AI bias is the risk that's invisible in the code and only shows in the outcomes — which is why it has to be tested for, not assumed away. This guide explains how to test an AI system for bias, the fairness metrics, and how to mitigate it. Part of our guide to AI testing.
AI bias testing is the process of measuring whether an AI system produces equitable outcomes across the groups it affects, and identifying where it discriminates. Because bias lives in outcomes rather than in the code, the only way to find it is to measure results across groups — you can't read it off the model.
Bias usually enters through the data: unrepresentative training data, historical patterns the model learns and amplifies, and proxy variables that stand in for protected attributes (a postcode standing in for ethnicity, say). Feedback loops can then reinforce it over time. None of this requires anyone to intend to discriminate.
Fairness is measured with specific metrics, and they can conflict — so you choose the ones that fit the use:
Open-source toolkits like Fairlearn and AIF360 help measure and compare fairness metrics across groups, and integrate into an evaluation pipeline. Tools measure; a human still decides what's acceptable in context.
Mitigation can happen at three stages: the data (rebalancing, removing proxies), the model (fairness constraints during training), and post-processing (adjusting outputs) — backed by human oversight for high-stakes decisions.
Fairness is a legal as well as an ethical requirement: the EU AI Act requires non-discrimination for high-risk systems, and India's DPDP and sector rules expect fair treatment. Documented bias testing is how you evidence it. See the EU AI Act.
AramGRC runs independent bias and fairness testing across the groups your AI affects, with documented results you can show regulators and buyers.
Measuring whether an AI system produces equitable outcomes across the groups it affects, and identifying where it discriminates — done by measuring results across groups, since bias lives in outcomes not code.
Identify the protected and proxy attributes, measure outcomes across those groups with fairness metrics, compare against a threshold, and document findings and mitigation.
Measures like demographic parity, equal opportunity, equalised odds and disparate impact that quantify whether outcomes are equitable across groups — they can conflict, so you pick the ones that fit the use.
Mitigate at the data (rebalancing, removing proxies), the model (fairness constraints), or post-processing (adjusting outputs), backed by human oversight for high-stakes decisions.
The types of AI testing — evaluation, red-teaming, security, bias and drift — and how to use them.
How to evaluate an AI model's performance, quality and metrics.
How to test an AI system for bias and unfair outcomes.
You're here
Adversarial testing, prompt injection and jailbreak defence for AI.
Monitoring AI models in production for drift, performance and bias.
The tools used to red-team and adversarially test AI systems.
A step-by-step process for testing AI models before and after release.
I red-teamed the same LLM with Garak, PyRIT and promptfoo — what broke, and which I'd pick.