About 58 percent of U.S. travelers now use AI to plan their trips, and over 60 percent of Millenial and Gen Z travelers have done so, according to a recent report by the travel research firm Phocuswright.
Yet the chatbots that produce travel plans—ranging from hotel and museum recommendations to detailed trip itineraries—have been found in at least one study to be error-free only 10 percent of the time, says Jungho Suh, a management professor at George Washington University School of Business, who closely follows the use of AI in travel. While he points out that the technology has improved since the study he cites was done in 2024, chatbots regularly make mistakes, he says, because they “draw on a fixed body of data that can be outdated and incomplete.”
Chatbots are essentially language simulators trained on gargantuan amounts of text; they string together words based only on which one is most likely to come next.
They can’t distinguish between something that is true or false, may not have been exposed to information relevant to your question, and are programmed to give authoritative-sounding responses to all queries whether they have the answers or not. If material is missing or vague, they make things up.
So while chatbots make trip planning infinitely easier by providing quick and seamless travel recommendations based on user queries, they inevitably make mistakes. But if you know what those errors are and how to deal with them, you can use AI to help plan safe and happy travels. Here’s a quick guide to chatbots’ most common mistakes and how to work around them.
Geography fails
When you ask a question that requires a chatbot to know precisely where something is, it usually draws only on words that have been written about a location, not on map data.
The programs that power AI “generate text based on something they may have seen on Reddit or wherever, but they don’t have access to hard logistical data,” explains Skyler Erickson, founder of Gondola, which uses AI to help travelers use their loyalty points to book hotels. “That really limits their usefulness” for planning tasks that require an awareness of distances and locations.
For example, I recently asked the chatbot Perplexity whether I could walk from Prince’s Paisley Park home and studio near Minneapolis to the (wonderfully named, by the way) restaurant Tequila Butcher. Since that specific question has probably never been answered in digital content, Perplexity fudged, saying only that they were “very close” and the walk was “short.” I checked with Google Maps and the walk is 38 minutes along a busy highway.
Chatbots can misunderstand the physical world in other alarming ways. Catherine Jackson Moynihan, an executive based in Charlottesville, Virginia, used ChatGPT to plan a campervan trip from the Outer Banks of North Carolina to Minnesota’s Boundary Waters, specifying less than six hours on the road each day. It did a decent job until it got to Lake Michigan, she says, which it recommended she simply drive straight across.
Later, she asked Google’s Gemini, which integrates Google Maps, to make the same plan. It successfully routed her around the lake—but flubbed by recommending a day of driving that was too long, a different type of distance failure.
Workaround
Check all distances and locations against Google Maps or Apple Maps. Google’s Gemini chatbot (and its AI Mode in search) integrate Google Maps, so they may be a better, but still imperfect, starting places for searches involving distances and geography.
Dated information
A chatbot doesn’t know anything that happened after the day its training stopped. For anything beyond that, it depends on information pulled from search engines. But this doesn’t mean it delivers the most current information in its answers.
Jake Peters, co-founder and chief product and technology officer of Fora Travel, a network of travel advisors, cites the restaurant Golden Poppy at Hôtel La Fantaisie in Paris, which ChatGPT confidently says is helmed by chef Dominique Crenn (whose restaurant Atelier Crenn in San Francisco earned three Michelin stars). Yet the Golden Poppy restaurant closed in June 2024.
“It was written about a lot when it opened [in 2023], and the AI was trained on all that information,” Peters says. “Not much has been written about the hotel restaurant” that replaced it. The result: ChatGPT reports the old news, with no idea it’s no longer accurate.
Workaround
Always verify information on restaurants, attractions, activities, hotels … essentially everything, against the primary website of the place itself. Then call.
Vibe blindness
Peters regularly tests new AI models by asking them to identify family-friendly restaurants in a Paris neighborhood he knows well. “I know these places, and I have kids, and some of them that show up on the lists are just not right” for children, he says. “Somehow [the chatbot] picks up a signal that tells it a place is kid-friendly, or it’s just making it up. It doesn’t have the first-hand, on-the-ground experience that a human does.”
Fora is developing a system that leverages AI and its advisors’ collective knowledge to overcome this problem—using the “hybrid” machine-and-human approach most AI experts recommend.
Workaround
Sometimes words that make perfect sense to you don’t direct a chatbot successfully. When you get a failed response, revise your prompt.
“Try something other than ‘family-friendly’ and see if something different comes back,” says Matt Artz, an AI product manager for the federal government and an anthropologist who studies human/AI interactions. “Tell it you have a 2-year-old and a 4-year-old,” or try “restaurants that treat children well.”
And the hybrid approach—doing a vibe check of AI recommendations with a real human—is always a great idea if you know a person with first-hand experience.
Recommending overtouristed locations
This isn’t an error as much as it is a consistent flaw you have to manage. Because the models predict words based largely on their prevalence in the content they’ve been exposed to, Artz explains, their answers tend to represent a sort of consensus of the accumulated writing on a topic.
In other words, unless nudged otherwise, chatbots are conventional wisdom generators. The chatbots have digested so many “best of” lists that they’ll suggest the most obvious, overvisited spots, which is probably not what you’re looking for.
Workaround
You may find more of what you want by asking for places and activities where “the locals go,” or where “to avoid crowds.”
Better yet, specify your own characteristics and preferences to get more personalized recommendations: i.e, “I’m a 35-year-old woman traveling solo, I like global cuisine, I want to stay downtown, and I need a list of places I can walk every day for exercise. I don’t like shopping.”
The bottom line
One of the best ways to minimize many flavors of chatbot failure, Artz says, is to let the chatbot write the prompt for you.
For example, type this into the chat: “Write a highly detailed prompt to generate a personalized travel itinerary for a [destination, how many days, kind of trip, what you like to do]. Ask me for specific pieces of information you need in order to create a detailed, personalized travel itinerary that meets my needs for my style of travel and budget. After I provide all those answers, write a prompt that will produce an itinerary for my trip to [destination].”
Answer the questions the chatbot asks, and when it produces the final prompt, tell it to use it to build an itinerary. Check it out. “Chances are, it will do a better job than you will” at crafting the trip plan you’re hoping for, and with fewer errors, says Artz. “Let the machine do what the machine does well.”
