How to use AI for sightseeing without it ruining the experience

AI in travel is mostly noise. Most "AI travel app" pitches add an LLM to a problem AI does not solve, ship a chatbot, and call it a feature. Done badly, AI flattens travel. It generates the same recommendations everyone else gets, hallucinates landmarks that do not exist, and trades the texture of a place for a smooth interface.

Done well, AI does three things in sightseeing that are genuinely useful. This guide walks through them.

What AI is actually good at in sightseeing

1. Personalising what you see, not telling you where to go

A travel app that funnels every user to the same five landmarks is a bad app, AI or not. AI's real strength is taking the messy edges of the world (every fountain, every plaque, every odd building tagged on the map) and surfacing the slice that matters to you specifically.

Concretely: instead of "top 10 things to see in Rome", the question becomes "what's interesting within forty metres of where I'm standing, given that I care about food history rather than cathedrals?" That's a problem AI is good at. It is also a problem most existing travel apps do not even try to solve.

2. Telling stories instead of reciting facts

Wikipedia is exhaustive but exhausting. A landmark's encyclopaedia entry is the kind of thing you read once and remember none of. AI is good at taking that source material and rewriting it as a 30-second story with a surprising angle and a payoff. The kind of fact you actually repeat at dinner.

The trick is constraining the AI to use the source material rather than inventing new "facts" — more on that in the next section.

3. Turning a vague intent into a concrete walk

"I want hidden chapels in the old town." "Top espresso bars locals use." "Spots where the architecture changes most dramatically." These are the kind of vague but interesting prompts no fixed-route tour app can answer. AI turns them into a list of specific places, with a route that works.

This is where AI in travel earns its weight. Not chatbot, not generic recommendations. Specific intent → specific places.

What AI ruins if you let it

The hallucination problem

The single biggest failure mode of AI travel apps is the AI confidently inventing landmarks that do not exist. A user sees "St. Anthony's Chapel, 14th century" on screen, walks to it, and finds an empty street. This happens because the LLM is generating from its training data without a check on what's actually there.

The fix is grounding: feeding the model the real artefacts at your location (from a map data source like OpenStreetMap) and the real Wikipedia entries for them, and constraining the model to only reference those. If an AI travel app does not say what it grounds in, treat its output as fiction until proven otherwise.

The flattening problem

AI tends toward the average. Ask it for "interesting places in Paris" and it will give you the same answer it gives everyone else. The flatness gets worse the more general the prompt.

The fix is specific input. The more specific your prompt — "passages couverts" instead of "Paris" — the less the AI can flatten. Apps that surface specific sub-categories (architecture, street art, food history, micro-neighbourhoods) get less generic results than apps that ask "what should I do in Paris today?"

The smoothing problem

A great AI travel interface can paradoxically remove the friction that makes travel memorable. Looking up at a building and wondering what the story is, before pulling out your phone, is part of the experience. An app that interrupts that moment with a notification before you've had time to wonder is worse than one that waits.

The fix is restraint. A good AI travel app should pull, not push. It should be there when you ask. It should shut up otherwise.

What to look for in an AI travel tool

• Does it say what it grounds in? "Powered by AI" is a red flag. "Pulls landmarks from OpenStreetMap, sources facts from Wikipedia" is a green flag.
• Does it personalise to specific intents, or only handle "things to do in [city]"? The former is hard, the latter is generic.
• Is the AI doing storytelling, or just retrieval? An app that summarises Wikipedia is not using AI well. An app that takes a Wikipedia entry and turns it into a 30-second story with an angle is.
• Does it leave you alone when you don't ask? Pull-based is better than push-based for sightseeing.

Naruho's approach (disclosure: I built it)

Naruho was designed around these constraints. Landmarks come from OpenStreetMap, so the app does not invent places. Stories are grounded in Wikipedia, so the AI cannot drift. Plans are AI-generated from your specific prompt, so "hidden chapels in the old town" gives a different result than "things to do in Rome". And the app stays quiet unless you tap or it picks up something within forty metres while you walk.

Try it free, or read how it works if you want the technical underpinning.