Human-Centered AI for Migrant Integration Through LLM and RAG Optimization

The enhancement of mechanisms to protect the rights of migrants and refugees within the European Union represents a critical area for human-centered artificial intelligence (HCAI). Traditionally, the focus on algorithms alone has shifted toward a more comprehensive understanding of AI’s potential to shape technology in ways which better serve human needs, particularly for disadvantaged groups. Large language models (LLMs) and retrieval-augmented generation (RAG) offer significant potential to bridging gaps for vulnerable populations, including immigrants, refugees, and individuals with disabilities. Implementing solutions based on these technologies involves critical factors which influence the pursuit of approaches aligning with humanitarian interests. This study presents a proof of concept utilizing the open LLM model LLAMA 3 and a linguistic corpus comprising legislative, regulatory, and assistance information from various European Union agencies concerning migrants. We evaluate generative metrics, energy efficiency metrics, and metrics for assessing contextually appropriate and non-discriminatory responses. Our proposal involves the optimal tuning of key hyperparameters for LLMs and RAG through multi-criteria decision-making (MCDM) methods to ensure the solutions are fair, equitable, and non-discriminatory. The optimal configurations resulted in a 20.1% reduction in carbon emissions, along with an 11.3% decrease in the metrics associated with bias. The findings suggest that by employing the appropriate methodologies and techniques, it is feasible to implement HCAI systems based on LLMs and RAG without undermining the social integration of vulnerable populations.