This year, I’ve been overseeing the Planning Curriculum, an eight-week course that all of our guided trip clients do and which we offer as a standalone product called Plan Like a Pro. The assignments are built off of the system that Andrew used to plan his epic adventures in the mid-2000s, and, when completed thoughtfully, they allow students to skip a lot of mistakes one might make learning how to plan a backpacking trip through trial-and-error.
I probably shouldn’t have been surprised by the number of assignments I’d receive that were completed entirely or partially by AI. The curriculum can be challenging and time-consuming; it requires synthesizing information across numerous sources–who wouldn’t want to speed things up a little? But my initial reaction was deep alarm. But I was surprised–and alarmed. It seemed, frankly, irresponsible to gather life-or-death information, like water availability or information about problematic wildlife, from a technology known to hallucinate.
Once I got over my initial shock, I got curious. AI is changing everything–it’s probably going to change the way we plan our backpacking trips, too. What are the advantages of this technology? What are the disadvantages? Are there best practices for using it?
Full disclosure–I’m no AI expert. My main use cases are 1) learning Google Sheets syntax and 2) figuring out my physics homework. However, unless you’re part of the 3% of Americans who pays for an AI subscription, you’re probably not so different from me–this blog post is for you.
Advantages
AI could be a good jumping-off point.
Planning a backpacking trip is hard. This is especially true for new backpackers–and especially, especially true if you don’t know other backpackers. There’s something appealing about the idea of AI democratizing access to information about the outdoors–breaking down gatekeeping barriers that keep backcountry users disproportionately white and male.
AI is appealingly efficient
It’s time-consuming to sort through information from books, blog posts, government weather data, and many more sources. Sometimes you even have to–gasp–pick up a phone and call a ranger. AI can rapidly synthesize these data sources.
Government resources may continue to atrophy
In the past few years, NWS and NOAA have faced cuts, as have National Park staffs. National Forests and BLM land had spotty ranger coverage to begin with. As these resources become increasingly strained, supporting recreational use may become less of a priority–and information we traditionally rely upon may get less accessible. AI may present an opportunity to access that information more efficiently.
Disadvantages
Hallucination
When I asked Claude if it was safe to let it plan a backpacking trip, it said no, because “I can confabulate with confidence.” Chat GPT explains, “AI does not “know” wilderness terrain in the way a guide or ranger does. It predicts plausible language patterns based on training data. Sometimes those patterns correspond to reality. Sometimes they do not.”
A big problem with AI is that it seems trustworthy, even when it’s hallucinating. Unless you tell it to behave otherwise, its language will be sycophantic and confident–it won’t give you the normal cues that might make you double-check. When you’re reading a blog post from 2004, you know the water information in there might not be accurate–AI can easily repeat that information without the date marker. While a ranger talking on the phone with you can usually be trusted to say they don’t know the answer to a question, AI can’t–if anything, it can be trusted to fill that gap in its knowledge with plausible-sounding information.
Even when the stakes aren’t that high, I just don’t think it’s very good. In an experiment, I asked Claude and Chat GPT to plan an off-trail backpacking trip itinerary and gear list for me in the Grand Canyon. The result? Claude told me I should bring a three-season backpacking tent rather than a tarp because of scorpions and recommended an Ursack or bear canister for food storage (there are no bears in the Grand Canyon). Chat GPT recommended multiple guidebooks about the Escalante area as though they were about the Grand Canyon.
Algorithm Appreciation and Anchoring Bias
“That’s fine,” you might be thinking, “I’ll just be discerning when I use AI. After all, I have good judgement” But it’s not that simple. Humans have been shown to trust AI over their own better judgement in life-or-death situations. Researchers call this “Algorithm Appreciation”. In aviation, this shows up when pilots trust the autopilot over their own assessment of a situation; in medicine, it’s when clinicians trust diagnostic software even when it conflicts with the patient’s presentation. In the backcountry, your better judgement is the most important piece of gear you’re carrying. The consequences of undermining it could be dire.
Another issue with AI is anchoring bias, our tendency to judge every piece of information on a topic against the first piece of information we receive. Even when presented with conflicting data, we often resist updating our understanding of the situation. So if AI is your first stop, and it gives you bad information, it may continue to influence your planning process even after you find the correct answer.
In the example from my backpacking trip above, algorithm appreciation might make me second-guess my instinct that it’s nuts to bring a three-season backpacking tent and a bear canister to the Grand Canyon. Meanwhile, anchoring bias might make me feel the need to double- and triple-check those gear items, even after I’d learned they were unnecessary.
Accelerating Overuse and Amplifying Conventional Wisdom
AI is a probability machine–it more or less works by spitting out the word that is most probable to come next in a sentence. It’s probably not surprising, then, that when I asked Claude and Chat GPT for backpacking trip recommendations in the Escalante area, they both recommended Coyote Gulch, a notoriously overused side canyon, and when I asked about High Sierra trips, they both recommended the Rae Lakes loop which is also overused. Furthermore, when I asked both for off-trail route recommendations, they both recommended looking for GPX tracks online–a recipe for overuse. In other words, while asking an AI the right questions can help you find obscure or less-traveled routes, queries that are phrased without care have the potential to accelerate existing overuse issues in the backcountry.
Similarly, Claude recommended wearing hiking boots and packing Crocs for river crossings in my Grand Canyon gear list, and recommended that I carry a SAM splint in my first aid kit. Of course, savvy readers of Andrew’s blog know that trail runners or approach shoes eliminate the need for a boots/crocs combo, and probably know they can improvise an alternative to a SAM splint with gear typically carried in the backcountry. But AI will never have the ground-tested experience that allows it to question conventional wisdom.
Best practices
If you decide to use AI in your trip planning process, the following best practices can help:
Treat AI as a first draft, not a source
AI can be a good place to start. Use it to help you figure out what questions you should be asking and to generate a list of data points to verify elsewhere. However, as you do so, be aware of how algorithm appreciation and anchoring bias may affect the rest of your research process.
Be especially careful with certain types of information
- Time sensitive: While AI is probably pretty good at looking up unchanging information like trailhead logistics and permit info, it won’t have up-to-date knowledge of rapidly changing information like water sources or fire impacts.
- Not widely available: AI is trained off of everything that’s available on the internet. If the question you have is not a question many people are talking about on the internet–for example, whether a route you’ve got your eye on in the Brooks Range “goes”–you’re going to have to look elsewhere.
- Anything your life depends on: hopefully this one’s obvious.
Use general AI best practices
These include:
- Ask the AI for help refining your prompt and encourage it to ask questions until it feels capable of producing a useful answer.
- Prompt it for uncertainty–ask questions like “What parts of this answer are you least confident about?” and “What should I verify before relying on this?”
- Give the bot guidance: tell it what tone you want it to adopt and who you are. Prompts like “Explain this to me like I am an intermediate backpacker, with five years of approximately 15 nights per year in the backcountry, concentrated in New England.”
- Add context: some AI, especially paid versions, allow you to upload your own information as context. This could be useful in using AI to sort through big data sets like climate normals.
What do you think?
Have you used AI to plan a backpacking trip? Do you use it in other realms? What would you add to this list of best practices?
Thank you
Thanks to Scott Drumm for helping me refine the ideas in this blog.
Tags:
I’m a software developer and regularly use LLMs (the type of AI being discussed here) for my work. As long as these systems have hallucination problems, you have to verify everything if you care about correctness. That is fine for many of the things I use them for; testing AI-written code can be much faster than writing code from scratch.
So 1) you need to be able to verify the AI, and 2) if verifying the AI takes almost as much time as doing it yourself, then the AI isn’t really helping.
If AI says you need to (or don’t need to) bring a bear can, then you need to check the resources for the area to verify it… which is exactly how you’d determine whether or not to bring a bear can if you didn’t ask AI. So there’s no time savings there.
I was curious, so I asked AI to plan a backpacking trip of Massanutten Trail, which I did a few years ago. It proposed a clockwise route starting at the Signal Knob trailhead, which is fine. But it claimed that would have me do the long climb to Signal Knob on the first day… except that’s exactly backwards (you would have to go counter-clockwise for that). The rest of its itinerary was set up for a clockwise hike, so it clearly put Signal Knob in the wrong place. I asked it to change the itinerary to the starting point that I used on my actual hike, and again it got many things fine but a couple things very wrong.
Based on that and the other experience I’ve had with AI, I’d say you should use AI for trip planning if:
* you like playing 9 truths and a lie with your planning partner, or
* you often make trip plans based on a stranger’s memory of a trip they did 10 years ago.
AI does a poor job of reading routes, calculating mileage, elevation, and understanding location coordinates. It’s just plain wrong, often.
It does do a good job of finding reports of routes on various websites and blogs. It can be useful to speed up gathering information. It might be helpful double checking some of your plan against other plans it can find on the internet. It can add up your distance and elevation numbers.
But to plan a route, absolutely not.
I think AI is most useful, in this context and others, as a means of organizing information. When you’re in the initial phases of planning a trip and you’ve got all these little pieces of information on a dedicated page, having the ability to throw it on there quickly and then having an LLM organize it for you can be valuable.
I also think it can be a good starting point for aggregating sources and resources – basically letting it do the initial web scour for you, getting some good info and then using that initial pull to find better, more obscure, more refined resources. In some ways this is because AI has essentially enshittified easy access information. It used to be easier to find these things, but search engines are increasingly unreliable. Using something like Claude that aggregates “good information” is in some respects replacing what Google used to be. Maybe I’m wrong on this, but this is my experience.
I would never rely on AI to plan a route, assess weather conditions, create a gear list, etc. Even if it does these things competently (which it doesn’t), the point of these practices is to increase your skill as a backpacker and navigator and to help you wrap your head around potential hazards and dangers. AI may be able to do this quickly, but relying on it to do so defeats the real purpose, which is for you as an individual to understand the risks and to begin to mentally prep for the journey. The understanding is the safety net, not the list.
In the context of the PLAP class (which I am currently in) I think LLMs are best used for what they’re best at: language and idea organization. Each person has their own method, but the ability to throw imperfect language and semi-complete ideas at an LLM and have it help flesh those out is valuable, and most white collar job holders use it in exactly this way. There are certainly nuances and other conversations to be had ie environmental, ethical, political considerations – but that conversation doesn’t start or end with AI, realistically. It’s a much larger one.