Where AI gets travel dangerously wrong

A client sends you a screenshot of an itinerary ChatGPT built for them and asks if it's "good to go." The route looks fine. The place names sound plausible. You have thirty seconds before the next call and you're tempted to just say yes.

That thirty-second temptation is where the next expensive mistake in your business lives. AI itinerary tools are confident, fluent, and wrong often enough that "chatgpt travel planning mistakes" is now a real search query, not a niche worry. A CNBC report from March 2026 found that a third of people who used AI to plan a trip received false information somewhere in the process, and the tool is popular enough that this is no longer a rounding error. CNBC's coverage puts AI trip-planning usage at roughly 24-30% of tourists already.

This post is not "don't use AI." It's the opposite problem: AI is now embedded in how your clients plan trips, and some of them will forward you its output expecting you to be the safety check. This is what that check needs to catch, with real incidents to show why it matters.

Why "ai hallucination travel examples" is now a real category of failure

A hallucination, in AI terms, is the model generating something that sounds correct but isn't: a place, a timing, a rule. It's predicting plausible-sounding text, not looking anything up. Travel is an unusually bad match for this. Place names, opening hours, permit rules and route distances are exactly the kind of specific, low-frequency facts that language models get wrong most often, and a wrong answer about a Ladakh permit or a ropeway timing reads exactly as confident as a correct one.

The same CNBC report that found the 33% false-information rate also found tourists trust these tools anyway. That's the actual danger. A wrong restaurant recommendation is a bad meal. A wrong altitude, ropeway closing time, or route distance can put a traveller somewhere they didn't plan to be, with no plan for what happens next.

The nonexistent destinations that nearly got people hurt

These aren't hypothetical edge cases. They're documented incidents from the last two years, and they share a pattern: the AI didn't say "I'm not sure." It invented a specific, named, confident answer.

• The "Sacred Canyon of Humantay," Peru. Tourists were en route to a canyon that does not exist, invented by an AI trip-planning tool, before a local guide intercepted them. They were headed toward terrain at roughly 4,000m altitude, without oxygen or phone signal, chasing a place that was never real. Futurism's reporting has the detail.
• Mt Misen, Japan. A couple got stranded on the mountain overnight after ChatGPT gave them the wrong ropeway closing time. The model didn't flag uncertainty; it gave a specific time, and it was wrong.
• The "Kuak Skyride," Malaysia. A couple drove roughly 370km to ride an attraction that existed only in a hyper-realistic, AI-generated video with an AI presenter narrating it. Nothing about the video signalled it was fabricated. Travel And Tour World has documented the same pattern of fake-destination AI videos pulling tourists toward nonexistent attractions in Spain, South Korea, Brazil and Singapore too. This is a global pattern now, not a one-off.

Careful: the common thread isn't exotic destinations or gullible tourists. It's that AI-generated content, text or video, carries no visible marker of confidence versus fabrication. A hallucinated canyon reads exactly like a real one. That's the whole problem with treating it as a source instead of a draft.

Why this is now also a legal question, not just an accuracy one

Here's the part that should change how you handle AI output in your own business, not just how you warn clients about theirs.

In Moffatt v. Air Canada (2024 BCCRT 149), a Canadian tribunal held Air Canada liable for its own website chatbot giving a customer incorrect bereavement-fare advice. The airline argued the chatbot was "a separate legal entity responsible for its own actions," and the tribunal rejected that outright. Air Canada paid out $650.88 in damages. The amount is trivial. The principle isn't: a company is responsible for what its AI tells a customer, full stop. "The AI said it" is not a defence.

Example: Say your team pastes a client's dates into ChatGPT, gets a full 6-day Vietnam itinerary back, tidies the formatting, and sends it out under your agency letterhead. If a listed hotel doesn't exist, a border-crossing rule is out of date, or a "2-hour" transfer is actually 5, that itinerary went out with your name on it, not the AI's. You forwarded it. You own it.

An Indian tour operator forwarding an AI-drafted itinerary is in the same position Air Canada was, minus the resources to absorb a bad outcome. A large airline can afford a tribunal ruling and a few hundred dollars in damages. A small agency dealing with a client stranded because of a wrong permit rule or a wrong altitude estimate is dealing with a refund demand at best and a consumer court case at worst, with none of the "the software said so" defence available, because your client didn't contract with the AI. They contracted with you.

The verification SOP: what never ships unchecked

The fix isn't banning AI from your itinerary process. AI-drafted itineraries are genuinely useful for structure, pacing and first drafts. The fix is treating every AI-generated fact as a claim to verify, not information to relay. Three categories need a human check before anything reaches a client, every time:

1. Timings and permits. Ropeway, monument, border, and park timings come from the official source (tourism department site, the property itself) or a very recent on-ground confirmation, never from the model's memory. Permit rules for high-altitude and border-area destinations change; verify current requirements each season, not once.
2. Distances and routes. Pull drive times and distances from a mapping tool, not from the AI's description. "It's a scenic 2-hour drive" is exactly the kind of specific-sounding, unverifiable claim that turns into a 5-hour drive or a road that doesn't exist.
3. Named places. If a place name in an AI-generated itinerary doesn't show up on a map or in an official tourism listing, it doesn't go in the itinerary, full stop, no matter how good the description sounds. Treat unfamiliar attraction names the way you'd treat an unfamiliar hotel: don't book or recommend it until you've confirmed it's real.

Signal in AI output	What to do before it ships
Specific timing (opening/closing hours, ropeway, permit office)	Confirm against the official source, same season
Specific distance or drive time	Check on a mapping tool, not the AI's description
Unfamiliar place name, attraction, or "hidden gem"	Confirm it exists via map or official tourism listing
Confident tone with no source given	Treat as a draft, not a fact, regardless of how it reads

This is a five-minute check per itinerary, not a research project. The cost of skipping it is a client at altitude with no signal, or a chargeback with your name on the paperwork.

Common questions

Can I trust ChatGPT for trip planning at all?

For structure, ideas and first drafts, yes. It's fast and often genuinely useful. For facts a client will act on (timings, permits, distances, whether a place exists), no. Not without checking. Treat it as a junior assistant who writes confidently and has never actually been anywhere.

What should I tell a client who shows up with an AI itinerary?

Don't dismiss it. Many of them are a reasonable starting point. Walk through it with them and flag exactly which parts you've verified versus which parts are still the AI's word: timings, permits and named attractions specifically. That conversation also happens to be a good sales moment. It's the clearest way to show a client what they're actually paying an agency for.

Are fake AI-generated travel videos a bigger risk than chatbot text?

In some ways, yes. A video with an AI presenter and drone-style footage of a nonexistent attraction is more persuasive than a wrong sentence in a chat window, and it spreads on social platforms with no fact-check attached. The Kuak Skyride case in Malaysia and similar incidents in Spain, South Korea, Brazil and Singapore all followed viral AI video content, not a chatbot conversation.

The short version

• A CNBC survey found a third of people using AI to plan trips got false information, and roughly a quarter to a third of tourists now use AI for planning. This is a mainstream problem, not an edge case.
• Real incidents include a nonexistent "Sacred Canyon of Humantay" in Peru that nearly led tourists into high-altitude terrain, a couple stranded overnight on Mt Misen after a wrong ropeway timing, and a 370km drive to a fake AI-video attraction in Malaysia.
• Moffatt v. Air Canada established that a company owns what its AI tells a customer: "the chatbot said it" is not a legal defence.
• If your agency forwards an AI-generated itinerary and something in it is wrong, you own that outcome the same way Air Canada did, with none of the resources to absorb a bad one.
• Never let AI output ship unchecked on three things: timings and permits, distances and routes, and whether a named place actually exists.
• Use AI for drafting structure and pacing. Use official sources, maps, and your own judgement for anything a client will act on.