Risk case for Google's hotel booking agent spans accountability, bias and workforce concerns, not just a taxonomy and a 2022 tribunal ruling

The firmest evidence in the account is a tribunal decision. In 2022, Air Canada's chatbot incorrectly told a passenger he could retrospectively claim a bereavement fare ^[6]. The airline later refused the refund and argued before a Canadian tribunal that the chatbot was effectively responsible for the information it had provided ^[7]. The tribunal rejected that argument and held Air Canada responsible for information presented through its website ^[8].

That system answered questions. The authors' argument is that accountability was already contested at that level, and that the stakes rise once systems take actions on a customer's behalf ^[9].

The rest of the piece is framework. The distinction it draws is between reactive systems, which answer a question or return a list of suitable hotels, and agentic ones, which operate with greater autonomy, adapt to changing circumstances and coordinate actions across systems with less continuous human direction ^[15]. Five roles is a research taxonomy, useful for deciding what to test; the testing still has to be done ^[5]. The article reports no error rate for AI Mode's hotel picks, and nothing on how the system weights price against review text and cancellation terms ^[17]. The claim that the ranking criteria sit outside both the traveler's and the hotel's control is therefore an argument about how such systems are built.

Two of the five named partners, Booking.com and Expedia, already run comparison engines; the other three sell their own rooms ^[10]. Both groups are now feeding inventory into a comparison step that a third party runs.

The bias examples are specific and unquantified. A voice-based concierge may understand some accents more reliably than others; a recommender might assume an older traveler wants only sedentary activities, or infer dietary or cultural preferences from crude proxies instead of what the guest requested ^[11]. The authors say hospitality research has already identified these risks ^[12]. Their narrower point about autonomy is the stronger one: a traveler can reject a recommendation, and a system that acts can carry the same assumption into which offers are presented and how complaints are handled ^[13].

The measurement that would settle the argument is simple. One property's booking log over a quarter would produce it: of bookings completed inside the assistant, the share that met every constraint the traveler stated, set beside the share a human agent working the same brief hit.