The arithmetic of the Shasta rescue points away from a model error. Nick Meyers, the lead climbing ranger there, told Business Insider that the roughly eight-hour estimate the party worked from was not generally wrong [11]. They had carried food and water for an eight-hour day and were still climbing at hour sixteen, twice the duration they had provisioned for, before the descent had begun [4][1]. The summit came seven hours after the turnaround time Meyers puts at around noon [2]. From the 3 a.m. start to the moment rangers reached them, the party was in the field for about thirty and a half hours [3]. A defensible average told them nothing about their own pace, and nothing in the plan was built to notice the gap opening.
The camper and the mapmaker are different defects, and separating them changes what a reader should take from the season. In two of the three encounters, the tool asserted something false about a physical place: a permission claim contradicted by posted signage [13], and a map that put features on the wrong sides of the mountain, which Meyers described as terrible [14]. The rescue is the inverse case, where the number was serviceable and the process around it was absent [11]. One of those calls for better grounding; the other calls for a planning step no model output can supply.
Novices have been under-provisioning and pushing past turnaround times for as long as there have been mountains, and Meyers says something close to this himself, calling AI not the nail in the coffin and describing a series of mistakes that added up [10]. The record cannot isolate AI's contribution from ordinary inexperience, but it does point to something narrower: the group's own account of itself, relayed by the Forest Service, was that it relied too much on AI rather than its own critical thinking [9], and a camper offered a chatbot as authority against a sign [13]. Once a tool becomes the justification a person gives an official, it has already displaced the check.
This is one ranger's tally, on one mountain, in one season [2], with no total incident count to set it against and no figure for what the response cost the Forest Service [19] -- not enough evidence to support a rate. Three cases describe the shape of a failure, and the shape is the part that transfers; the trend line is not yet available.
Both vendors publish the right instruction. Google's site says Gemini can make mistakes and hallucinate and that users should double-check its responses [15]. OpenAI's says ChatGPT can respond confidently even when it is wrong, and advises using it as a first draft rather than a final source [16]. Those cautions land at the moment of the query, when the cost of skipping the check is nothing, and the cost itself arrives at hour sixteen with a dead phone and no spare food [6]. Meyers's own remedy is a phone call to a person, using the updated Shasta website and a ranger on the line [17][18]. For anyone deciding this quarter where an assistant sits in front of users who cannot audit its output, the lesson from Shasta is that a plausible answer can quietly remove the step where someone asks what happens if the number is off by half, and that step belongs in the process rather than in a disclaimer.