Using Artificial Intelligence to Improve the Use of Evidence in Social Work Practice: A Rapid-Cycle Development Project

Research question: Can a bespoke Artificial Intelligence (AI) tool help social workers and managers to make better use of research evidence in children s social care? Background: Like many professionals, social workers rarely use research in practice–finding and understanding relevant texts takes too long, and applying complex findings to individual cases is theoretically and practically challenging. AI excels at rapidly summarising high-volume information, and anecdotally social workers already use it. AI will change the evidence-practice relationship–yet non-trivial challenges and risks are manifest. Existing AI is not specifically trained on social care data and does not always produce ethical or useful summaries. Furthermore, workers needs of AI, and how they might use (or misuse) it in practice, is currently unknown. Aims: To understand social workers evidence needs, to support the development of an ethical, reliable, and helpful AI tool (the 'Social Worker Evidence Engagement Tool', SWEET) to facilitate evidence use in practice. Objectives: To: Understand workers perspectives and priorities for accessing and utilising evidence. Identify situations and purposes for which social workers want to access and apply evidence. Test example social worker queries and expert and AI responses, to inform tool design. Develop key design concepts and features to support the mitigation of bias and errors, include public voice, and facilitate responsible use. Collate credible, relevant sources of social care evidence in a custom-built database. Use outcomes of the above to create SWEET. Pilot and refine SWEET. Evaluate the accuracy and feasibility of SWEET. Methods: We will work in close collaboration with our partner local authority (Neath Port Talbot) through an iterative rapid-cycle 3-stage design ethnography. Stage 1 will investigate social workers evidence needs and use of evidence in decision-making. Stage 2 will develop SWEET, explore how it is used and improve it. Stage 3 will test the final version of SWEET, to understand its utility and limitations. Ethnographic methods, including observations and interviews, will be used to understand the needs of workers and managers, and their uses of evidence. Focus group discussions will elucidate views of parents and young people with lived-experience of children s social care. We will analyse these data, using grounded theory and thematic analysis, to inform development of design concepts. SWEET will be a Retrieval-Augmented Generation system, drawing on a high-quality social care evidence database. SWEET will be trained on responses to social worker evidence queries written by a panel of academic experts, and social workers quality ratings. Model performance will be refined and evaluated using an automated Retrieval-Augmented Generation Assessment framework, whereas humans will evaluate factual accuracy, quality and user experiences. Timelines for delivery: The study will be delivered over a period of two years. Anticipated impact and dissemination: We will disseminate findings of evidence needs and responsible AI in children s social care with practitioners, policy makers, and the public, through webinars, blogs and AI Summits . Ultimately, this research seeks to produce a not-for-profit AI tool to directly improve social workers use of evidence in practice.