A woman on Sentinel Bridge stopped me before I started talking. She had her phone in one hand, her kid's hand in the other, and she said, "I already asked ChatGPT about this trail, but I wanted to hear what you think."

That sentence contains the entire future of my profession in twelve words. She asked the machine first. She came to me second. And she still came. Why? What do I have that the machine does not? I have been thinking about that question more than I expected this season.

The machines are already here. That is not a provocation. It is a status report.

Yosemite Conservancy is funding an AI model this year to predict bear-human encounters by processing GPS collar data alongside dietary analysis. A project called BearID uses facial recognition to identify individual bears at roughly 84% accuracy. At Glacier Bay, BirdNET identifies bird calls from acoustic monitoring stations. The NPS App has surpassed six million downloads and averages 60,000 daily users. AI-driven referral traffic to travel booking sites is up 3,500% year over year. A facility management team reportedly fed a GPT years of deferred maintenance backlog and cleared it in a fraction of the time it would have taken staff. Digital passes launched on Recreation.gov in January.

This is not a future story. The future arrived while everyone was still debating whether it should.

The bear chapter is, to my mind, the best case for AI in national parks.

Bear incidents in Yosemite dropped from 1,584 in 1998 to under 35 in recent years. That is a 99% reduction, and it was accomplished by humans: food storage orders, bear boxes, education, hazing, law enforcement. But the remaining incidents are the hardest ones. The bears that still get into trouble are the edge cases, the animals that have learned workarounds, the encounters driven by drought or unusual food availability or campground configurations that nobody anticipated.

AI can process movement data, weather patterns, campground occupancy, trail use density, and mast crop yields simultaneously. It can flag a convergence of risk factors three days before a bear and a cooler full of salami end up in the same campsite. Japan's bear warning system, built on similar principles, cuts alert times by 30 minutes. Romania's camera trap network achieves near-perfect detection rates for large carnivores entering human zones.

I have watched these bears for twenty years. I have sat in meadows at dusk and watched sows teach cubs which logs to turn over. I need to be clear about what is at stake. Yosemite does not relocate problem bears. It stopped doing that years ago because relocation does not work. A bear that has been conditioned to seek human food will seek human food wherever you put it. When a bear in this park crosses the line, when it breaks into a car or a tent or becomes aggressive around people, the outcome is euthanasia. The bear is killed. That is the policy, and it is the policy because the alternatives failed.

That is why the AI matters here more than anywhere else in this conversation. Better data means fewer bears get conditioned in the first place. Fewer conditioned bears means fewer bears put down. The math is not abstract. It is an animal that is alive or is not. I am not going to argue against a tool that keeps bears alive.

A company called GuestOS is already here. The five-person AI startup, whose founder's family has owned Yosemite Resorts since 1977, has partnered with the National Park Service to bring AI-powered phone service to the public information office. If you call the park and a voice answers your question about trail conditions or campground availability, there is a growing chance that voice is not a person. It is a language model trained on park data, handling the calls that used to go to staff who are no longer there to pick up.

This is not speculation. It is happening, and it is happening in part because the phones have to be answered and there are not enough people left to answer them. The Park Service has lost 24% of its permanent workforce since January 2025. When positions go unfilled, the work does not disappear. It migrates to whatever can absorb it. Right now, that means AI. GuestOS is a small example of a large pattern, and the public information office is just the first place where visitors will hear the difference.

Fourteen million international visitors come to national parks each year, many of them non-English speakers navigating complex trail conditions and safety information in a language they do not fully command. AI translation could serve them in twenty or more languages, in real time, with context-specific accuracy that printed brochures cannot match. Disabled visitors could receive real-time accessibility updates. A first-time visitor from Osaka or Oaxaca could close the knowledge gap with a tenth-generation Fresno family that has been coming to the Valley every summer since the kids were born.

The equity case for AI in parks is strong. It might be the strongest case of all. And when you cannot hire enough staff to answer the phones in English, the case for a system that answers in twenty languages starts to look less like innovation and more like necessity.

Freeman Tilden wrote in 1957 that interpretation is not instruction but provocation. He meant that the job of a naturalist is not to fill a visitor's head with facts. It is to make them care about something they did not care about when they walked up.

AI instructs beautifully. It can tell you the name of the bird, the height of the waterfall, the geological epoch that carved the Valley, and the best time to visit to avoid crowds. It can do this in any language, with citations, with a map, with a suggested itinerary. It is, in many measurable ways, better at instruction than I am.

It cannot look at your face and decide to tell you a different story than the one it planned because something about your expression said you needed it. It cannot kneel next to a four-year-old and pick up a granite pebble and put it in her hand and say, "Feel how cold that is? It was under the glacier." It cannot hold silence at a campfire and let the fire crackle while thirty people sit with an idea that is too large for language. It cannot notice the teenager in the back row who has not looked up from his shoes and find the one sentence that makes him look up.

Interpretation is a human technology. The oldest one. I do not think it is in danger of being replaced. I think it is in danger of being defunded.

I mentioned the staffing losses earlier. One in four permanent positions gone since January 2025. That is not a technology story. It is a funding story. But the two stories are about to collide, and the collision is what keeps me up at night.

If AI fills the gaps that budget cuts created, the gaps become permanent. The seasonal naturalist position that went unfilled this year gets filled by a chatbot next year. The chatbot works fine. It costs less. It does not need housing. It does not file workers' comp claims. It does not ask for a raise. And the argument for refilling that human position erodes a little more each fiscal year until it is gone entirely.

Technology deployed in an institution that is being defunded is not neutral. It either supplements the workforce or replaces the argument for having one. Those are two different futures, and the distance between them is a policy decision that nobody seems to be making on purpose.

The parallel I keep returning to is cars. Yellowstone banned automobiles until 1915, and other parks debated them for years. Then they were allowed. Then they became the reason roads were built. Then the roads became the reason more cars came. Then the cars became the reason parks needed more traffic management, more pavement, more infrastructure, more budget. Technology adoption has momentum. It does not stop where you planned. It does not respect the boundary you drew on the whiteboard in the planning meeting. It goes where the path of least resistance takes it, and in a system under financial pressure, the path of least resistance almost always leads toward cheaper.

The Wilderness Act of 1964 uses the word "untrammeled." It means unmanipulated, not interfered with, left to its own processes. Every generation redefines what that word requires.

Every generation's technology debate in parks follows the same pattern: resist, debate, adapt, normalize. The Park Service placed a moratorium on webcams in 2000, worried they would diminish the experience of visiting in person. Today, the Old Faithful webcam is beloved. Cell towers were fought bitterly. Now visitors complain when they do not have signal.

A 2026 paper in AI & Society proposes a "Digital Wilderness Act" to establish governance frameworks for technology in protected landscapes. The Department of the Interior's Secretary's Order 3444 is already pushing AI integration across land management agencies. And the physical infrastructure of the digital world is not abstract. Data centers are encroaching on the viewsheds and watersheds of more than 70 national parks. The digital world is not just arriving in the wilderness as software. It is building next door as concrete and cooling towers.

I keep coming back to the woman on Sentinel Bridge. She asked ChatGPT first. She came to me second. And she still came.

That is the data point that matters most this season. The machine gave her information. She came to me for something else. Call it meaning. Call it connection. Call it the thing that happens when a human being who loves a place stands in that place with another human being who is seeing it for the first time. When the light does something on Half Dome that it will never do again in exactly that way, and two strangers look at each other and one of them knows what to say about it and the other one needed to hear it.

Whatever that thing is, it is not optimizable. It is not scalable. It is not efficient. And it is the entire point of having national parks in the first place.

The machines are here. They are going to stay. Some of them are going to do genuine good. The question is not whether to use them. The question is whether the people who fund the parks will remember why the parks exist, and whether they will keep paying for the thing that no machine can do.

The woman's kid tugged her hand. I started talking. The sun went down behind Half Dome. Nobody checked their phone for an hour. Not even me.