Xiaoning handed over her four-day, three-night Seoul trip plan entirely to Gemini.
On the first night, Gemini recommended that she visit a 70-year-old samgyetang (ginseng chicken soup) restaurant in Myeongdong, giving the address down to the exact street number. According to the AI’s description, there should be an elderly lady working there, and it earnestly suggested she order the samgyetang set with a cup of omija tea.
She took a taxi there and stared at the address for five minutes, only to find that it was now a cosmetics store. When she asked the staff inside, they said the restaurant had been closed for over two years.
The next morning, Gemini suggested she start with mung bean pancakes at Gwangjang Market for breakfast, saying that she could then conveniently stroll over to Tongin Market. But when she got there, she realized the two markets were four kilometers apart, and required a bus transfer—hardly convenient.
Later, she searched Xiaohongshu and found that, over the past three months, many travelers using Gemini for outbound trips had run into similar problems. European and American backpackers complained that it invented Japanese train schedules, while Southeast Asian bloggers noted that about half the restaurants it recommended had already closed.
What AI really appears to do is scan the keywords in your query, look for the terms that most frequently appear together in its dataset, and then output the highest-frequency combination.
That may be why it suggested eating breakfast at Gwangjang Market and then “conveniently” visiting Tongin Market—because in the travel guides it had read, the two market names may often appear in the same article. But the four-kilometer distance and the need for a bus transfer? That is where it falls short.
AI knows how to make recommendations. But human behavior is far more complex. If your prompts are even slightly unclear, the itinerary it generates can go wildly off course.
