Interviewing children: using training software to combat false memories

Training with "Virtual Kids": In future, public prosecutors and police
Training with "Virtual Kids": In future, public prosecutors and policewomen will be able to practise their questioning techniques with Matteo, one of the virtual characters.

Questioning children in criminal proceedings is challenging. Carelessly asked questions distort the testimony and can harm the young interviewee. But there is a lack of practice opportunities. An interdisciplinary team from Lucerne University of Applied Sciences and Arts and the ZHAW School of Engineering wants to make this possible in the future with virtual characters and AI.

Statements made by children can be crucial for criminal proceedings. However, children are particularly vulnerable in interview situations, which requires a high degree of care and appropriate training for the interviewer. This qualification has been provided for many years as part of further training at the Lucerne University of Applied Sciences and Arts (HSLU). However, there has so far been a lack of realistic training opportunities for specialists from the judiciary and police - a gap that an interdisciplinary team from the HSLU’s Departments of Computer Science and Social Work and the ZHAW School of Engineering aims to close.

Children are easily influenced

"Children are easily influenced by adults, especially by authority figures such as a policewoman or a public prosecutor," says project leader and HSLU professor Susanna Niehaus. At worst, this can lead to false memories: These are memories of things that seem real but which the child has never experienced. They can not only lead to a misjudgment, but also have fatal consequences for the children themselves. "False memories can cause significant trauma-related disorders and lead to the breakdown of important relationships," says the legal psychologist.

Training on the computer

The training software developed by experts in legal psychology, computer science, game design, animation and computer linguistics is intended to create the most realistic training scenario possible for investigators. This is because training with real children is not possible for ethical reasons. Training is therefore carried out with "Virtual Kids" - a fitting name for the project. Child characters equipped with their own "experience memory" are interviewed via computer. An AI gives different answers depending on the content and type of question. Suggestive questions, for example, also lead to incorrect answers during training that do not correspond to the original memory of the child characters questioned.

Creating AI that expresses itself like a child

Programming an AI for this was no easy task. "A five-year-old child would not use convoluted sentences and the same vocabulary as adults," says ZHAW computational linguistics expert Don Tuggener. The system therefore first had to learn to express itself like a child. In a survey, however, it is not only the answers that play a role, but also the body language. Therefore, a second challenge was to give the characters appropriate facial expressions in response to certain questions, for example when a question triggers stress or fear in the child.

Pilot phase in cooperation with the police

The next stage of the training software will be tested by prospective investigators. The one-year pilot phase will begin in the fall. "We hope that the systematic tests will provide answers to the questions of under which conditions the greatest training effects can be seen and how the improvement in questioning technique can be maintained in the long term," says Susanna Niehaus. The team is also testing whether people would subsequently interview real children better as a result of training with the software.

In future, investigators throughout German-speaking Switzerland could use "Virtual Kids" to train their questioning techniques. The project team is also not ruling out future use in neighboring German-speaking countries - or even teaching the characters other languages.