AI systems such as chatbots are increasingly deployed in sensitive sectors including public services and healthcare. To be trustworthy, these systems must not only avoid unsafe behaviour but also respond in ways that align with human values such as safety and autonomy. That places specific demands on how such systems handle the diversity of needs and preferences among users.

A structural challenge is that AI systems learn from vast quantities of online data, which causes them to reflect the most common opinions and ways of thinking rather than adapting to the individual. In everyday situations this can result in a generic answer that does not quite fit someone’s circumstances. In more sensitive contexts, the consequences can be more serious. Ahead of the most recent Dutch elections, the Dutch Data Protection Authority observed that popular chatbots paid disproportionate attention to the largest political parties while smaller parties were rarely mentioned.

Liscio argues that the key to addressing this problem lies in recognising human values. Values such as autonomy, safety, fairness, and respect for tradition play an important role in the choices people make, yet they are difficult for AI to identify. “Values differ from person to person, change depending on the situation, and are often not expressed explicitly,” Liscio says. “People are also not always aware of their own values until they are prompted to reflect on them.”

In earlier research, Liscio developed a prototype together with an interaction designer to explore how users could reflect on and communicate their values through chatbot interfaces. Building on those insights, his Veni project will develop new methods that enable language models to infer users’ values from conversations and generate responses that better match what matters to them, while keeping the system transparent and leaving users in control. The research combines technical AI development with experiments involving real users.

Liscio will test his approach in two application areas. The first is voting advice: political preferences are often linked to different value priorities, and a better understanding of those values can help provide more transparent guidance that supports people in reflecting on their own preferences. The second focuses on lifestyle coaching for people with diabetes, where personal trade-offs between health, independence, and daily routines are central. An AI system that understands these differences can offer more targeted and relevant support in both contexts.