AI travel itineraries: what works and what does not

Asking an AI "plan me three days in Palermo" always works, in the sense that an answer arrives. The point is another: is it right? A good itinerary is not a well-written list, it is a sequence of real places, open, reachable in the order it proposes. And that is where things get complicated.

What AI does well#

• It orders. It turns ten scattered ideas into a day with a logical thread.
• It adapts. If you say "I am with kids" or "I only have half a day", the angle changes.
• It explains. It tells you why a place is worth it, not just where it is.
• It rewrites. If a ticket falls through or it rains, it redoes the programme in seconds.

What it does badly#

• It invents with confidence. A generic model can suggest a restaurant that has closed, a museum that does not exist, a wrong opening time.
• It ignores geography. It can line up stops that are half an hour apart on foot on the map.
• It does not know it is raining today. Unless it receives the data, it does not adapt the programme to the day.
• It flattens the rules. Information that changes by season, weekday or event is often squeezed into one sentence that is true "in general".

Why does it happen? The short version#

A paper by researchers at OpenAI and Georgia Tech describes hallucinations as plausible but false statements. According to the authors they persist because training and evaluation reward guessing more than admitting uncertainty: on a right-or-wrong test, an "I don't know" scores zero while a guess can score a point. It is not a flaw of one product: it is a behaviour to know about and work around with method. Anthropic's guide to reducing hallucinations reaches the same conclusion from the practical side: its techniques significantly reduce errors but do not eliminate them, and critical information should always be validated.

How to use it without getting hurt#

1. Give precise context: area, days, pace, who you travel with.
2. Ask for close stops, grouped by neighbourhood.
3. Give the AI permission not to know. Anthropic's guide lists this as the first step: allowing the model to say it does not know the answer can sharply reduce false information.
4. Ask for the source of every rigid fact (hours, price, rule) and ask it to flag the ones it cannot confirm.
5. Check every stop on a map: does it exist, where is it, how far from the previous one.
6. Check opening hours on the official source.
7. Ask for a plan B for rain or tiredness.

Prompts you can copy#

The first builds the itinerary. Paste it and replace the parts in square brackets:

The second is the audit, to run in the same conversation:

The third is plan B, which works well because it starts from a plan you have already checked:

Cases where AI gets it wrong, and how to spot them#

Here are four situations you can demonstrate with official pages. You do not need statistics: examples are enough.

Hours change by season. On the Colosseum Archaeological Park site, for the period from 25 October 2026 to 28 February 2027 entry is listed from 8:30 to 16:30, with last entry at 15:30. A generic summary saying "until sunset" or "until 7 pm" is right in one period and wrong in another. Red flag: round hours with no date attached.

Entry rules are not in guidebook summaries. Colosseum tickets are named and you show an ID at the gate. An itinerary that only says "buy tickets" leaves out the point that can stop you from getting in. Red flag: no mention of ID, the name on the ticket or booking.

One place, several uses. The Pantheon is an active church: the Ministry of Culture page lists Mass at specific times and says that during services entry is free and limited to worshippers. An AI that slots it in "Saturday at 5 pm" ignores that constraint. Red flag: places of worship with no mention of service times.

Sources disagree. On that same Pantheon page, the Italian and English texts give different limits on tickets per purchase (25 versus 10). A model drawing on a single version will hand you a confident number that may be wrong. Red flag: a precise figure with no source next to it.

And the Vatican Museums: the official page lists recurring closures (Sundays, except the last of the month, plus several dates during the year). A plan that puts the museum on "Sunday morning" is plausible and wrong, unless it is the last Sunday of the month.

Checklist before you leave#

• Every stop shows on a map with its exact name and a sensible position.
• You looked at the official site for hours, closures and day of the week.
• You read the entry rules: named tickets, ID, dress code, luggage.
• Consecutive stops are reachable in the time planned, calculated on a map and not by eye.
• There is a real lunch break, not a "free lunch" line between two museums.
• You have a plan B for rain and one for tiredness.
• Figures (prices, times) are marked as indicative and have a source.

A note on maps: even Google Maps hours are not an official source. The help page explains that hours can be suggested by users and that edits are reviewed and may take time to appear. They are fine for getting your bearings, not for deciding whether to arrive at 16:20 at a site with last entry at 15:30.

When a guide is enough (and you do not need AI)#

AI is at its best when you have to choose and order. If instead you are going to Rome to see three places with strict hours and named tickets, you need one piece of information: what the official site says. Then a good guide, maybe the one you are reading, like Rome in two days, is worth more than an answer generated on the fly. Use AI for the rest: timing, order, alternatives and ideas for the evening.

Why real places make the difference#

A system designed for travel starts from the opposite end: real places first, with their coordinates, then the AI that orders and narrates them. That way it cannot suggest a place that is not there, because it does not fish it out of its memory. It is the idea behind UNNIVAI: a route made of real places, built around your time and your tastes, and rewritten when the day changes.

Concrete examples#

The rule holds for every city: try it on the guides Rome in two days and Ortigia in one day, where every stop has a place, an area and a time. For a well-made plan B, read what to do in Catania when it rains.

The five-minute test#

Before trusting any itinerary, even your own: pick three stops at random and check them on the map. If all three exist, are close and are open, the rest usually holds. If even one fails, start again.

If you want to try a different approach, join the UNNIVAI waitlist.

Verification note: the examples on hours and rules were read on the official pages of the Colosseum Archaeological Park, the Ministry of Culture (Pantheon) and the Vatican Museums on 29 October 2026; the statements about language models come from the Anthropic guide and the Kalai et al. paper cited above. We did not measure how often individual AI assistants make mistakes and we give no percentages. Hours and rules change: always check the official source.