Scaling fair and responsible AI: Collaboration with Meta

Our department has developed a long-term collaboration with Meta, delivering impactful research and scalable machine learning solutions.

The work has supported large-scale systems, including models for predicting social media content popularity and active learning methods that reduce labelling costs while improving model performance. The current focus of the collaboration is multicalibration, a fairness framework ensuring predictions remain accurate and equitable across diverse and overlapping user groups. This capability is increasingly critical for organisations deploying AI in high-stakes or regulated environments.

The work has produced significant outcomes within Meta’s production ecosystem, including MCGrad, a novel scalable multicalibration algorithm supporting hundreds of models and generating over one million multicalibrated real-time predictions per second. It has also delivered key research outcomes, such as the algorithm’s development, theoretical guarantees for gradient boosting integration, and open-source tools enabling organisations to adopt responsible, fairness-aware AI practices efficiently.