Fairness in AI concerns whether a system's decisions, predictions, and outputs distribute benefits and harms justly across different people and groups. It operates on several levels: individual fairness (treating similar people similarly), group fairness (ensuring comparable outcomes across demographic groups), and procedural fairness (ensuring the process by which decisions are reached is equitable and transparent).

There is no single agreed definition of algorithmic fairness. Mathematically, different fairness criteria — demographic parity, equalized odds, predictive parity, calibration — are often mutually incompatible. This means that designing a "fair" system requires explicit choices about which group or which kind of harm takes priority, choices that are inherently value-laden and contested.

Fairness problems typically emerge at the intersection of data (what was collected, and about whom), models (what patterns were learned), and deployment (what decisions the system is used to make). The issue is compounded when AI is used as a decision-making layer that obscures human responsibility.

Why it matters

AI systems are increasingly embedded in high-stakes decision-making across domains such as healthcare, finance, criminal justice, and employment, and evidence has accumulated showing that these systems can reproduce and amplify structural inequities, leading to ethical, social, and technical concerns.

When AI systems are unfair, they make mistakes in a patterned, structural way that compounds existing social disadvantage. Individuals already marginalised by historical discrimination face compounded harm when automated systems - which carry the appearance of objectivity - replicate or worsen that discrimination at scale and speed no human bureaucracy could match.

Impacts & consequences

When AI fairness norms break down, the consequences span individuals, institutions, and society:

• Individual harm. People are wrongly denied jobs, credit, bail, healthcare, or welfare benefits. The harm is often invisible; the person simply receives a rejection with no explanation and no recourse.

• Amplification of inequality. Biased systems trained on historically unequal data perpetuate and entrench the same patterns. Biases can perpetuate systemic discrimination and inequality, with detrimental effects on individuals and communities in areas like hiring, lending, and criminal justice.

• Erosion of institutional trust. When governments and corporations are shown to use opaque, discriminatory systems, public trust in both AI and the institutions deploying it is damaged - often irreversibly.

Causes & contributing factors

Fairness issues in AI rarely stem from malicious intent; instead, they are usually caused by:

• Historical data bias. Training data that reflects past human prejudices, structural inequalities, or uneven societal practices.

• Representative bias. Data sampling that underrepresents or misrepresents certain demographic groups (e.g., facial recognition trained primarily on lighter skin tones).

• Proxy variables. Removing protected attributes (like race) but keeping data points strongly correlated with them (like zip/post codes or neighbourhood data), leading to indirect discrimination.

• Flawed objective functions. Optimising an algorithm solely for efficiency or profit without constraining it for equitable outcomes.