Periodic Reporting for period 2 - AEQUITAS (ASSESSMENT AND ENGINEERING OF EQUITABLE, UNBIASED, IMPARTIAL AND TRUSTWORTHY AI SYSTEMS)
Periodo di rendicontazione: 2024-05-01 al 2025-10-31
As AI-based decision support systems spread across public and private sectors in a context of growing inequalities and intersectional discrimination, AI fairness and Trustworthy AI have become central EU priorities because, despite fairness being a core EU principle, there was a major gap in concrete technical, socio-legal and organisational methods to ensure, test and repair non-discriminatory, rights-respecting AI systems
AEQUITAS has addressed this gap by developing a comprehensive, controlled experimentation and governance environment for fair and trustworthy AI, delivering a Fair-by-Design methodology and blocks (covering fairness, fundamental rights and compliance across the AI lifecycle), the Information Flow Model (IFM) for socio-technical analysis, an integrated Experimenter to design and document fairness-focused tests, a synthetic data methodology and engine, diagnostic tools with new bias metrics, mitigation methods, and new benchmark datasets for realistic high-risk domains. These components have been validated in real-world use cases in healthcare, HR, education and services for vulnerable groups, and are now accessible both via the AI-on-Demand platform and as an on-premises prototype, enabling organisations to assess and improve AI fairness directly on their own sensitive data in line with privacy, security and regulatory requirements.
Beyond the project, AEQUITAS has been taken up in EU-level guidance for regulatory sandboxes for Member States, where its methodologies are used as building blocks for fairness- and rights-oriented sandbox design, and it is being integrated into the Italian “AI Factory” initiative as a reference framework for experimentation and pre-commercial validation of trustworthy AI solutions.
Finally, AEQUITAS has produced a substantial body of educational material and actionable knowledge tailored to different stakeholder communities: guidance and tools for civil society and vulnerable groups to understand and contest AI decisions; practical handbooks and frameworks for policymakers and regulators; methodologies, metrics and software for data scientists and developers; and governance models, templates and training pathways for the corporate and industrial sectors. Through this combination of methods, software, benchmarks, policy integration and capacity-building, AEQUITAS has established a reference framework for fair and trustworthy AI, contributing to scientific progress, regulatory readiness and practical impact in high-risk AI domains.
AEQUITAS has addressed this gap by developing a comprehensive, controlled experimentation and governance environment for fair and trustworthy AI, delivering a Fair-by-Design methodology and blocks (covering fairness, fundamental rights and compliance across the AI lifecycle), the Information Flow Model (IFM) for socio-technical analysis, an integrated Experimenter to design and document fairness-focused tests, a synthetic data methodology and engine, diagnostic tools with new bias metrics, mitigation methods, and new benchmark datasets for realistic high-risk domains. These components have been validated in real-world use cases in healthcare, HR, education and services for vulnerable groups, and are now accessible both via the AI-on-Demand platform and as an on-premises prototype, enabling organisations to assess and improve AI fairness directly on their own sensitive data in line with privacy, security and regulatory requirements.
Beyond the project, AEQUITAS has been taken up in EU-level guidance for regulatory sandboxes for Member States, where its methodologies are used as building blocks for fairness- and rights-oriented sandbox design, and it is being integrated into the Italian “AI Factory” initiative as a reference framework for experimentation and pre-commercial validation of trustworthy AI solutions.
Finally, AEQUITAS has produced a substantial body of educational material and actionable knowledge tailored to different stakeholder communities: guidance and tools for civil society and vulnerable groups to understand and contest AI decisions; practical handbooks and frameworks for policymakers and regulators; methodologies, metrics and software for data scientists and developers; and governance models, templates and training pathways for the corporate and industrial sectors. Through this combination of methods, software, benchmarks, policy integration and capacity-building, AEQUITAS has established a reference framework for fair and trustworthy AI, contributing to scientific progress, regulatory readiness and practical impact in high-risk AI domains.
AEQUITAS has progressed from concept to a fully implemented and validated framework for fair and trustworthy AI, combining socio-technical analysis, diagnostic tools, mitigation methods, synthetic data generation and a controlled experimentation environment.
Technically, the project has delivered the AEQUITAS Experimenter, available both as an on-premises prototype and as a service integrated with the AI-on-Demand platform. The Experimenter orchestrates all core components: data and model upload, multi-metric fairness assessment, selection and application of mitigation strategies, synthetic data generation and export of audit-ready reports. This is underpinned by a modular architecture and standard interfaces, allowing integration with existing AI pipelines.
On the diagnostic side, AEQUITAS developed the Information Flow Model (IFM) as a socio-technical methodology to map how data, decisions and responsibility flow across human and algorithmic actors. IFM is complemented by an Awareness & Diagnosis Engine, which implements families of fairness and bias metrics capturing disparities across groups, over time and at different stages of the decision pipeline. Together, these allow multi-layer bias detection that links statistical patterns to their organisational and social context.
For bias mitigation, the project implemented an extensible Mitigation Engine (aequitas-lib), covering pre-, in- and post-processing techniques, and introduced novel methods such as FaUCI, FAiRDAS and the FairBridge reasoning layer for constraint-based, ranking-oriented and hybrid symbolic–statistical fairness enforcement. These methods are accessible through the Experimenter, enabling users to compare alternative mitigation strategies and quantify trade-offs between accuracy and fairness.
AEQUITAS also defined a comprehensive Fair-by-Design (FbD) methodology, operationalised via the FairBridge engine and a set of sub-methodologies.
Finally, the project designed and implemented an AEQUITAS synthetic data generation methodology and engine, supporting tabular, textual and image data. This has been used to create realistic, privacy-preserving and fairness-configurable datasets for all use cases, enabling stress-testing and experimentation where access to real data is constrained.
Technically, the project has delivered the AEQUITAS Experimenter, available both as an on-premises prototype and as a service integrated with the AI-on-Demand platform. The Experimenter orchestrates all core components: data and model upload, multi-metric fairness assessment, selection and application of mitigation strategies, synthetic data generation and export of audit-ready reports. This is underpinned by a modular architecture and standard interfaces, allowing integration with existing AI pipelines.
On the diagnostic side, AEQUITAS developed the Information Flow Model (IFM) as a socio-technical methodology to map how data, decisions and responsibility flow across human and algorithmic actors. IFM is complemented by an Awareness & Diagnosis Engine, which implements families of fairness and bias metrics capturing disparities across groups, over time and at different stages of the decision pipeline. Together, these allow multi-layer bias detection that links statistical patterns to their organisational and social context.
For bias mitigation, the project implemented an extensible Mitigation Engine (aequitas-lib), covering pre-, in- and post-processing techniques, and introduced novel methods such as FaUCI, FAiRDAS and the FairBridge reasoning layer for constraint-based, ranking-oriented and hybrid symbolic–statistical fairness enforcement. These methods are accessible through the Experimenter, enabling users to compare alternative mitigation strategies and quantify trade-offs between accuracy and fairness.
AEQUITAS also defined a comprehensive Fair-by-Design (FbD) methodology, operationalised via the FairBridge engine and a set of sub-methodologies.
Finally, the project designed and implemented an AEQUITAS synthetic data generation methodology and engine, supporting tabular, textual and image data. This has been used to create realistic, privacy-preserving and fairness-configurable datasets for all use cases, enabling stress-testing and experimentation where access to real data is constrained.
AEQUITAS has advanced the state of the art by delivering an integrated framework that combines novel fairness algorithms (e.g. FaUCI, FAiRDAS, FairBridge), a socio-technical methodology (IFM), a full Fair-by-Design lifecycle, and a controlled experimentation environment (Experimenter with synthetic data engine) validated on real high-risk use cases. These results provide a coherent toolchain that links fundamental rights and AI-Act requirements to concrete metrics, mitigation methods and governance procedures, already influencing EU sandbox guidelines and the Italian AI Factory. To ensure further uptake and impact, key needs include: continued research and demonstration in additional domains (e.g. GenAI and multi-modal systems), product-level hardening and scaling of the software, stronger integration into standards and certification schemes, access to finance and industrial partners for commercialisation, and sustained support for capacity-building and education so that regulators, industry and civil society can effectively apply the AEQUITAS framework.