An AI system can pass security and privacy checks yet still produce worse outcomes for particular groups. Fairness is a way to assess whether treatment and outcomes are equitable, but no single definition settles the question. A group-level check can miss unequal treatment of similar individuals, and different formal definitions can conflict. section 6.0 section 6.2.2
Choose fairness lenses according to the harm a system could cause and who bears the cost of mistakes. In an insurance claims system, group fairness might mean examining denial rates by age band, ethnicity, or postcode; individual fairness asks whether similar claims receive different outcomes. Appeal success rates can reveal another difference in treatment. section 6.2.2
Table 6.2 Fairness lenses in GenAI: definitions and examples of unfair treatments
|
Fairness Type
|
Core Question
|
Example
|
|
Procedural
|
Is the
process consistent and transparent?
|
An HR tool
uses hidden criteria, ranks similar employees inconsistently, and offers no
appeal[59];
a moderation system permanently bans creators without notice or a clear path
to review.
|
|
Distributive
|
Is the
allocation of scarce resources aligned with needs?
|
A utility
restores power to dense urban areas first regardless of medically dependent
rural customers.
|
|
Equality
|
Is everyone held to the same standard?
|
A content moderation system applies looser controls to
users on certain browsers or device configurations, so two people posting
identical content are policed differently based on the device they happen to
use.
|
|
Proportionality
|
Do rewards
reflect contributions or effort?
|
A loyalty
program over-rewards large occasional baskets while ignoring frequent small
purchases from budget-constrained families. A ranking algorithm amplifies
certain creators far beyond what their engagement merits.
|
|
Group
Fairness
|
Do
demographic groups receive similar outcomes? Note: Groups can overlap;
optimizing fairness for one group may worsen fairness for another.
|
Black
borrowers with the same credit score as white borrowers are approved less
often for loans. A face-recognition system misclassifies darker-skinned faces
more frequently than lighter-skinned ones.
|
|
Individual
Fairness
|
Do similar
individuals get similar results?
|
Two
borrowers with near-identical credit and income receive different decisions
because the model overweights irrelevant features (e.g., education,
postcode).
|
|
Counterfactual
|
Would the
outcome change if a sensitive attribute does (e.g., gender, race, age)?
|
Job
applicants with similar profiles and different gender do not have the same
chance of getting hired. A counterfactual test flips only the gender field
for one applicant. If the decision switches from “hire” to “reject,” the
system violates counterfactual fairness.
|
Compare fairness lenses through questions about process, allocation, and equal standards. section 6.2.2
Turn the chosen lenses into engineering controls:
- Define specific fairness questions and review them with business, legal, and, where possible, affected-community representatives. Document the goals in a Responsible AI Charter. section 6.2.4 section 6.2.2
- Review model cards and test candidates against relevant evaluation datasets. BOLD tests bias in open-ended generation, WinoBias targets gender stereotypes, and BBQ tests reliance on stereotypes in ambiguous contexts. Poor results may require stronger downstream controls. section 6.2.4
- Use specific prompt instructions rather than a generic instruction to be fair. Treat prompts as a first filter, not a guarantee. section 6.2.4
- Inspect training and reference data for skewed language or historically disadvantaged outcomes. Check outcome rates by relevant groups, decide whether to rebalance, document the decision, and rerun tests after changes. section 6.2.4
Conditional approval for a clinical feature
Derm-Cam recommends whether a patient should see a dermatologist. Approval was conditional on a fairness test and other safeguards, including clinician fallback when the system could not use an approved patient-facing message. Findings had owners and deadlines so completion could be verified. section 3.3.3
Use fairness metrics to inform launch decisions and monitor them afterward. Repeat assessments when models, data, or deployment contexts change. Reliability is necessary but not sufficient: a stable system can still systematically disadvantage a group. Test input variations too, since dialect or language differences can reduce answer quality for reasons unrelated to a user’s needs. section 6.8 section 6.1.2 section 6.1.3













