Artificial Intelligence to redact sensitive information

Overview — This proof of concept examines whether AI can assist with the redaction of sensitive information within Scotland's Children's Hearings System, the statutory system of care and justice that makes decisions about compulsory measures of supervision for children and young people in need of care, support and protection. It forms part of SCRA's wider AI programme, which identifies and tests potential applications of AI in the Children's Hearings System, based on engagement with staff about how AI might address common challenges and free up staff to better support children and families.

How it is used — SCRA safeguards children partly through non-disclosure provisions, which allow information that would otherwise be shared to be withheld from an individual, for example withholding address information where a child's whereabouts must not be disclosed to a parent, by redacting it from materials issued for Hearings. In this proof of concept, the AI scans the text of multiple documents to recognise patterns of words and phrases that should not be disclosed and highlights them as suggestions to the staff carrying out redaction. It draws on the non-disclosure information held in SCRA's case-management system, such as a child's, parent's or carer's name or address, and places such as their school or medical centre, and learns to identify patterns indicating the whereabouts of a child, applying this across other reports and documents.

Human oversight — The redaction itself is not automatic. Staff check the AI's suggestions and remain responsible for reading through reports to identify any information not previously considered for redaction, such as the name of a sports or youth group a young person attends. Overall accountability for redaction remains with human decision-makers.

Assessment and impact — The tool will be tested on 50-100 cases over three months to establish how long AI-based redaction takes, how accurately it identifies the required information, the number and causes of any data breaches, and staff confidence in using it. As with the related proof of concept, a wider research project is gathering the views of children, families, carers, social workers, safeguarders, advocates, solicitors, SCRA staff and panel members on the ethics, rights and legalities of using AI, and equality and human rights, children's rights and data-protection impact assessments have been undertaken.

• Better protects child safety by reducing the risk that sensitive details are missed during redaction: information such as a child's home address, school, or the sports and youth clubs they attend, any of which could reveal where a child lives.
• Reduces the time and manual effort of redaction, which is currently done by staff carefully reading each report and using software to black out words, and is highly labour-intensive across documents that can run to hundreds of pages.
• Helps SCRA cope with high and complex volumes: more than 71,000 instances in the last year, often with multiple documents per child and additional documents where siblings are subject to legal orders.