On Saturday, August 29, three young men from Roseville, California made camp at 8,400 feet on Mount Shasta (14,162 feet per USGS). At 3 a.m. Sunday they left for the summit carrying daypacks. The plan said eight hours to the top.

They summited at 7 p.m., seven hours past the noon turnaround the Mount Shasta Avalanche Center recommends for summit day. An hour into the descent, in the dark, they got lost and called sheriff’s dispatch to ask for directions. They then wandered off the Clear Creek route into the Mud Creek Canyon drainage on the mountain’s south side; one of them fell and injured a knee, and the group bivouacked in the steep drainage overnight. On Monday morning, a U.S. Forest Service climbing rangers team and sheriff’s search and rescue volunteers walked them out (timeline pieced together from CBS Sacramento and KRCR).

After the rescue, the three told deputies that for route information and packing lists, they had mostly asked Gemini. The Siskiyou County Sheriff’s Office called this a “critical misstep” in its Facebook statement: 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.” ABC News says it reached out to Google for comment; its report contains no response.

I included this in Saturday’s briefing. Here I want to take it apart, because the error wasn’t in any single fact. The whole estimate pointed in the wrong direction.

Diagnosis: how far off is eight hours

The Mount Shasta Avalanche Center (run by the USFS Shasta-Trinity National Forest) describes the Clear Creek route this way: “Those of strong mind and body can do the Clear Creek route in a day, but most should plan a delightful two, or even three day trip.” In summer and fall the surface is loose ash, scree, and gravel, and there is spring water near the camp at 8,600 feet.

In other words, these novices (ABC News’s word) received a plan that assumed the single-day pace the official description reserves for those of “strong mind and body.” Once the plan is wrong, everything pegged to it fails with it: food and water packed for 8 hours, daylight budgeted for 8 hours, a turnaround time reduced to decoration. The sheriff’s statement puts its finger on it with “especially when their planned 8 hour ascent became a multi-day ordeal.” The supplies may have been a fair answer to the 8-hour question. They were no answer to the actual mountain.

This is not an isolated case. In July, Lions Bay Search and Rescue in British Columbia rescued a group that had planned the Howe Sound Crest Trail with Google Maps, then posted a warning on its own site and described the case to The Narwhal: Maps quoted a walking time of 5 hours 13 minutes for a route that actually takes 8 to 14 hours, with well over 1,300 meters of elevation gain and sustained scrambling. Google’s response was that walking navigation on Maps is designed for streets and standard walking speeds, and that hikers should use maps built for hiking. Search manager Maria Masiar was blunter about AI planning tools: “We have this case where AI has given false information, false directions. It’s pulling from unknown sources, and those sources aren’t always correct.”

Mechanism: why the model is most optimistic where it should be most conservative

One boundary first: what exactly the three asked Gemini, and what Gemini answered verbatim, does not appear in any of the public reporting I checked, so I could not verify it. Another KRCR report notes they also consulted YouTube videos and AllTrails. “Advised to bring far less food and water” is the hikers’ own account to deputies. But however well or badly their prompts were written, the case points to the same set of structural problems.

First, the model’s clarifying questions are not a gate. You might object that when you use these tools, they do ask follow-ups. True. In my own use, current Gemini and ChatGPT will often open a planning question by asking about dates, party size, destination. So the title deserves a precise restatement: the ranger’s questions are procedure, the model’s questions are probability. A ranger has to know how many of you there are, what you have climbed before, and when you are going before there is any targeted advice to give; the questions come first, and they force you to face variables you had not thought about. On the model’s side, nothing is guaranteed. In my experience, whether it asks depends on your wording and the model version, sometimes even on the run; and if it asks and you answer vaguely, it usually answers anyway. I have not tested Gemini on this route, and I know of no mechanism that guarantees it will withhold an answer because key information is missing. In early August I wrote about the MIT Sloan study on AI financial advice, which found the same pattern: the advice was only as good as the questions people asked. Vague prompts with no context got simplistic rule-of-thumb advice; prompts that spelled out the asker’s circumstances got measurably better plans. The variables that mattered most here (that day’s conditions on the mountain, whether late-season scree or leftover snow; this particular group’s fitness; and altitude acclimatization) are precisely the ones novices do not know to volunteer. Adjusting to thin air takes days: the CDC’s travel medicine guidance puts acute acclimatization in the first 3–5 days after ascent and advises against going from low elevation to a sleeping altitude above 9,000 feet in a single day. The three slept at 8,400 feet their first night and pushed for 14,000 the next morning, exactly the pattern that guidance warns against; per KRCR, their actual climb took about 16 hours. No medical assessment appears in any of the reporting I have seen, so I cannot confirm acclimatization is what turned 8 hours into 16, but of all these variables it is the one a novice is most likely to leave out. The user does not volunteer the information, the model does not insist on it, and an answer comes out regardless.

Second, the times posted online come from people who are not you. The people who write up their climbing times are probably skewed toward those who climb often and climb fast (my inference; I have no statistics), and trip reports that say “doable in a day” assume readers of the same kind. A model that learns its numbers from texts like these does not know which kind of climber is asking. How Gemini arrived at 8 hours cannot be replayed. What can be checked is that the official description is itself a distribution: strong climbers one day, most people two or three. Compress a distribution into a single number and the part most easily dropped is the “most people” half. Google Maps’ 5 hours 13 minutes is a purer specimen of the same error: by Google’s own account, walking ETAs assume standard street walking speeds, which have nothing to say about 1,300 meters of climbing and hands-on scrambling.

I ran into the model-doesn’t-know-who-you-are problem myself this year, at far lower stakes. In late May I spent a week in Alaska. The trip came together on a week’s notice, so I had ChatGPT lay out two itineraries, one with a rental car and one without. The without-car version looked right, I locked it in within two hours, and my travel companion and I booked flights, hotels, and the train. Only after booking did I notice the recommended lodging ran to lodges: pricier and nicer than the budget hotels I normally book, a mismatch with how I actually travel. Too lazy to redo the itinerary, I kept the bookings. The model’s default plan fits some typical user, not necessarily the person asking. My mismatch stopped at a hotel bill. The Shasta three’s mismatch walked into a two-day rescue.

Third, and most important: safety planning does not want an accurate estimate. It wants an estimate with margin. Carrying two extra liters of water costs you a heavier pack; carrying two liters too few can cost a life. When the losses are that asymmetric, the correct move is to provision for the bad case. Ranger-station advice always sounds fussy and conservative because search and rescue has seen every way the plan fails, and the sheriff’s office putting out a public statement after this rescue is that feedback loop doing its work. I know of no comparable loop on the model’s side, no path by which an accident on this mountain flows back to correct its estimate of the route. In my experience it defaults to a middle-of-the-road, confident answer and adds no safety margin unless you explicitly ask for one. For booking restaurants and sequencing a city itinerary, a middle-of-the-road answer is what makes the tool useful. On a 14,000-foot volcano, the same behavior is what makes it dangerous.

What to do: take over the asking yourself

Sheriff Jeremiah LaRue’s version, in an interview with local TV, is milder than the written statement: “I think that AI can be used as a research tool, but when you’re going to go put yourself in a potentially life or death scenario, just make sure that you’re getting all of the information that you can from the experts.” I agree with where he draws the line, and I would add four specifics:

  1. Feed it every variable yourself. Dates, route, party size, each person’s experience and fitness, any history at altitude, sunrise and sunset. The model may ask, or may not; you cannot bet on it asking everything for you. This is the same conclusion as the financial-advice piece.
  2. Demand the bad case. Ask “if we move at half the guidebook pace, how does the supply list change” and “list the assumptions behind this estimate.” In my experience it does not volunteer that list of assumptions; you have to ask.
  3. Hard rules do not come from a chat. A noon turnaround exists precisely to make the decision for you at the moment you are most depleted and most tempted to gamble. No chat output should override it.
  4. Check the numbers against the authoritative source. Times, water sources, current conditions: call the ranger station or read the managing agency’s page. Treat AI output as a draft. The point of the draft is to send you into that phone call with better questions, not to spare you the call.

Beyond those four, there is a plainer point: planning with AI is no different from planning with any other tool. Cross-check, adapt on the ground, never follow it wholesale. Turn-by-turn navigation has been mainstream for many years and the low-comedy failures never stopped. In 2016 an Ontario driver followed GPS down a boat launch and straight into Tobermory’s harbour, rolled down the window and swam to shore. This June, a Seattle driver drove an SUV onto the elevated light rail tracks at Mount Baker station and continued along the guideway, shutting down light rail service for hours that evening; per KOMO, she told police she had been following GPS, and officers described her as confused. What the navigation actually instructed in each case, the reports do not say. What can be established is that the thing ahead was visibly water, visibly rail track, and the final push on the pedal was human. AI is the same: it can produce a plan that looks complete, and the judgment at the decisive moment still has to be yours.

One last judgment. This accident and the AI financial-advice failures come from the same combination: key information missing, a model that answers without insisting the information be complete, asymmetric losses, and a confident output all the same. In personal finance that combination produces chronic losses; on a mountain, acute ones. Recognizing the combination is worth more than memorizing “don’t plan hikes with AI.” The next place it shows up probably won’t be on a mountain.

References