Invented huts, life-threatening routes – and a few useful ideas: when AI plans the hiking route

The plan looks great: three days of hiking, two overnight stays – at the Legler Hut and the Rugghubel Hut. Both offer double rooms. Along the way, there will be chamois and glacial lakes, and alpine cheese fondue at the destination. All this is promised by the chatbot that planned the route.

If you look for the huts on the map, you'll be disappointed: there aren't the promised six hours of travel between them, but 50 kilometers as the crow flies.

Artificial intelligence can translate, compose emails, and often knows the right answer to even the most difficult questions . If the promises of tech companies are to be believed, it will soon take over highly skilled jobs. But how does it perform in truly useful tasks—questions whose answers can't be found online with just a few clicks? Like planning a hike? We put it to the test.

The prompt, i.e. the request to all tested chatbots, was as follows:

I want to go on a three-day hike with two overnight stays in alpine huts in Switzerland. I only want to stay in double rooms. Please select suitable huts and hiking trails:

• two huts for a three-day hike,
• with double room,
• maximum eight hours travel time between the huts,
• less than three hours from Zurich.

We submitted the request to a total of nine chatbots from the providers Anthropic, Deepseek, Google, Open AI, Manus, Microsoft, and Perplexity. Seven of these were free and two were paid.

We enabled internet search and, where available, research and reflection features to get the best possible results from the AI.

The chatbot answered brief queries. The chatbots each made one to three suggestions, which we then evaluated. In most cases, the results were hair-raising.

First of all, every chatbot had ideas. In an enthusiastic tone, they announced great hikes with beautiful views and cozy double rooms.

When reviewing the responses, however, it quickly became clear: The AI often advertised double rooms where none existed. The free version of Anthropic's AI, Claude, invented double rooms for both suggested cabins, while the paid version still did so for three out of six cabins.

And while searching for the "Oberhornseehütte" suggested by Deepseek, it became clear that the chatbot had completely fabricated it. The fact that it also suggested a hike in the closed Lötschental valley seems like a forgivable slip-up.

How do these fabricated details come about? The initial suspicion was that the chatbots had accessed real descriptions of multi-day hikes, but couldn't determine whether the trailside huts actually had double rooms. This hypothesis turned out to be false.

The existence of a hiking route is harder to verify than that of a double room. Theoretically, you can hike routes that no one has yet described online. In practice, it turns out that if you can't find a route between two huts online, that's a serious warning sign.

The free version of the AI Claude from Anthropic, for example, suggested hiking in five and a half hours from Grindelwald over the "Scheidegg-Wetterhorn" to the hut of the same name ("2061 m, view of the Eiger, Mönch and Jungfrau").

The AI ignored the fact that the Scheideggwetterhorn is 3,360 meters high and can only be reached by hours of climbing at a high level of difficulty, as well as the fact that there is no Scheideggwetterhorn hut. The chatbot even provided the hut's phone number. At least the mountain rescue number (1414) listed is correct. Anyone who relies on this chatbot's information might need it.

While the paid version of Claude only suggested real huts, the day hikes between them were impossible in two out of three cases. Both the Glecksteinhütte and the Rotstockhütte, as well as the Blüemlisalphütte and the Lämmerenhütte, are 20 kilometers as the crow flies, and include several mountains and valleys. The chatbot's time, distance, and elevation information were incorrect .

Only the third suggested hike in the Appenzell region was feasible based on the key data, although it was incorrectly described.

The Chinese provider Deepseek, Microsoft's Copilot and the AI search engine Perplexity also suggested hiking routes that do not exist.

Apparently, all these AIs have memorized the names of huts and mountains in Switzerland and know what a route description sounds like. However, this information isn't presented in an organized fashion, but rather in a colorful mix, creating a seemingly plausible hiking route that doesn't exist.

But it's not as if today's chatbots aren't up to the task at all. The best chatbots deliver usable results.

The winner in our test is Google's chatbot Gemini. In the free version's research mode, it suggested three hikes, all of which met the specified criteria. It didn't invent double rooms or hiking trails.

This chatbot was also the easiest to control, as it provided a direct link to a source for each sentence. This is likely the reason why Gemini achieved such good results. The bot relies heavily on internet searches.

Gemini followed hiking routes that had already been described on blogs, ensuring that the trail exists and has even been recommended as a hike before.

Google has incorporated its strength in finding information online into Gemini's research mode – in this case, with encouraging results. The search process must have been quite laborious; the AI took more than ten minutes to complete.

Second place is more surprising: It goes to the chatbot Manus from the Chinese company Monica, based in Singapore. Manus caused a stir three months ago when the first testers reported enthusiastically about it. In fact, it performed quite well in our test. The chatbot suggested a feasible route between accommodations with double rooms.

Unlike all other chatbots, Manus initially misinterpreted our prompt. He interpreted the travel time from Zurich not as the start of the hike, but as the arrival at the first hut. As a result, the first day of hiking was eliminated. After an improved query, the results were quite good.

Manus suggested two feasible routes in the Alpstein, with overnight stays in mountain inns. One of them was very suitable as a three-day tour, the second a bit less so, due to the rather large number of descents and ascents between the two huts.

And then there's Chat-GPT, the chatbot that has become synonymous with generative AI for most people—and is still by far the most widely used . This position isn't entirely undeserved, at least according to our testing. Both Open AI AIs suggested feasible routes that passed by accommodations with double rooms.

The free version of Chat-GPT received negative points for fictitious details about the route. And the paid version, similar to Manus, first leads down into a valley and then back up the other side.

An internet search shows that the chatbots have likely combined these routes themselves. They can't be found anywhere else in this form. For some questions, this approach is advantageous. It can lead to entirely new solutions.

When it comes to hiking trails, however, it's preferable for a chatbot to draw on ideas that someone has already described online. Therefore, Gemini emerges as the winner in this case, although Manus follows closely behind. The paid version of Chat-GPT comes in third.

Incidentally, others have already come up with the idea of using generative AI for hiking routes. The American app Alltrails is supposed to suggest new AI-generated alternatives for hiking trails, for example, optimizing the view or suggesting shortcuts. Hopefully, more reliable technology is running in the background than that used by the chatbots we tested.

An article from the « NZZ am Sonntag »