AI bias often creeps in after a model is built, as its training data drifts away from the real population. Yao Li, Global Chief Product Officer at Experian, showed how to catch it in a live demo at aiblLIVE London 2026.
Experian compares a model’s training data with population reference data to find groups that are under or oversampled. It then traces each dataset to the models and people that depend on it, so the most critical fixes come first.
All figures as stated on stage by Yao Li.
The population a model learned from stops matching the population it serves. In Experian’s demo, a 2023 baseline was compared with later data by income and age. The biggest change was in the lowest income band, a group Li says is traditionally overlooked. A model trained on the old picture serves them worst.
The population that are most affected is actually the ones that need it the most.
Yao Li, Global Chief Product Officer, Experian
Yes. Li says teams must check for two problems at once.
Undersampling means telling a model too little about a group, as happened with women and people of colour in early facial recognition. Oversampling means giving it far more data on one group than it needs. Both skew what the model learns.
Fix first where critical users depend on the data. Experian maps each drifting dataset to the models and people it feeds.
In the hospital demo, one dataset fed admissions, ambulance arrival and triage models. The admissions model serves clinicians, billing and compliance. Critical users like these justify extra spend, while a prediction that is still good enough can wait.
Yes, by comparing at population level. Li says Experian’s monitoring uses only aggregate data, such as the share of a group in a postcode, and never individual records. That reference data works as a mirror held up to a training set, showing whether a group is under or overrepresented and how that changes.
Yao Li is Global Chief Product Officer at Experian, and an engineer and AI researcher by training. She spoke on the Customer track at aiblLIVE London on 20 October 2026.
If your models were trained a year or more ago, this is the check to run. The full PDF walks through the live demo step by step.
For data and analytics leaders, heads of AI, risk and compliance teams, and product owners of AI models. Five pages, about a seven-minute read.
