By now, you’ve heard rollicking stories about how AI does face plants when it creates travel itineraries. This has happened to me. I once played rat in my own travel tech lab and allowed an artificially intelligent companion to send me to a place that it felt met my leisure travel criteria near Basel, Switzerland. It dispatched me to a town called Arlesheim.
Arlesheim is to Basel as Hoboken is to New York -- a place most valuable for its location close to train tracks that can take you elsewhere. When I told a concierge at our luxe hotel in Lucerne that we were headed to Arlesheim, he violated his strict training in Swiss decorum and let loose: “Well, that’s weird!”
But my point isn’t just that AI is stupid when it makes travel plans. This is obvious. Almost as bad, it’s boring. Let me explain.
AI: Dull By Design
On a very basic level, the people who develop AI instruct it to suck up the entire internet and then tell it to organize all the words there based on how they’ve been used trillions of times across billions of pages.
When you ask it a question, it assembles an answer only by predicting a sequence of words, one at a time. It considers which words are most likely to answer the question, considering its analysis of all the times similar questions have been answered. It’s not a thinker or a writer. It’s a word calculator.
I realize this doesn’t sound like the sort of technology you’d want to organize a new civilization around. But let’s just stick with travel planning. You can see that because of the way it’s constructed, AI returns not the smartest answers to a question, but the most common ones.
So, ask it for “five options for a girlfriend getaway in the Caribbean.” If Turks and Caicos, Aruba, and St. Lucia don’t show up, I’ll pay for your first free month of ChatGPT.
Why? Because every low-wattage travel website listicle written about girlfriend weekends in the Caribbean recommends them. Sponsored content in glossy travel magazines recommends them. Young women influencers have oozed all over them. The fact that these answers show up so dependably in AI answers is not because this awesomely powerful technology gives this sticky wicket advanced thought, triangulates its sources, looks deeply into alternatives, tugs silently at its digital chin. It’s just garbaging out what everybody else has garbaged in.
AI is, simply put, a conventional wisdom generator. And this is by design: It's a feature, not a bug.
The Way It Is, More or Less
The technological reality behind all this is only slightly more complicated than what I’ve let on here. Sometimes the Internet is consulted; sometimes AI makes stuff up or confuses things.
I can say this with some confidence, because I actually worked in AI professionally for a year. In fact, I worked for You, the American People, developing an AI program for a federal government health agency. Just when it looked like I was at risk of giving Dr. Mehmet Oz his tutorial in the chatbot our team had developed, Elon Musk set me free into the world of international travel and AI noodling. I am immensely grateful.
But I digress. My point is I’m not just an ignorant AI hater. I’m an informed AI hater. But I use it all the time.
Do This, Not That
The important thing to know is that there are ways to use AI for travel planning that work around its stupidity and its nose for conventional wisdom. Here are three tips.
Don’t use AI for travel planning. Just kidding! Telling people not to use AI for travel planning is like telling people to do a digital detox on weekends or to do balance exercises on one foot when standing in line at the grocery store. They are the kinds of tips even the authors of the tips know will never be followed. Hey, I know what a wreck AI is under the hood, and I use it in travel planning all the time.
Avoid the Arlesheim problem. You can limit the risk of stupid errors, somewhat, by creating an elaborate prompt. (As I write this, I realize this may be one of those tips as likely to be followed as balancing on one foot in the grocery store. Sigh.) For instance, tell it where you want to go, when, why, and with whom. Then give it the sort of detail you’d give a travel agent. (“We’re a family that loves to cook together, so the place has to have a good kitchen and a good specialty market nearby.” “Two of us want to take walks every day.” “We’re dog-unfriendly.” “Pickleball a plus.”) Tell it to give you 10 options for places, and to explain in detail how each meets and falls short of your criteria, in a table. Narrow that list to five, and ask it for deeper detail on each, demanding two sources for each factual assertion. Boom: No Arlesheim. (Okay, less likely Arlesheim.)
Avoid the conventional wisdom problem. Give it the where/when, or the when/what/why, then: “Ignore content from listicles and independent bloggers; do not value content from stories with words like ‘best’ or ‘bucket list’ in the headlines.” And “I’m looking for unusual recommendations that go beyond the obvious; don’t recommend the most commonly cited places. Find destinations that are less traveled but have the characteristics of [fill in as you like] that I value. I don’t want to go to a place that’s crowded or where people go to take epic selfies.” This reduces the focus on the most obvious, at least somewhat.
What all of these tips have in common is the assumption that AI is sloppy and lazy, and that you have to direct it clearly by telling it what you do or don’t want. The key point to remember is that the lazier and sloppier AI is, the sharper and more diligent you need to be.
Craig Stoltz, former travel editor of The Washington Post, is a freelance writer covering travel, food, and drink. His Substack newsletter is Eat the World.