Post by Università Bocconi

What if the real power of #AI assistants lies not in the answers they give, but in the questions they ask? #AIassistants do more than respond to users, as Assistant Professor Martino Banchio explains. They structure conversations, select options to show, decide their order and, often before we notice it, shape the space in which decisions are made. Think of a simple restaurant recommendation. An assistant may ask whether you prefer meat or fish, whether you want something elegant or informal, and what budget you have in mind. Each question seems neutral. Yet, by deciding which questions come first and which options appear later, the assistant is already guiding the choice.    This mechanism can be useful when interests are aligned. If the assistant is truly trying to identify the best restaurant for the user, its ability to narrow down options becomes a strength. The problem emerges when incentives diverge. A platform that monetizes through advertising may not have the same objective as the person asking for advice. A system designed to maximize engagement may benefit from prolonging the conversation rather than helping the user reach the best decision. The key issue is not that #algorithms choose for us. The final decision remains with the user. Their power lies in controlling which alternatives are visible, in what order and through which sequence of questions. This is why conversational architecture matters.    Banchio’s analysis also highlights a counterintuitive point: randomization can make these systems more effective. An algorithm that always follows the same logic is fragile, because users have different and often unknown preferences. Like a waiter who always recommends the most expensive dish, it may persuade some people and lose others. Varying the order of questions and recommendations can help the system work across users. This has important implications for #AlgorithmicTransparency. A fully predictable algorithm may seem easier to scrutinize, but it can also become less adaptable to the unpredictability of human preferences. The question is not only what content algorithms produce, but who designs the structure of the interaction, with which objectives and constraints.    The issue extends beyond restaurants. When an AI assistant orders sources for a question about current events, it influences how opinions are formed. When it ranks products in an online shop, the order of presentation already becomes persuasive. Regulating algorithmic content remains necessary, but it is not enough. As #ArtificialIntelligence becomes more embedded in everyday decisions, we also need to examine how questions are framed, how options are ordered and how conversational paths are designed.    The next time an AI assistant asks what you prefer, the most important choice may already have been made: which questions to ask, and in what order.