In September 2025, two hikers had ChatGPT plan a route to the Schmalhorn, a 1,953-meter-high peak in the Allgäu Alps. Instead of the well-known approach from Einödsbach, the AI suggested a route via Spielmannsau, along an old hunters’ trail. About 100 meters below the summit, they were unable to proceed any further, and one of them suffered a panic attack. The mountain rescue team had to climb up to them and fly them down by helicopter. “In fact, there is no official trail there,” David Lacher of the Oberstdorf Mountain Rescue Service later told the Allgäuer Zeitung. Was this an isolated incident, or is it the norm when you let AI plan the hike?

Why Komoot’s AI Shouldn’t Be Allowed to Make Things Up

Since February 18, 2026, Komoot has been integrated into ChatGPT as a standalone app, accessible via the tag “@komoot”: A phrase like “a smooth three-hour loop with views in Eastern Switzerland” is all it takes for the AI to search Komoot’s database of over 7 million routes that have already been completed. What used to require several filter clicks can now be described in a single sentence.

The difference from the Schmalhorn incident is this reliance on real map data. David Lacher of the Mountain Rescue Service warned after the incident that anyone unfamiliar with an area should never rely solely on AI. Komoot’s routing technology, on the other hand, compiles real, mapped sections of trails for new suggestions—not made-up routes. The German Alpine Club also advises combining digital route suggestions with a map and compass.

According to its own statements, Strava has been structuring its AI route suggestions in a similar way since May 2025, using the movement data of millions of athletes. The same limitation applies to both companies: The AI may only use what it finds in Komoot’s or Strava’s own database. It is not allowed to invent anything on its own. While errors can still occur here, the “hallucination” problem can nevertheless be mitigated. But what if a route still leads someone astray and they take a fall? Can the AI then be held liable? No. Artificial intelligence is not a legal entity. It cannot stand trial and cannot be held liable for any damages.

When no one takes responsibility

Only a person or a company can be held liable. The Hamm ruling from February 2026 demonstrated what happens when such a responsible party exists. On a mountain bike flow trail, a rider fell at a blind spot where the correct line was not visible and which the operator had secured only with warning signs and loosely hanging caution tape. The Hamm Higher Regional Court awarded her 5,000 euros in compensation for pain and suffering and split liability equally between the two parties. The operator was held liable because it had not adequately secured the hazardous area. The rider was also held liable because she failed to adjust her speed despite the warning signs. The court was able to rule this way because there was a responsible party: the club that operated the trail and could have secured it better.

On the other hand, someone who describes a trail in a mountain bike guidebook or suggests a route is unlikely to face legal action, because the rider must assess for themselves whether to follow it. This personal responsibility is even more critical in the mountains, and it applied long before AI. As it stands, an AI suggestion is merely advice like any other. That will change as of December 9, 2026. Starting then, under a new EU regulation, software and AI will be considered products—and a product can be defective. Anyone who incorporates it into their offering is liable for its errors just like a manufacturer. Whether a route that leads nowhere will be considered a product defect in court remains to be seen. Until then, the driver alone is liable for the AI’s misstep.