What is AI bias?
AI bias refers to systematic distortions in AI systems that favour or disadvantage certain groups because training data or model design reflect human prejudices.
DEFINITION
AI bias does not arise from malicious intent but from the training data and the context in which a model is built. A model trained on historical applicant data showing that men were more often hired for certain positions learns that tendency further. The result: an AI recruiting tool favours men — not because that was intended, but because the bias was embedded in the training data. AI bias appears in different forms: representation bias arises when certain groups are underrepresented in training data. Measurement bias arises when the feature used is skewed. Confirmation bias arises from the model assumptions themselves. For decision-makers, AI bias is a legal risk, an ethical problem, and a quality problem at once. Countermeasures: review training data, audit results regularly, and involve diverse teams in AI development.
CONNECTIONS
Leadership
Leaders bear responsibility for AI-supported decisions in their area. Those who do not know about AI bias risk making personnel decisions or performance evaluations based on distorted algorithms without noticing.
Agility
In agile teams that develop or deploy AI systems, the Definition of Done must include explicit bias checks. Without this step, AI bias is systematically declared “done”.
Project Management
AI bias is a measurable project risk that should be recorded in the risk register. Especially in AI projects affecting people (recruiting, evaluation, steering), bias mitigation must be planned as an explicit measure.
KEY POINTS
- AI bias arises from distorted training data or model design.
- It reproduces and amplifies human prejudices systematically.
- Three main types: representation bias, measurement bias, confirmation bias.
- AI bias is a legal, ethical, and quality risk.
- Audits of model results are the most important countermeasure.
EXAMPLE
A company uses an AI scoring system for pre-selecting job applications. After six months, HR finds that female applicants are systematically rated lower. Analysis shows: the model was trained on historical hiring data in which men appeared more frequently in leadership positions. The model adopted this skew. The company pauses the tool and conducts a manual audit.
MISCONCEPTIONS
Is AI more objective than human decisions?
Not automatically. AI can adopt human prejudices from training data and apply them systematically at scale, which can amplify bias rather than reduce it.
Can I completely avoid AI bias?
Complete avoidance is hardly possible because data always carries historical context. The goal is to recognise, measure, and reduce bias as far as possible.