We know that algorithmic bias is a serious problem. But what can we concretely do to address it? The answer is that there is no magic solution, but there are approaches that, when combined, can significantly reduce the risk. They fall into three categories: interventions on data, interventions on the algorithm and interventions on the process.
Interventions on data aim to correct distortions in the training dataset. This may mean balancing the representation of different groups, removing variables that act as proxies for protected characteristics (such as race or gender), or applying "data augmentation" techniques that generate synthetic data for under-represented groups. These interventions are necessary but not sufficient: data is always an imperfect reflection of reality.
Interventions on the algorithm include so-called "fairness constraints" — mathematical constraints that require the model to produce fair results according to a specific definition. There are dozens of fairness metrics: demographic parity (equal proportions for each group), equalised odds (same error rate for each group), calibration (same prediction accuracy for each group). As we have seen, these definitions can be incompatible. Choosing which to apply is an ethical, not technical, decision.
But the most important interventions are on the process. The algorithmic audit — a systematic and independent analysis of a system's outputs — is the fundamental tool for identifying and documenting bias. Organisations like the Algorithmic Justice League, founded by Joy Buolamwini at MIT, have demonstrated through audits that Amazon, Microsoft and IBM's facial recognition systems had error rates of up to 34.7% for dark-skinned women, compared to less than 1% for light-skinned men.
Participatory design is another fundamental approach: involving affected communities in the design of systems that concern them. Not as "test users" at the end of the process, but as co-designers from the start. Those who experience the consequences of algorithmic discrimination possess knowledge that no engineer, however brilliant, has.
Finally, transparency and accountability. Companies that develop and deploy AI systems should publish "model cards" and "datasheets" documenting how the model was built, on what data, for what purpose, with what known limitations. And they should be legally liable when their systems cause harm. Without consequences, there is no incentive for change.