Confidently Wrong: The Trouble With AI Trip Planning

(7 minute read)

Over the past couple of years, two friends have shown me itineraries an AI built for them. The first was for a road trip, and it read beautifully. Every day had a shape, every stop had a sentence about why it deserved your time, and the morning's stop was four hours' drive from the afternoon's, a distance the plan crossed in a single line break, as if the drive were a formality.

The second had my friend doing the Louvre in an afternoon. You can give the Louvre only an afternoon, and plenty of people do, but not while seeing everything he'd told the tool he wanted to see. It didn't mention that. It handed him an afternoon and let him find out for himself.

He isn't unusual. Katie Collins of CNET asked for a Scotland itinerary and got one packed with good activities that had her driving hours back to Edinburgh for every meal. It's plausible on the page and absurd on the ground, and it's being handed to a lot of people. Deloitte's 2026 travel research found AI use for travel rose from 15 to 25 percent in a year. In August, the travel concierge company Greetwell surveyed 1,000 American adults, 485 of whom had used or tried AI for travel. More than half of those had hit at least one bad recommendation, and about one in six had been recommended a place, tour, or activity that doesn't exist.

We aren't going to argue that a machine can't know what we know. Our argument is narrower, and harder to wave away: a model produces a plan, but it doesn't verify one, and nobody answers for it when it's wrong.

In Fairness, We Tried It

We used these tools for research in the past, assuming they'd save us time, and mostly they gave us homework. The tool we used recommended a restaurant that had been closed for more than two years, and when we asked what to do in one particular village, it suggested something nowhere near that village. We stopped, and not because we're precious about the craft. The errors cost more time than the tool saved.

They're fine for brainstorming and outlining, and I still use AI, for cooking, working on French, training plans, and whatever I happen to be curious about that week. It's mostly useful, which is exactly what makes the rest of it so maddening. I also use it to help refine the writing for this blog, our website, and our newsletter, though never to plan anyone's trip. If a tool could reliably take over the tedious parts of planning, the route math and the timetable comparisons, I'd take the hours back gladly.

A chatbot will ask you questions now, too, if you give it the chance. That doesn't mean it asks the right ones: how fast you like to move, whether stairs are a problem, whether you'd rather have a guide or be left alone with a map. And I've found the ones it does ask can be leading, pushing you toward its own presumptions, and it can draw the wrong conclusions from your answers.

Every Sentence Sounds Right

The error rate on its own would be manageable. What makes it unmanageable is the voice. In my experience, ChatGPT, Gemini, Claude, and the rest get things wrong often, and they deliver every answer in the same assured register, the right ones and the wrong ones alike. It's like traveling with a friend who has never once said "I'm not sure," and who, when you finally prove him wrong, apologizes beautifully and then tells you some other false bit of information with exactly the same confidence.

A wrong answer that announced itself would be easy to live with. One that reads exactly like a right answer gives you no signal about which sentences to check, which means, if you're careful, checking all of them. At that point the plan hasn't saved you any work. It's just handed you a longer to-do list.

It Was Never Written Down

The obvious objection is that these tools search the live web now, so a closed restaurant is a stale-data problem that fresher data will fix. That assumes a restaurant tells the internet when it closes, and plenty never do. The listing stays up, the old reviews stay up, and the only way to learn the kitchen has been dark for two years is a phone call. But the harder problem in rural Europe isn't information that has gone out of date. It's information that was never written down.

A client of ours traveled with her dog. She was staying outside Paris and making day trips, with more planned in Brittany, and almost none of the museums and châteaux on her list said anything online about whether a dog could come in. The only way to know was to ask each of them. Near us, in rural Burgundy, asking usually means hunting down a phone number and calling someone who, often as not, doesn't speak English.

The vendange (the grape harvest) is the same. It moves every year, and when it falls depends on the weather of the months before it. No website can tell you in the spring, because nothing has been decided yet. It's being settled in the vines, and I hear how the season is going from the vignerons themselves, because I'm at their domaines buying wine anyway, which makes it the most agreeable research I never set out to do. The holiday calendar is written down, of course, but it can't tell you how the region moves around it.

A tool that searches the web can only find what somebody posted, and out here, a great deal of what matters never gets posted at all. The best travel advice we ever ignored came from a man in a Corsican cafe who told us it was far too hot for the hike we were about to start. He was right, we stupidly went anyway, and you won't find him on any website.

Our Names Are on It

Then there's the question of who answers for it. Build a day around one thing, have that thing turn out to be closed or under renovation, and the day is ruined as planned. Put someone in a hotel nobody verified, let them have a miserable week, and somebody ought to answer for that. When a chatbot built the plan, nobody does. The tool won't be embarrassed. It won't lose sleep or a reputation.

Our names are on everything we recommend, and we buy from the people we recommend. We buy wine from Michel Goubard et Fils in Givry, and from Cindy and Jérémy at Domaine Vervier in Solutré-Pouilly. We've taken private tours with guides, and been walked through hotels, gîtes (furnished holiday rentals), and chambres d'hôtes (bed and breakfasts). We haven't slept in every room or eaten in every restaurant we'd point someone toward, and no honest agency has. But we've looked, we've asked the questions that matter, and we know how these people work and how they treat their guests. If we have a bad experience with someone, we don't forget that either.

And plans change whether anyone wants them to or not. When a client missed trains, we were on the other end of the phone, and having missed the last train out of Paris myself, I took those calls without judgment. When a client found something worth following, the plan bent around it.

Five Years From Now

People keep asking what these tools will look like in five years. I have no idea, and neither, as far as I can tell, does anyone else, though plenty of people will tell you with total confidence, which rather brings us back to the point. My own hunch is that they won't come as far as people assume. They'll still know only what the internet tells them, and we'll still be here, buying bread from the same boulangerie most mornings and asking vignerons about the weather, however much the landscape of European travel shifts around us.

We're not neutral on this, obviously. We live here. But it's a question worth asking before your next trip: would you rather it were built by an algorithm working from whatever happens to be written down, or by a creative human being on the ground in the place you're going?