ROBUSTIFAI · ROBUSTIFYING GENERATIVE AI THROUGH HUMAN-CENTRIC INTEGRATION OF NEURAL AND SYMBOLIC METHODS

Generative AI (GenAI), such as foundation models, represents a powerful and transformative class of AI capable of learning patterns from data and generating new content. However, GenAI has notable shortcomings that can lead to misuse or hinder its widespread adoption and positive societal and economic impact. These shortcomings stem from its lack of robustness in three key areas: technical, operational, and user robustness. Addressing these challenges in foundation models, especially in the context of human cyber-physical systems (HCPS)—the most demanding GenAI applications in terms of robustness—will pave the way for solutions applicable across various domains, unlocking GenAI's full potential.Building on the EU’s competitiveness in constructing and assuring dependable complex systems, RobustifAI, a 3-year project with a budget around €9M, brings together 18 leading partners spread over 10 EU countries but also Switzerland and India, to tackle the above three-dimensional robustness challenge in GenAI systems. We aim to make a step change to the existing GenAI system development paradigm by developing and promoting a rigorous design and deployment methodology for building robust GenAI systems. The methodology is based on the following three orthogonal innovative axes: (1) techniques to understand, express, and embed human-centric needs within the neural model, (2) principled methods for integrating neural models and symbolic techniques, and (3) enabling the adaptivity of GenAI systems to environmental changes and user variations. RobustifAI will actively contribute to on-going EU initiatives on AI, such as the AI-BOOST project on AI challenges and other EU projects on AI efficiency, autonomous vehicles, or service robots. Its successful execution will secure for the EU a distinct and leading position in more sustainable and socially beneficial AI advancements, and strengthen EU’s vision that technical advances and societal benefits can be achieved simultaneously.