As artificial intelligence becomes an integral part of how institutions draft, analyze, and manage official documents, concerns around fairness and discrimination are gaining increasing attention. Language technologies, and especially large language models (LLMs), can unintentionally reproduce gender bias that exists in historical data and established institutional practices. addressing this challenge is at the core of Fair-legal, an applied research and innovation project that aims to develop ai systems capable of identifying and eliminating gender bias in official texts.

Fair-legal focuses on two areas where language has a direct and tangible impact on equality and access to rights and opportunities. the first concerns public-sector legal texts, including legislation and jurisprudence, where even subtle linguistic bias can influence interpretation and long-term outcomes. The second targets public and private sector job notices and announcements, where gendered language can affect who feels encouraged to apply and who is excluded before the selection process even begins.

The project develops an end-to-end approach that goes beyond simple text correction. At its core is the fine-tuning of a large language model specifically adapted to these domains. Rather than addressing bias only at the final output stage, Fair-legal intervenes earlier in the AI pipeline, examining how bias is encoded in language representations and reducing it before the model generates text. This multi-layered strategy allows the system to produce gender-fair alternatives while preserving the meaning, tone, and formal structure required in legal and institutional communication.

A key methodological innovation of Fair-legal lies in its treatment of language data. the project combines advanced text pre-processing and data augmentation techniques with the creation of word and phrase embeddings that reflect the semantic structure of the source texts. These embeddings are then systematically analyzed and adjusted using geometric and semantic methods, including knowledge-based representations, to minimize hidden correlations that give rise to gender bias. The resulting debiased representations form the basis for training the project’s large language model in a controlled and transparent environment.

Following the model fine-tuning phase, Fair-legal applies intelligent prompt-based techniques as a complementary refinement step. This ensures consistent and reliable behavior across different types of texts and supports the model’s ability to generalize beyond clearly visible cases of bias. The outcome is an AI system that can recognize discriminatory patterns even when they are subtle or implicit, offering corrected formulations that align with modern standards of equality and inclusiveness. The effectiveness of the approach is validated through two pilot applications corresponding to the project’s target domains. In legal texts, the system is evaluated on its ability to maintain legal accuracy while eliminating biased formulations. In job notices and public announcements, the focus is on producing inclusive language that supports fair access to employment opportunities without altering the professional or formal character of the text.

Fair-legal is implemented through a collaboration between the University of the Aegean, ITML, and Lioncode, combining academic research expertise with industrial-scale development and deployment capabilities. Ethical and legal considerations are embedded throughout the project, in line with European principles for trustworthy artificial intelligence and full compliance with data protection regulations. Human oversight, transparency, and non-discrimination are treated as core design requirements rather than optional features.

By the end of the project, Fair-legal is expected to deliver a practical AI capability that institutions can integrate into their digital workflows, supporting fairer communication in high-impact contexts. more broadly, the project demonstrates how advanced language technologies can be aligned with ethical principles and societal values, offering a concrete example of how AI innovation can contribute to equality, accountability, and trust.