The Power of Framing Questions

Try asking an AI assistant to recommend a restaurant. Instead of answering, it will ask you if you prefer meat or fish, if you are looking for something elegant or informal, if you have a budget. All reasonable questions, and yet, in the end, you will have the distinct feeling that it was not you who chose the restaurant. It was the assistant, with its incessant questions, that has chosen for you.

We often focus on the quality of answers, but today the real power lies in the questions. In a recent study with colleagues from Stanford and Reddit, we build on this insight that is familiar to many users of ChatGPT and similar tools. AI assistants do more than provide answers: they channel, select what to show us, in what order and how to structure the space of possibilities. And this power of steering is enough to shape the outcome of our conversations and our decisions.

So far, nothing alarming. On the contrary, if AI truly has our interests at heart, this mechanism is remarkably effective. An assistant seeking our ideal restaurant is a powerful ally. But can we be sure that AI’s objectives align with our own? Do those who monetize through advertising have an interest in showing us the best restaurant or the one that has paid the most? Are platforms that benefit from user engagement interested in helping us reach the right choice or prolonging the conversation? The same mechanism that, when interests are aligned, guides us towards the best decision can, when they diverge, lead us elsewhere.

Controlling the Options

However, AI assistants cannot choose for us. Even when interests are misaligned, the final decision remains with the user, who evaluates and selects from the options presented. The power of these platforms lies in controlling which options to present and in what order. This power has clear limits, forcing AI to decide how to steer a conversation without fully knowing the preferences of the person it is engaging with. 

Randomization

For this reason, our research shows that an algorithm that always follows the same logic is fragile. Consider a waiter who always starts by recommending the most expensive dish: some customers may be persuaded, while others will simply disregard the suggestion and choose independently. The solution, counterintuitive as it may seem, is randomization. Varying the order of questions and recommendations is essential precisely because the user’s preferences are unknown. In this context, randomness is not a flaw; rather, it is the only way to ensure a minimum level of effectiveness regardless of the interlocutor’s tastes.

This has direct implications for the call for full algorithmic transparency, a topic that dominates public debate. Our research suggests that a fully predictable algorithm is also systematically fragile, unable to adapt to our unpredictability. The implications extend far beyond the restaurant example. When an AI assistant selects and orders sources to answer a question about current events, it is shaping the path through which we form our opinions. When it recommends us products in an online shop, the order of the presentation already acts as a persuasive device.

A Choice Made at the Start

Regulating the content produced by algorithms is necessary, but not sufficient. It may be even more important to ask who designs the architecture of conversations, with what objectives and under what constraints. And this must be done with the awareness that a degree of unpredictability is an essential condition for their proper functioning.

The next time an AI assistant asks whether you prefer meat or fish, remember that the most important choice has already been made: which questions to ask you, and in what order.