AI-informed avatar to improve child protection

Publicly released:
Australia; QLD

An avatar collaborative project involving Griffith University researchers, led by SimulaMet and OsloMet in Norway, will use AI to improve child protection.

Media release

From: Griffith University

An avatar collaborative project involving Griffith University researchers, led by SimulaMet and OsloMet in Norway, will use AI to improve child protection.

The system has been designed to improve interviews by police and child protective services working with maltreated children, using an avatar to train users in effective investigative interviewing techniques.

The avatar is a young girl, who responds in real time, on a case-by-case basis, with her responses informed by thousands of transcribed human interactions.

She’ll also be able to change with age and speak any language.

The development of the avatar’s response style is being overseen by Professor Martine Powell of Griffith Criminology Institute’s Centre for Investigative Interviewing.

She is a world-leading expert in investigative interviewing, focusing on obtaining accurate and detailed information from vulnerable interviewees and questioning about sensitive topics.

Professor Powell said the biggest gap in investigative interviewing of children is adequate training.

“One of the most important components is simulated practise interviews, but they’ve historically been difficult to stage and can be very costly,” she said.

“Our team excel in conducting this training, with post-doctoral staff acting as the child and providing feedback to those being trained, but these sessions need to be timed and staged, plus they come with the mental health impacts of having to consistently play that role.

“This is the first dynamic avatar of its kind in the world, with the child able to respond instantly without human prompting or intervention, and rewarding well-structured or open questions with more information.

“Having a standardised tool that’s able to facilitate this at any time of day or night, in a way that promotes learning and improvement is going to boost the efficiency of training worldwide.”

With child abuse being a global concern, studies have estimated 22.6 per cent of children experience physical abuse and 11.8 per cent are subjected to sexual abuse before the age of 18, with children being the only witnesses for 70 per cent of sexual abuse cases.

It’s also been found that children are reliable witnesses when interviewed properly and in line with best-practice recommendations.

Researcher and Project Manager Gunn Astrid Baugerud of OsloMet’s Faculty of Social Sciences said the impact of investigative interviews by police and Child Protective Services on abused children could be profound, making effective training vital.

“In Norway we have world-leading experts in AI at SimulaMet, and we’ve been working closely with a team of them as well as Microsoft to harness GPT-3 capabilities since January 2022 – long before it was launched to the public.

“We’re using what’s called a multi-model avatar which has different modalities that need to be integrated.

“We’ve trained our system based on thousands of human interactions, modifying and refining the ‘brain’ of the avatar to create a virtual child who can act as a real child would in that situation.

“There will also be different versions of the child, so someone who is less experienced and needs to strengthen their skills in asking open-ended questions could use one version, and more experienced practitioners would have a more challenging experience, with an avatar disclosing fewer details.

“Some abused children will be very talkative and others will not be – so we try to prompt ways of adhering to best practice to get the best results.”

The avatar project has been financed by the Research Council of Norway, OsloMet and SimulaMet as well as a foundation in Norway for children at risk in partnership with the Centre for Investigative Interviewing.

Journal/
conference:
Organisation/s: Griffith University, OsloMet, SimulaMet, Research Council of Norway
Funder: This project was funded by the Research Council of Norway, OsloMet and SimulaMet.
Media Contact/s
Contact details are only visible to registered journalists.