Ethical AI through Neurosymbolic Methods

Project Description

In this project, we develop a principled neurosymbolic framework for embedding ethical and legal constraints directly into neural models. Motivated by the growing influence of AI systems in high-stakes domains such as hiring, healthcare, and public administration, the work addresses a central problem of contemporary AI: ethical guidelines are typically implemented implicitly by a small group of technical experts and private actors, rendering normative control opaque and inaccessible to broader stakeholders. We argue that trustworthy AI requires transparent, controllable, and formally grounded mechanisms for aligning model behavior with societal values.
The core contribution of our work is the design of different Neurosymbolic architectures that integrate neural learning with symbolic reasoning over logical constraints. The proposed approaches introduce an explicit interface between neural latent spaces and declarative ethical specifications formalized as logic-based rules or structural causal models. Thereby, they enable interpretable, modular, and adaptable alignment mechanisms that remain agnostic to specific definitions of abstract concepts, such as fairness or harm.
The framework is validated through several application domains: e.g. in the scenario of automated job interview analysis, where symbolic constraints will be used to mitigate demographic bias in multimodal hiring systems. Another use case is safety in large language models, where institutional guidelines as well as legal and ethical norms will be translated into formal constraints and enforced within embedding spaces.
Overall, this projects contributes a unifying technical and conceptual framework for ethical AI alignment, advancing neurosymbolic methods as a viable foundation for transparent, accountable, and democratically governable AI systems.

Contact Persons

Leo Kestel

Send an email

More

Prof. Dr. Christoph Kern

Social Data Science and Statistical Learning