An experiment involving six AI agents produced strikingly similar recommendations for a two-week trip through China, raising questions about whether automated travel planning could amplify the same “Lonely Planet effect” that sends travelers to the same destinations, stores and restaurants. The test also examined how the agents handled a late-night request from a user struggling with insomnia.
Six AI agents, one China itinerary
The test asked each AI agent to create a first-draft itinerary for a possible two-week adventure in China late the following year. The brief included flights, hotels, nightlife and dining recommendations, while asking for a route that combined historical attractions with the atmosphere of a major city. The traveler wanted to use no more than four hotels because frequent changes of location were undesirable.
Remarkably, every agent proposed essentially the same route through China: Beijing, Xi’an, Chengdu and Shanghai. The suggested journey began with imperial history, continued with Sichuan cuisine and pandas, and ended with Shanghai’s nightlife and metropolitan energy.
Muse
Muse: 5/5. Muse produced an itinerary that was easy to read, supported by visual charts and practical, easy-to-follow advice.
“We will leave SFO the week after Golden Week, take a direct flight to Beijing, and then take a direct flight home from Shanghai. Four hotels, no turning back. Start with imperial history, middle with Sichuan cuisine and pandas, and end with Shanghai nightlife.”
Dot
Dot: 4/5. Dot’s draft was strong but excessively long, reaching 24 pages.
AI agent snippet: “My recommendation is to fly into Beijing, take the high-speed train to Xi’an and Chengdu, then fly back to Shanghai and return home. Enjoy imperial history, terracotta warriors, Sichuan cuisine, and a metropolitan finale without overnight trains or additional hotel transfers.”
Grokbot
Grokbot: 4/5. Grokbot’s output was readable and filled with practical tips and suggestions.
“The first option is a 13-night trip for two people to Beijing, Xi’an, Chengdu, and Shanghai, departing from SFO on October 13th and returning to Japan on October 27th after Golden Week.”
Gemini Spark
Gemini Spark: 3/5. The itinerary was acceptable, but it offered less detail and was harder to read than some of the other drafts.
“A route for first-time travelers that combines imperial history, Sichuan gastronomy, vibrant nightlife, and the comfort of four hotel hubs.”
Instinct
Instinct: 3/5. Instinct’s itinerary was decent and useful, although its layout was less appealing than those of several other agents.
“It is headquartered in Beijing and headquartered in Shanghai, with Xi’an and Chengdu in between. With four locations, there’s no turning back. History moves forward and the food and bars get better each time I go.”
Claude Duet
Claude Duet: 3/5. Claude Duet created a solid itinerary, although its output did not include images or other visual elements.
“It will arrive in Beijing on Saturday, October 9, 2027, then travel to Beijing → Xi’an → Chengdu → Shanghai, and return from Shanghai on Saturday, October 23. That’s 13 nights in 4 hotels, 2 bullet trains and 1 short flight.”
Testing the agents during insomnia
The late-night test followed an episode of poor sleep, although Vietnamese coffee may also have contributed. It explored what happens when AI tools receive broad, open-ended requests—situations in which people may need the most help but may not know where to begin.
A human assistant might be able to respond to messages during the night, and the test examined whether digital helpers could offer a similarly appropriate response. Two agents, Gemini Spark and Claude Duet, appeared to lack the sensitivity and situational context expected from a real person.
Grokbot
Grokbot: 5/5. This response felt coherent and practical enough that the tester said they would probably follow it if the situation were real.
“I’ll set a reminder to check on your mom later this morning to see how her legs are doing, so there’s nothing to worry about for now.”
Dot
Dot: 4/5. Dot and Gemini Spark offered similar suggestions, essentially advising the user to write down what was on their mind. Dot’s wording felt more collaborative and direct.
“Send me three things that are taking up space in your head. It doesn’t matter if it’s messy. Keep it a short checklist with one clear first step. After that, you can put down the phone.”
Instinct
Instinct: 4/5. Compared with some of the agent’s other responses, this suggestion seemed reasonable and useful.
“Here’s what you can do from bed. Please take a quick look at the Chinese draft and tell us one thing you would like to change. I’ll have v2 ready when you wake up.”
Muse
Muse: 4/5. Muse delivered another strong response. The agent appeared designed to improve over time and become better aligned with what helps individual users.
“Check today’s calendar against the email reminders you sent yourself and look for any discrepancies or items you shouldn’t have missed. This is a check that should be done properly yesterday and should only be shown at appropriate times unless there is an emergency.”
Gemini Spark
Gemini Spark: 2/5. The general advice could be useful for someone dealing with sleep problems, but comments about an empty morning calendar were irritating. Insomnia is not necessarily related to Google Calendar events.
“The easiest task at the moment is to get what’s looping in your head down on paper so your brain can let it go. Especially today, my calendar is completely clear and I don’t have any morning appointments waiting for me, so I have to be careful.”
Claude Duet
Claude Duet: 2/5. This was one of Claude Duet’s weakest responses. Although the two tasks were technically unfinished items, ordering lunch or reviewing work projects did not seem especially relevant in the middle of the night.
“Let’s close up two loose ends from yesterday. Thursday’s pitch was delivered and the lunch cart was covered with a DragonEats order. Please say the words. Mark both as complete. Then go back to sleep.”
OpenAI Dot delivered the strongest overall performance
Dot remained behind an expensive paywall, so most people may want to wait until OpenAI releases the agent for free or offers it at a price of at least $20 per month. Despite weaknesses on tasks closely related to the tester’s area of expertise, Dot performed best overall in these tests.
A month after its release, Dot already felt more polished. The tester preferred its separate app to the current ChatGPT integration, although the smartphone widget offered a useful compromise.
Meta Muse was the most accessible
Muse performed well in the tests and was the easiest agent to set up. It was also free. Meta says the service handles user information responsibly in secure virtual machines, but concerns remain because of the company’s previous mishandling of user data.
Claude Duet offered the best interface
Claude Duet was tested as a research preview just one day after its official release, so some rough edges were expected. Although it received middling scores in most tests, its dedicated app was easy to use for navigating and following up on ongoing tasks.
Users focused on productivity and streamlined software experiences may become particularly interested in Claude Duet in the coming months. Access currently requires joining Anthropic’s waiting list, and the service has a $20 monthly fee.
Instinct created the most frustration
Some users may appreciate Instinct’s sparse design and full integration with Messenger, but that approach did not work well in this test. Other WIRED writers have praised Instinct enough to say they are willing to overlook privacy concerns, while the text-only experience felt limiting here.
The tester came away feeling too removed from the agent’s actions and was frustrated by an emoji reaction—described as resembling a grimace—after reporting a missed scheduling conflict.
Source: www.wired.com