AI Fairness 360 (AIF360) is IBM Research’s open-source toolkit for detecting, understanding and mitigating bias in machine learning datasets and models. It bundles a comprehensive library of fairness metrics, explanations of what those metrics mean, and bias-mitigation algorithms that can be applied before, during or after training. The project has close to 3,000 stars on GitHub and remains the reference toolkit in the responsible-AI field.
Fairness problems in ML are subtle: a model can be accurate on average while systematically disadvantaging a protected group. AIF360 gives practitioners a disciplined way to audit that risk. It ships with example datasets used in fairness research, dozens of metrics computed for datasets, models and predictions, and a suite of algorithms โ from reweighing training data and learning fair representations to post-processing model outputs โ that trade a small amount of accuracy for measurably fairer outcomes. Interactive demos on the project site explain each concept visually, which has made AIF360 a standard teaching resource as well as a practical auditing tool. A companion R package serves the same community. The toolkit is framework-agnostic: it works with models trained in scikit-learn, TensorFlow, PyTorch or anything else that produces predictions.
AI Fairness 360 is completely free and open source under the Apache-2.0 license. There is no commercial edition or paid service attached to the toolkit.
AIF360 is for data science teams, model risk and compliance specialists, and researchers who need to audit ML systems for bias and demonstrate โ with established metrics โ that mitigation steps actually worked.