Algorithmic Bias

Algorithmic bias occurs when an AI system produces systematically different outcomes for different groups of people based on protected characteristics like race, gender, age, or socioeconomic status. It can originate in training data, in feature engineering that introduces proxy variables, or in threshold decisions that were never tested across the full population the system serves. 

In LLM-based systems, bias surfaces as outputs that consistently misrepresent or disadvantage certain groups, or as retrieval pipelines that return skewed information. In agentic systems, bias can emerge across multi-step decision chains where no single step appears problematic but the cumulative effect systematically disadvantages certain populations. 

Bias is not a theoretical risk: it is a measurable property of a system's outputs.  

At Eticas, we test for bias in production, using the system's actual outputs with the populations it actually serves, not on benchmark datasets that may not reflect deployment conditions. Our open-source library provides the statistical framework for demographic benchmarking, fairness metrics, and drift detection across privileged and underprivileged groups.