The Pulse
ChatGPT Hallucinations Put Customers at Odds With Service Workers
The Verge reports that customers are bringing AI-generated errors into restaurants, parks and childcare conversations. Workers say some customers trust ChatGPT or Claude over staff who know the menu, terrain or child in their care.
AI.info Team · Written with AI ·
A New York City server says diners with shellfish allergies have challenged her warnings about dishes after consulting ChatGPT. The broth accompanying one fish dish contained shellfish, she told them, but some customers trusted the chatbot’s contrary answer. “You’re allergic to shellfish, and ChatGPT says there’s no shellfish in this, and we are telling you, yes there is!” Madison, whose surname The Verge withheld, told the publication.
The exchange captures a pattern in The Verge’s report by Gaby Del Valle, published October 2: workers in customer-facing jobs say people increasingly treat AI-generated answers as authoritative, even when those answers conflict with what staff can see or verify. The accounts span restaurants, parks, childcare and retail. They are interviews with individual workers, not a measure of how widespread the problem is.
When ChatGPT invents a wine
Madison, who worked at a Michelin-recommended Italian restaurant, told The Verge that diners sometimes dismiss sommeliers in favor of ChatGPT. Some then ask for bottles the restaurant does not stock—or wines that do not exist. To avoid embarrassing customers, she and her colleagues may suggest that a recommendation reflects a wine list the restaurant does not carry, rather than say the chatbot fabricated a brand.
The friction goes beyond wine. Del Valle says movie theater workers, baristas and Apple Store employees described customers asking for products or features that do not exist. Those accounts point to a practical problem for service staff: correcting an AI-generated answer can become an argument about whose information deserves trust.
Bad itineraries, defended against the map
A river ranger identified by the pseudonym Jason told The Verge that visitors have arrived with AI-generated itineraries that place campsites in the wrong locations or arrange stops in an impossible order. Some routes would require people to row upstream on a river where that was not possible. Jason said groups remained skeptical even after rangers explained the problem while pointing to a map.
In New York, an after-school teacher identified as Casey described a family relying on a ChatGPT-generated bedtime schedule. She believed the child needed two to three more hours of sleep each night, based on his behavior in the program, but said the parents insisted the schedule was fine because ChatGPT had approved it. Casey told The Verge that conversations become difficult once parents have decided the AI is the more authoritative source.
AI agents can muddle the reservation
Lauren Wilford, who has worked in food service for 15 years, said customers first used ChatGPT or Claude to interpret menus and choose wine pairings. Over the past year, she said, her restaurant has also received calls and emails from agents trying to book on customers’ behalf. The technology can make ordinary coordination less clear when workers cannot tell whether a request reflects what the customer actually wants.
Wilford described a client who made last-minute requests for a podium, audiovisual equipment and catering changes ahead of an event at the Providence, Rhode Island, restaurant she manages. The next day, the clients arrived without mentioning those requests. Wilford suspects an AI agent may have made them, but said she cannot prove that was the case.
“I’m sympathetic to the reasons people use AI agents,” Wilford told The Verge. She drew a distinction between administrative tasks and situations where two people need to work out details together, warning that customers may not notice what gets lost when they try to skip that exchange. In her account, the immediate cost of an AI mistake falls on workers who must correct it—and on customers who may act on advice that is wrong.