About the Project

FairLegal is an ambitious and pioneering research initiative focused on developing a Large Language Model (LLM) powered by Artificial Intelligence (AI), capable of identifying and eliminating gender bias in legal and administrative texts across the public and private sectors.

The project addresses the growing demand for fair and unbiased technological support in socially significant areas such as jurisprudence, recruitment, and public procurement. As AI becomes increasingly embedded in decision-making and daily life, the development of trustworthy and ethically designed systems becomes critical. FairLegal fills this gap and actively contributes to the EU’s goal of “Trustworthy AI,” one of the seven pillars of the European regulatory framework.

FairLegal combines Natural Language Processing (NLP), deep learning, and Fair AI theories to identify and reduce bias in both data and algorithms.

The project follows a dual approach: first, mitigating bias at the data level using word embeddings and knowledge graphs; second, performing fine-tuning of an LLM on bias-reduced datasets, followed by prompt engineering. This two-step strategy is groundbreaking, as it combines cutting-edge technologies and extends beyond typical one-stage methodologies. Tools like BERT, spaCy, RDF/OWL, and GANs enhance the innovative nature of the project.

The project focuses on two pilot scenarios: (a) legal texts from the public sector (e.g., case law, legislation, circulars), and (b) job or project announcements from public and private entities. In each case, data will be collected and analyzed from specialized sources (such as the National Printing Office, Kariera.gr, Skywalker, Indeed), and a structured database with metadata will be created. The project’s goals include developing technology to detect biased language, producing a functional AI system capable of correcting such phenomena, and delivering a market-ready product for public bodies, companies, and organizations.

The project is implemented by a consortium of three partners with high expertise: the University of the Aegean (i-Lab), ITML, and Lioncode. Roles are clearly defined based on scientific and technical specialization: the University leads coordination and research strategy, while the companies focus on technological development and implementation. The project follows PRINCE2 methodology, using a three-level governance structure that includes an ethics and compliance committee. Expected outcomes are broad: institutional, societal, technological, and business benefits – both for the consortium and for stakeholders in the public and private sector.

FairLegal will carry out the following key actions to achieve its ambitious goals:

• Collect and organize large volumes of legal and administrative texts from public and private sources.
• Analyze how gender bias appears in official texts.
• Build a structured dataset capturing biased content to support effective model training.
• Apply advanced preprocessing and data augmentation techniques to the collected texts.
• Develop and fine-tune word embeddings (WEs) and knowledge graphs (KGs) to detect and reduce bias at the semantic level.
• Perform fine-tuning of a Large Language Model (LLM) using bias-free data.
• Enhance the trained model using prompt engineering techniques for further bias mitigation.
• Pilot the system on two scenarios: (a) public legal texts and (b) public-private sector announcements.
• Continuously evaluate performance throughout the development lifecycle.
• Deliver a market-ready product and define a dissemination strategy targeting public and private organizations.