Can AI plan a trip

Language models are good at structure, which is the part most of us find hardest. Knowing that Kyoto splits into an eastern temple day and a western Arashiyama day, or that four days in Iceland means the Golden Circle and the South Coast, is pattern knowledge worth having. It buys back the blank-page hour, which is the hour most trips die in.

What they get wrong is anything specific and current: opening hours, seasonal closures, whether a permit system exists this year, how long a mountain road really takes, whether a named restaurant is still in business. A confident wrong answer looks exactly like a right one, and that is the whole problem, because you cannot tell which one you are holding.

Check that the place exists and is open before you do anything else, that the travel time between consecutive stops matches what the plan assumes, and that anything needing a timed ticket is available on your date. Three checks, ten minutes, and you will catch nearly everything that would otherwise cost you a morning.

A planner that looks each suggested place up in real venue data will tell you the address exists, while dropping the list on a map makes a two-hour gap between stops obvious. Pasting a ChatGPT itinerary onto a map is the version of this most of us already have to hand.

Ask for a four-day shape with named venues and neighborhoods, and then paste the list onto a map. Delete anything that fails to geocode, anything whose hours collide with your dates, and any pair of consecutive stops more than a 25-minute hop apart unless you meant it. What survives is a draft you can book against.

A confident itinerary for Seoul that puts a palace, a market, and a hill neighborhood on one Tuesday reads well in a chat window, but it collapses on a map. Geography is the check language models skip. Organizing by day applies that check to whoever wrote the list, because a human list scatters the same way.

Booking a table at a restaurant that closed last year, or a hike that needs a permit the model never mentioned, is the usual failure. The draft was fine, so the reservation was the mistake. Run the three checks on anything that takes a card, and leave the rest as a maybe list until you are standing in the city.

Models will also invent a small, little-known bar with a completely plausible name. If you cannot find the address in a maps app, it stays off the trip, which is why Google Maps lists make a better save file than a chat transcript.

Keep the shape: which neighborhoods pair, how many days a city wants, what a sensible Tuesday looks like. Throw away every specific hour and price, because those are the parts that go stale between the training run and your trip.

Verify existence, travel time, and timed tickets on anything that takes a card, because everything else can be wrong at no cost to your week. That is the line, and it is easy to hold once you have said it out loud.