AI-enhanced ethical decision-making in social work: development and testing of an application-specific language model

We develop and test an application-specific language model for supporting ethical decision-making in social work. The model is built using Gemini 2.5 pro with seven agents, six of which are responsible for information aggregation of possible actions, relevant ethical issues, principles, standards, laws and values while the seventh is responsible for summarising the information and recommending and justifying issue-related decisions. We further test the model via 70 simulated cases. In evaluating the appropriateness of social workers’ case-specific actions, the model’s judgements showed 80% alignment with the answers from the casebook. It also provided sufficient reasons or appropriate remedial plans in nearly half of the cases. Errors were due chiefly to inaccuracies in understanding legal compliance requirements, inadequate consideration of case-specific situations, and logical fallacies. Conversely, in some cases the model provided responses that were more appropriate than those from the casebook. Accordingly, we see considerable scope to apply artificial intelligence in supporting social work ethical decision-making.