An AI model learns from past data, and past data reflects past decisions, including unfair ones. A model trained on historical hiring might learn to prefer the profiles that were hired before. A credit model might penalise postcodes that stand in for income or community. None of this needs bad intent; it happens by default unless someone checks.
What an audit looks at
- Data: who is represented in the training data, who is missing, and whether labels carry old judgements.
- Outcomes: whether results differ across groups such as gender, age, region or language, beyond what the task justifies.
- Proxies: features that quietly stand in for protected characteristics.
- Process: who can override the model, how complaints are handled and how often the system is re-checked.
Which systems to audit first
Prioritise any AI that affects people's access to jobs, credit, insurance, education, healthcare or public services. Customer-facing chatbots and voice agents deserve a check too: they should work as well for someone speaking Hindi or Gujarati as for someone speaking English.
After the audit
A good audit ends with named risks and fixes you can schedule, not just a score. Typical fixes include rebalancing data, removing proxy features, adding human review for borderline cases and setting up regular monitoring, because models drift as the world changes.