Three young hikers were rescued from Mount Shasta in California after planning their ascent with the help of Gemini, Google's chatbot. The Siskiyou County sheriff's office report contains the key detail: the model recommended they carry far less food and water than the group needed.

The chronology speaks for itself. They set off at three in the morning. The mountain's rule is simple and it is on every signpost - if you are not at the summit by noon, you turn back. They reached the summit at seven in the evening. Then they tried to descend in the dark, called the sheriff's office to ask where to go, spent the night in Mud Creek canyon and were brought out the following morning by rangers and volunteers.

Is the chatbot to blame for this? Not entirely, and that is worth saying out loud. Nobody forced them to keep going up seven hours past the cut-off. But an eight-hour ascent that turned into a multi-day ordeal did not go wrong near the summit - it went wrong the moment someone asked how much water to carry and got a confident answer from a machine that has never been there.

"It is always advisable to call the local Mount Shasta ranger station before your trip to get the most accurate information, and never to rely solely on artificial intelligence when planning a trip," the sheriff's office said. The sentence sounds like a banal safety notice. In fact, this is the first time an American rescue service has had to state officially that a chatbot is not a source of mountaineering information.

The problem is not that the model got it wrong. The problem is how it got it wrong - without a single sign of uncertainty. A ranger at a counter will look at you, ask how much water you are carrying and raise an eyebrow. The model gives you the number in the same calm tone it uses to explain a recipe. This summer the Balkans have plenty of their own mountains, their own unmarked trails and plenty of people who type their first question about a route into an app rather than into the phone list of a hiking club.