check it.
Tatra 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 guideTonggongjian, China
A chatbot matched a 1,500m peak to a man with no hiking experience, and never mentioned gear. He arrived in skate shoes.
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.