What AI gets right (and spectacularly wrong) about travel recommendations — Budge

Travel related questions require holding several variables simultaneously — distances, transport options, time, accommodation logic — and AI handles them quickly and accurately.

Ask ChatGPT for the best ramen shop in Shinjuku and it will give you a name, a neighbourhood, a description of the broth style, and probably the approximate price. It will do this with the same confident tone it uses to explain quantum physics. There is a reasonable chance the restaurant closed in 2022.

This is the specific failure mode that makes AI travel recommendations genuinely dangerous rather than just imperfect: the confidence is uniform regardless of reliability. The system doesn't know what it doesn't know, so it doesn't signal uncertainty where uncertainty is warranted. You get the same flat authoritative voice for "Tokyo is the capital of Japan" and "this specific ramen shop in a specific alley is still open and worth the queue."

Understanding where AI is reliable and where it isn't — specifically, structurally, rather than vaguely — is what makes the difference between using it well and getting burned by it. The broader case for AI in travel planning covers the framework; this post goes deeper on the specific right-and-wrong taxonomy.

What AI does well in travel planning

Logistics reasoning with multiple variables

This is where AI genuinely outperforms the alternative. Is the itinerary Tokyo–Nikko–Kyoto–Nara–Osaka–back to Tokyo physically achievable in ten days? What's the sensible direction to travel a ring route in New Zealand? Should you base yourself in Split or Dubrovnik if you want to visit Hvar and Korčula? These questions require holding several variables simultaneously — distances, transport options, time, accommodation logic — and AI handles them quickly and accurately.

The specific value is that it catches things you didn't think to check. Describe your itinerary and ask "what's wrong with this plan" — a good AI response will surface: your Kyoto–Nara–Osaka day is physically possible but exhausting, you'll miss the morning crowds at Nara if you're coming from Kyoto, and the accommodation you've picked is on the wrong side of Osaka for your next day's start point.

Synthesising destination information

Researching a destination used to mean reading eight sources and mentally reconciling them. AI does that synthesis in seconds. The question "what should I know about Lisbon that most travel guides skip" produces genuinely useful output — not a ranked list of attractions but the contextual knowledge that shapes how you approach a place. Which neighbourhood to stay in and why. What the public transport really covers and where it doesn't. What time of year the famous things get crowded enough to ruin them.

The reliability here is high for stable, structural information: geography, transport systems, price levels, seasonal patterns, cultural context. These change slowly and AI training data captures them well.

Budget frameworks

"What does a mid-range week in Lisbon cost including accommodation, food, and activities" — AI handles this accurately enough to be useful for initial planning. Not to the decimal, and not current pricing, but the order-of-magnitude framework is reliable. You'll know whether you're looking at a €80/day destination or a €200/day destination, which is what you need at the planning stage.

Visa and entry requirement research

AI is good at telling you the general framework — what visa category applies to your nationality, whether you need to apply in advance, what the general processing time is, what documentation is typically required. It's not a substitute for checking your government's official travel advisory because requirements change, but it's a fast way to understand what questions to ask before you go to the official source.

Generating options you didn't know existed

The most underused AI capability in travel research is expansion of the option set. Ask "what are alternatives to Santorini for someone who wants volcanic Aegean scenery without the crowds?" and you get Milos. Ask "what's the best base for exploring the Peloponnese that isn't Nafplio?" and you get Kalamata or Gythio. Ask "what's between Florence and Rome that most people skip?" and you get Orvieto, Civita di Bagnoregio, the Etruscan towns. These suggestions aren't obvious from standard searches and they're often the best part of a trip.

What AI does poorly

Specific local restaurant recommendations

This is the clearest failure case and the most important to understand. Ask AI for a specific restaurant — particularly a small independent one in a non-English-speaking country — and you are getting a name from training data that may be years out of date. Restaurants close, change ownership, decline in quality, change their concept, or move without leaving a trace in AI training data. The AI has no mechanism to know this has happened and no signal to indicate uncertainty.

The problem compounds because AI doesn't distinguish between "this restaurant is famous and verified by thousands of sources over many years" (relatively reliable) and "this restaurant appeared in a blog post from 2021 that trained my model" (completely unreliable for current status). Both come out in the same confident voice.

For restaurant recommendations, use: Google Maps sorted by "Newest" reviews, Tabelog (Japan), The Infatuation (major Western cities), local food blogs with recent dates, or ask someone who's been there in the last year. AI can tell you which neighbourhood to look in and what kind of restaurant to find — the specific name is the part you need to verify.

Real-time pricing

AI has a training cutoff and no live booking system access unless explicitly connected to one. When AI tells you a flight from London to Tokyo costs around £820, that's a figure from training data, not today's fare. When it says a room at a specific Kyoto ryokan costs ¥25,000 a night, that may be the 2023 price, the 2020 price, or an average that never applied to a specific room category.

Use AI for order-of-magnitude planning (is this a €100/day or €250/day destination). Use actual booking platforms for real numbers that you'll commit money to.

Cultural nuance in safety advice

AI tends toward over-cautious, liability-hedging safety advice that isn't calibrated to the actual risk level of a specific destination, neighbourhood, or activity. "Exercise caution" applied uniformly to a neighbourhood in Lisbon and a neighbourhood in Caracas is not useful information. Specific, nuanced safety advice requires current, local knowledge that AI training data captures poorly — and AI has strong incentives toward caution rather than accuracy here.

For genuine safety information, use your government's travel advisory (which has specific risk levels), recent forum posts from travellers who've actually been there, and the judgement of local operators and accommodation staff when you arrive.

Understanding personal travel pace

AI cannot infer how travel-tired you get without being told explicitly. It will build you an itinerary that is physically achievable for someone with moderate fitness and high travel energy, because that's what itineraries look like in the training data. If you have a bad knee, need a slow morning after a late night, travel with someone who needs more rest time than you, or simply find intensive daily sightseeing exhausting after day four — you have to say this explicitly, and even then AI doesn't fully model what a day of continuous novelty actually costs in cognitive energy.

The fix is explicit prompting: "I travel best with one major activity per day and unhurried afternoons" or "my partner needs at least one slow morning in three." These constraints change the output substantially. But AI won't ask for them without prompting.

What AI simply cannot do

Know what a place feels like.

The smell of the spice market in Marrakech, the specific quality of light in Lisbon at 6pm in October, the way a Greek island empties out when the last ferry leaves and the atmosphere transforms — these are not in training data. AI can tell you these things are true; it can't convey what they feel like or help you know whether you'll respond to them. This is fine — it's not what AI is for. But it matters as a calibration: AI research gets you to a destination; being there is what fills in the rest.

Catch what you haven't thought to ask.

A good travel agent or well-travelled friend will tell you things you didn't ask about. Kyoto accommodation in cherry blossom season sells out six months in advance — this surfaces unprompted from experience. The particular coastal road you've planned is notorious for car break-ins — local knowledge. Your planned hiking trail washed out in last autumn's storms and hasn't reopened — current information. AI answers questions; it doesn't (reliably) volunteer what you've missed unless you ask "what am I not asking about."

How different tools handle these failure modes is covered in the AI travel planners comparison — the short version is that purpose-built travel AI is more likely to surface what you haven't asked than a general-purpose assistant.

Using AI accurately

The frame that produces the best results: AI as a research accelerator, not a travel expert. The distinction matters because people who treat it as an expert follow its specific recommendations without verification and occasionally end up at a restaurant that closed eighteen months ago. People who treat it as a research accelerator use it to synthesise, reason, and expand the option set, then verify the specifics with sources that have current local knowledge.

If you want to go deeper on any part of this, Budge is essentially a travel researcher you can have a conversation with — it's what I built because I was tired of piecing together 12 tabs.

Practical rules that follow from this:

Use AI freely for logistics, structure, destination context, budget frameworks, and option generation. These are the tasks where it's fast and reliable.

Verify AI restaurant and accommodation recommendations with current sources — at minimum, a Google Maps search showing recent reviews.

Don't use AI pricing as anything more than order-of-magnitude guidance. Always verify against actual booking platforms.

Give AI explicit personal context — your travel style, energy level, what you're trying to get out of the specific trip — before asking for recommendations. The gap between generic AI output and AI output calibrated to a specific person is substantial.

The tool is genuinely useful. It's useful in the way a very well-read research assistant is useful — excellent at synthesis and reasoning, needs supervision on anything current or local.