ľ�ϸ���Ӱ��

Have you been scammed on Marktplaats, or bought something online that was never delivered? If you've reported fraud online, there's a good chance you used the assistance tool developed by Daphne Odekerken. She will defend her PhD thesis on 6 January, focusing on the design of an AI system that simplifies filing a complete report.

The police receive hundreds of online fraud reports every day. The problem, however, is that people often don’t know what information is required to complete the form, explains PhD candidate Daphne Odekerken. “Sometimes the police have to exchange emails back and forth to fill in the missing details, which causes significant delays.”

To make the process more efficient and reliable, Odekerken contributed to the development of a decision-support tool. “It’s like an intelligent reporting form. It can automatically identify missing information in the text provided by a citizen and ask targeted questions. For instance, if someone states, ‘I didn’t receive anything,’ but doesn’t mention whether payment was made, the AI system—a type of chatbot—will ask, ‘Did you pay?’ This ensures the report is complete before it is submitted.” With the assistance tool, citizens receive clarity more quickly, and the police can focus on cases that genuinely need investigation.

Explainable outcomes

The assistance tool is based on rule-based AI. This means it operates using predefined rules, such as legal articles. “A rule-based system only makes decisions when it has gathered enough information to draw a conclusion.” According to Odekerken, this approach has a key advantage: it makes outcomes explainable. “That’s critical for applications like police reports. You can demonstrate that the system works correctly, which builds trust among both citizens and police officers. That’s why I’m not a fan of self-learning AI. It’s essential to understand precisely how systems work and why they arrive at specific conclusions.”

Rule-based AI can only reason with complete information, which is a challenge in the context of police reports, as these are often incomplete. “For developing the tool, we couldn’t rely on existing rule-based AI techniques. We had to create a system capable of reasoning with incomplete information—in the sense that it can identify what it doesn’t yet know.” Odekerken developed new algorithms to determine what additional information is relevant and retrieve it by asking the citizen targeted questions. “Only when there’s no more relevant information to gather does the tool provide a substantiated recommendation.”

It’s essential to understand precisely how systems work and why they arrive at specific conclusions

In addition to the assistance tool, Odekerken is working on general methods that can be applied to other areas of police work, such as assessing unreliable webshops. “Here, the AI system analyses characteristics of webshops previously deemed reliable or unreliable. If a new webshop scores better than one classified as legitimate, we can consider it trustworthy too.” If the system finds no comparable cases, the final decision remains with a human. “For complex decisions, human oversight is always necessary. It’s vital that AI supports humans, not replaces them.”

Odekerken conducted her research within the National Police AI Lab, a collaboration between the police, Utrecht ľ�ϸ���Ӱ��, and the universities of Amsterdam (UvA) and Delft (TU Delft). In this lab, PhD candidates work simultaneously for both the police and the university. This dual role suits Odekerken perfectly: “I was looking for something meaningful, and in this role, I can thoroughly explore a topic while ensuring it has a clear practical application.”

After completing her PhD, she plans to continue working with the police while spending one day a week as a guest researcher at Utrecht ľ�ϸ���Ӱ��. “I want to do something that isn’t just theoretical but also practical and makes a difference for people. At the same time, I value staying connected with researchers and students to keep up with developments in the field.”