Go-JuST

Go-Just

AI scenarios - experiment

FPTest05

Methodology

The methodological approach of this study was initially based on the collection and analysis of four distinct texts, each corresponding to a regional unit of Thessaly: Larissa, Trikala, Magnesia, and Karditsa. These texts contained documented descriptions of social and institutional injustices observed in the respective areas and served as the core empirical material for the experimental procedure. Their selection was guided by criteria of representativeness and thematic diversity, thereby ensuring the validity and reliability of the research design.

In the second stage, the empirical material was processed by five different large language models: OpenEuroLLM, Mistral, DeepSeek-70b, GPT-120b, and GPT-5. A common research question was posed to each model, namely the generation of policy and social intervention scenarios aimed at addressing and mitigating the identified injustices. This procedure pursued a dual objective: on the one hand, to exploit the variation in model architectures and algorithmic strategies, and on the other hand, to enable a comparative assessment of the outputs in terms of their theoretical and practical contributions.

Finally, the scenarios produced by the models were consolidated and subjected to a systematic evaluation by the research team. The assessment was conducted on the basis of qualitative indicators such as clarity, innovativeness, and realism, as well as quantitative indicators relating to completeness and structural coherence. Through this comparative process, both the potential and the limitations of large language models were highlighted, particularly with regard to their capacity to support the development of strategies aimed at alleviating social inequalities at the regional level.

Report