AI4Gov makes a step forward towards identifying bias that seriously affects decisions based on AI algorithms. Bias is a cognitive phenomenon that significantly shapes individuals' perceptions, judgments, and decision-making across diverse situations. The emergence of AI bias occurs when algorithms generate systematically prejudiced outputs due to biased assumptions during development or within the training data. For this purpose, AI4Gov has developed a Bias Detector Toolkit which is a holistic application focused on explaining AI bias and equipping developers with an easy-to-navigate and visually organized catalogue.
In addition, the aspects of explainability, interpretability, and transparency in the modern policymaking domain are essential for the provision of citizen-centered and democratic decisions. The AI4Gov project seeks to leverage the potential derived from the utilization of SotA and cutting-edge eXplainable AI (XAI) techniques and approaches, such as LIME and SHAP that can be relative useful on the provision of preference importance ranking in the explanation stage. However, the project goes also beyond these techniques through the introduction of Situation Aware eXplainability (SAX) techniques that are evolutionary XAI techniques applied to business processes. They aim at tackling the shortcomings of contemporary XAI techniques when applied to business processes, such as their inability to express the business process model constraints and to include the richness of contextual situations that affect process outcomes. While their explanations are usually not given in a human-interpretable form that can ease the understanding by humans. A situation-aware eXplanation is a causal sound explanation that takes into account the process context in which the explanandum occurred, including relevant background knowledge, constraints, and goals. Under this scope Large Language Models (LLMs) are also utilized that are a subset of foundation models that can perform a variety of natural language processing (NLP) tasks and are capable generating narratives for improved process outcome explanations. In addition, a SAX4BPM library was also developed to include a set of services to support the different aspects of SAX explanations, taking into account contextual information classified into three different types: completeness, soundness, and synthesis.
Complementary to these novel approaches, a decentralised blockchain infrastructure has been designed and initially implemented within this first reporting period to enhance the transparency and traceability of data and policies storage and business logic execution. In deeper detail, off-chain policies are expected to govern core operational characteristics of the network and enforce data policies that are not expected or allowed to change in the future. On-chain policies, on the other hand, involve the whole consortium, both pilots of AI4Gov and future adopters, and allow the collaboration in changes needed in the implemented business logic and/or the data flow scenarios. Furthermore, considerations regarding GDPR’s accountability guidelines, the right of individuals to control their data, the right to be forgotten and its compatibility with the immutable nature of decentralised ledgers were taken into account contributing to the introduction of a set of novel and validated decentralised business models. These models further enable public and private organizations to monetize their assets and boost the trustworthiness of the AI-based policy development process for citizens, businesses, and public administrations.
