As more and more tasks and decisions are delegated to AI-enabled computers, mobile devices, and autonomous systems, it is crucial to understand the impacts this may have on people and that AI treats people ethically. Among other topics, we are working on:
1) Value-based and explainable AI, where we are developing AI models that are able to reason about human values, so that AI models act according to them. We also work on making AI models more transparent and explainable, so that users can better understand what they do and why. We have already proven that, in some specific recommendation domains, making AI value-aligned and explainable leads to more accpetable and satisfying recommendations. This also allows for a better way to scrutinise AI models in general.
2) AI Discrimination, where users may be treated unfairly or just differently based on their personal characteristics (e.g. gender, ethnicity, religion, etc.). Interestingly, AI very often reproduces existing instances of discrimination in the offline world by either inheriting the biases of prior decision makers, or simply reflecting widespread prejudices in society. Therefore, by developing methods to study AI discrimination, this also enables us to understand instances of human discrimination. For instance, we have applied our methods to discover biases in natural language processing models to discover dangerous prejudices in online communities using their own language (usually containing slang).
Our research in this domain often involves cross-disciplinary collaborations, including colleagues from the social sciences, digital humanities, law, ethics and policy/governance.
See more publications on this topic here