Some wrong travel advice costs you a missed lunch. Some of it costs you a day, a tank of fuel, or a mountain rescue. The difference usually comes down to whether the recommendation was ever checked against the ground.
In September 2025, a Peruvian mountain guide, Miguel Angel Gongora Meza, overheard two tourists planning an unguided trek to the “Sacred Canyon of Humantay” in the Andes. They had a confident screenshot. They had paid about $160 each. They were heading for high altitude with no oxygen and no phone signal. “There is no Sacred Canyon of Humantay!” he told the BBC. The name was two unrelated places stitched together by a chatbot into something that read as real.
That is the failure mode worth worrying about. Not clunky AI prose, which most of us can smell a paragraph away, but fluent AI prose about a place that does not exist, or one that exists under rules the model never learned.
Why a hallucination reads like the best guide you have ever found
Large language models generate the statistically plausible next sentence, not the verified one. As the BBC’s reporting on AI-planned trips explains, the same system that produces an accurate answer produces an invented one in the same tone, with the same confidence and the same adjectives. Researchers call it hallucination, and Google’s chief executive Sundar Pichai has described it as a likely inherent feature of the technology, which means the problem is not being patched away.
The scale is now measurable. An August 2026 survey of 1,000 US leisure travellers by Greetwell found that of the 485 who had used AI for travel, 55% had hit at least one recommendation that was inaccurate, unavailable, closed or simply wrong. Roughly one in six, or 16%, had been sent to a place, tour or activity that did not exist at all.
Scotland is not somehow immune. In January 2026, Anne Hardy, an adjunct professor of tourism at Southern Cross University, told CNN that empirical research suggests around 90% of AI-generated itineraries contain at least one error. Her examples were exactly the kind of details that decide whether a Highland day works: wrong walk lengths, wrong difficulty ratings, wrong weather assumptions. In remote glens with no mobile coverage, those are not annoyances. They are the whole risk profile.
Scotland is unusually easy to fake badly
The Highlands are a near-ideal surface for a hallucinating model. Place names are Gaelic, often with no English equivalent the model has seen often enough to ground. Many sites are documented in fragments across land manager pages, local blogs and forum threads rather than one authoritative listing. Access changes with season, weather and land management. Public transport to trailheads is thin, so a small distance error becomes a long detour.
Now add demand. The Old Man of Storr on Skye drew about 304,000 visitors in 2025, up from 205,056 in 2022, according to a Highland Council operational report from August 2026. The North Coast 500 route now generates tens of millions of pounds a year and a formal economic and environmental impact study, prompted by overtourism complaints, was commissioned in late 2025. When a machine scrapes the loudest content about these places, it tends to serve you the busiest, most photographed version with no sense that the communities around them are stretched.
VisitScotland already knows this. Its own content team once asked ChatGPT for places to visit in Scotland and found the model pushed the same overcrowded honeypots, ignored responsible tourism, and lacked any local insight. The organisation’s public guidance now takes a firm line: never publish content that is 100% AI-generated, and always fact-check specific mentions of locations and landmarks before using them.
Two minutes of text checks, and one honest caveat
Before you trust a Scotland guide, read a couple of paragraphs as if you were grading them. AI text tends to run in a flat, even rhythm, to reach for the “not this but that” construction, and to praise places in generic terms (“exceptionally warm”, “a must-visit”) rather than naming a landlord, a bothy, a bus route or a specific hill. Real travel writing is messy in a way machines avoid: it mentions the midges, the closed cafe, the missed ferry.
Reviews have their own tells. Clusters of near-identical praise posted on the same day, accounts with one contribution and a number in the username, or the same factual error repeated across several glowing posts are classic bot-farm signals. You can paste a suspicious paragraph into an AI checker to get a likelihood score, and it can be a useful first pass.
But treat the score as a hint, not a verdict. A peer-reviewed evaluation published in the journal Information in October 2025 tested six widely used AI content detectors and found accuracy ranging from 14.3% to 71.4%, with precision never exceeding 11.1% and results that flipped after minor edits. None of these tools is reliable enough to be the final word, in either direction.
Where to check any claim, in order of reliability
The reason a detector is a weak instrument is that it only tells you how a text was made. It cannot tell you whether the bothy exists, whether the car park closes at dusk, or whether the ridge walk is a grade you can handle. For that you need ground truth.
| Source | Best for | Weakness |
|---|---|---|
| Land manager or government body (Historic Environment Scotland, NatureScot, National Trust for Scotland) | Sites, access, seasonal closures, booking rules | Sparse on cafes, parking and logistics |
| VisitScotland | Regional overviews and responsible-travel framing | Promotional by design |
| Ordnance Survey or a current paper trail map | Distances, gradients, actual paths | Silent on closures and conditions |
| The business’s own booking page or a phone call | Opening hours, availability, tour spaces | Requires time and, sometimes, a call |
| Recent reviews with specific detail | Real conditions, recent closures, staffing | Polluted by fake and AI reviews |
Source order reflects how each is maintained and when it was last checked, as of October 2026. Start at the top and work down only as far as you need.
Two habits do most of the work. First, cross-check the place name against an official listing; if a “hidden spring” or “secret viewpoint” appears in one blog and nowhere else, it probably does not exist. Second, reverse-image search the photo. AI-generated scenery is now used freely in guides, and VisitScotland’s own generative AI policy states plainly that it will not use AI-generated or over-enhanced images to represent the country, precisely because authenticity is treated as the asset.
The honest case for AI travel guides
It would be easy to write off AI planning entirely, and that would be lazy. The technology is genuinely good at the jobs that do not require truth: brainstorming, translating a rough wish list into a route shape, suggesting a starting point you then verify. The same Greetwell survey found that 34% of AI-travel users said their trust in it had increased over the previous year, and proponents argue that used well it can spread visitors toward lesser-known places rather than concentrating them further.
My read is that the risk is not evenly distributed. An AI itinerary for a city break is a low-stakes draft; a bad ferry time costs you an afternoon. An AI itinerary for a solo ridge walk or a remote glen in winter is a different category of object, because the same confident voice that invents a hot spring will also invent a safe-looking route. I would happily use AI to start planning Scotland. I would not let it be the last thing that touched the plan.
What would change my mind? Evidence that the models had stopped hallucinating specific, checkable facts at scale. The current data runs the other way, so I will keep verifying.
A Scotland pre-trip verification checklist
- Confirm every named attraction exists on a land manager or VisitScotland page, not just on a blog.
- Check opening dates and ticketing directly. Historic Environment Scotland recommends booking online in advance for Edinburgh Castle, Stirling Castle, Urquhart Castle, Skara Brae and Maeshowe, and warns that a sold-out online day leaves no walk-up tickets.
- Match any walking route to a current map and a land manager’s access notes, then recheck the forecast the day before.
- Call or message any small tour, distillery or boat operator you plan to use. Availability is the detail AI gets wrong most often.
- Reverse-image search one photo from the guide. If it exists nowhere else, treat the whole page as suspect.
- Read the Scottish Outdoor Access Code and note responsible-travel rules for the specific area.
Frequently asked questions
How can I tell if a Scotland travel blog was written by AI?
Look for uniform sentence rhythm, generic praise with no named people or places, the “not this but that” construction, and details that cannot be verified anywhere else. One unverifiable landmark is a stronger warning sign than any style tic.
Are AI content detectors reliable enough to trust?
No. A 2025 evaluation in the journal Information found accuracy across six popular tools ranging from 14.3% to 71.4%, with results that changed after small edits. Use a detector as one signal among several, never as proof.
What is the single best source to confirm a Scottish attraction exists?
The body that manages the land or site. Historic Environment Scotland, NatureScot and the National Trust for Scotland all publish current access, ticketing and seasonal information for the places they look after, and none of them invent a location to fill a listicle.
Is it safe to plan a Highland hike from an AI itinerary?
Not on its own. Verify the route against a current map and the relevant land manager, check weather and conditions, and treat any walk in remote terrain as something to plan with local advice. Scottish Mountain Rescue and the Scottish Outdoor Access Code are the starting references.
Do I need to worry about fake reviews for Scottish hotels and tours?
Yes. Tripadvisor’s 2025 Transparency Report said it removed more than 2.7 million fraudulent reviews, including 214,000 AI-generated ones, and the platform has faced criticism over AI summaries that downplayed traveller safety concerns. Favour reviews that mention specific, checkable detail.
How this article was put together
I relied on primary and first-hand reporting rather than SEO roundups: the BBC’s investigation into AI-planned trips, CNN’s coverage of the Weldborough hot springs case, Originality.ai’s analysis of AI reviews on TripAdvisor, a peer-reviewed detector study in the journal Information, VisitScotland’s published AI policy, the Greetwell AI Travel Report 2026, and Highland Council’s operational report on the Old Man of Storr. Visitor figures are for 2025 and were current when that report was published in August 2026. Review-fraud numbers come from Tripadvisor’s own transparency reporting and should be treated as self-reported. Where I give a personal view, it is marked as mine.
The short version
A detector can tell you a paragraph looks machine-made. It cannot tell you the glen is real. The only test that survives a fake hot spring, a closed car park or a phantom ridge is ground truth from the people who manage the place, and that test costs you a handful of searches and, occasionally, a phone call. On a trip where the wrong turn can mean hours in the cold, that is the cheapest insurance on the market.