Fairness fix for AI systems that companies can bolt on without retraining

Researchers developed a method that adjusts AI predictions after they're made, removing bias tied to protected attributes like race or gender without requiring changes to the underlying model. Companies can now apply this to existing AI systems in production — no retraining, no access to the original training code needed.

Until now, fixing bias in deployed AI required either rebuilding the model from scratch or having full access to its architecture — both difficult in real-world systems where companies use third-party models or can't afford to retrain. This post-processing method removes that constraint. It means bias mitigation becomes something you can apply to any AI system in operation, not just ones you built yourself. The practical question is whether this actually gets used when the simpler option is to do nothing — adoption depends on whether liability concerns, regulation, or market pressure force companies to deploy it.