AI 不会替你爬山AI Won't Climb the Mountain for You
三个二十岁的人在沙斯塔山上把脑子关掉了,这是救援负责人 Nick Meyers 的原话。他们没带够食物,高估了自己的体力,下午没登顶也不回头,晚上七点半才到顶,摸黑下撤,第二天早上迷路,一个人磕伤了膝盖。事后他们承认,计划有一部分是 Gemini 给的。
这不是 AI 骗人。这是人主动把解释权让渡出去了,还让得心甘情愿。
Flathmann 给这个现象起了名字:cognitive surrender。你问一个聊天机器人,它答得又快又笃定,你没法反驳,因为你不懂。你越不懂,它越像专家;它越像专家,你越不问。等到山里天黑风大,你已经没有能力把判断力重新拿回来了。这不是技术问题,是主体性的问题——第一章讲的表达,就是你如何判断、如何决定、如何在场;你把这一整套交给一个语言模型,等于把你的存在性外包了。
AI 的假最优解表达,是扮演一个全知的向导。它不告诉你它不知道,它把 Reddit 帖子、过期 YouTube 视频、甚至编出来的东西,用权威的语气端给你。你听了很舒服,因为你想要的就是有人告诉你“你可以的”。它给你的是鼓励,不是知识。这和浪漫爱叙事、和偶像产业是同一个入口:它喂给你你想听的话,然后你把它内化成自己的判断。区别只在于,这次内化的是登山路线。
但要说清楚:这不是新问题。十年前是 Instagram,再往前是 Google Maps 和手机信号。Jim Moss 说得对,过去二十年媒体一直在推人出门,出门的人多了,出事的人自然多。Meyers 的办公室一年平均救十二次,2026 年看起来是个正常年份;怀俄明的救援呼叫 2026 年到目前为止是 312 次,比 2025 年的 466 次和 2024 年的 377 次都少。数字没有爆炸。爆炸的是人把判断交出去的速度。
那个在“禁止露营”牌子旁边扎帐篷的女人,坚持说 AI 告诉她可以。那个拿着 AI 生成地图在登山口转悠的男人,地图上明显的地标标在了山的另一边。不列颠哥伦比亚那个十六岁的少年,被直升机从 Crown Mountain 的岩架上救下来,新闻全在说 Claude,他后来说他也查了 Google Maps 和 AllTrails。这些故事里没有受害者,只有一群把“我想去”当成了“我能去”的人。
山不读你的聊天记录。它只认你带了多少食物、几点掉头、膝盖还能不能弯。
真正该问的不是 AI 会不会害死人,而是为什么这么多人愿意让一个机器人替自己思考。答案不复杂:因为自己思考太累了,因为承认“我不知道”太难受了,因为一个笃定的声音比一片沉默的山容易对付。AI 只是把这个缺口照得更亮了一点。
Meyers 的建议很朴素:查权威地图,查步道协会的网站,找向导,找护林员,实在不行问一个真人。这话放在任何领域都成立。你把判断力交给谁,谁就替你活。山只是把这件事的代价,标得比较清楚而已。
Three twenty-year-olds on Mount Shasta turned their brains off. That's the rescue chief's phrase, not mine. They didn't bring enough food, overestimated their fitness, didn't turn back in the late afternoon, summited at 7:30 p.m., descended in the dark, got lost by morning, and one of them banged up a knee. They later admitted part of their planning came from Gemini.
This isn't AI lying. This is people handing over the interpretative authority willingly, and handing it over gladly.
Flathmann calls it cognitive surrender. You ask a chatbot, it answers fast and certain, and you can't call BS because you don't have the expertise. The less you know, the more it sounds like an expert; the more it sounds like an expert, the less you ask. By the time it's dark and cold on the mountain, you've lost the capacity to take your judgment back. This isn't a tech problem. It's a problem of subjectivity. Expression, as I've argued, is how you judge, decide, and be present. Hand that whole set to a language model and you've outsourced your existence.
AI's fake optimal expression is to play the all-knowing guide. It doesn't tell you what it doesn't know. It takes Reddit threads, out-of-date YouTube videos, and outright fabrications, and serves them in a voice of authority. You like it, because what you wanted was someone to tell you that you can. It gives you encouragement, not knowledge. Same entry point as the romance narrative, same as the idol industry: it feeds you what you want to hear, and you internalize it as your own judgment. The only difference is that this time what gets internalized is a hiking route.
But let's be clear: this isn't new. A decade ago it was Instagram. Before that it was Google Maps and cell signal. Jim Moss is right that two decades of media pushing people outdoors means more people outdoors, and more people getting into trouble. Meyers's office averages a dozen rescues a year, and 2026 looks average. Search and rescue calls in Wyoming are actually down: 312 so far in 2026, compared to 466 in 2025 and 377 in 2024. The numbers haven't exploded. What's exploded is the speed at which people hand over their judgment.
The woman who pitched her tent next to a No Camping sign insisted AI told her she could. The man wandering a trailhead with an AI-generated map that put prominent features on the wrong side of the mountain. The sixteen-year-old in British Columbia helicoptered off a ledge on Crown Mountain, with news stories all about Claude, though he later said he also checked Google Maps and AllTrails. There are no victims in these stories, only people who mistook I want to go for I can go.
The mountain doesn't read your chat history. It only knows how much food you brought, when you turned around, and whether your knee still bends.
The real question isn't whether AI will kill people. It's why so many people are willing to let a robot think for them. The answer isn't complicated: thinking is tiring, admitting I don't know is unpleasant, and a confident voice is easier to deal with than a silent mountain. AI just shines a brighter light on that gap.
Meyers's advice is plain: check authoritative maps, trail association websites, talk to a guide, talk to a ranger, talk to a real human. That advice holds in any field. Whoever you hand your judgment to, that's who lives your life for you. The mountain just makes the price tag easier to read.