Learn AI /Responsible AI, Human Rights and Safety /Fairness

Fairness

Intermediate
Responsible AI, Human Rights and Safety
Human rights practitioners
Policy officers
Programme & data teams

Also called: equity in AI.

Detail

In five seconds

What it is

Designing and checking AI systems so they do not produce avoidable, unjust differences in how people or groups are treated or affected.

In plain words

Fairness in AI means checking whether a system works reasonably well for different people and groups, and changing it if it creates avoidable or unjust disadvantages.

An everyday example

A shop installs an automatic door that opens reliably for most adults walking towards it. Staff then notice that it reacts less consistently to shorter people, wheelchair users and people approaching slowly. The door is not fair merely because it opens correctly most of the time. Its sensor placement and timing need to be tested for the different people expected to use it. The settings can then be adjusted, difficult cases checked again, and failures monitored after installation. The goal is dependable access without avoidable barriers, not identical movement by every customer.

Where the story stops working

An exam is designed openly by people who can inspect each question directly. Many AI systems are more complex, may update over time, and can affect people at much larger scale. Also, fairness in AI is not only about test performance. It can involve allocation of resources, ranking, moderation, surveillance risks, and legal or human-rights questions that an exam analogy does not capture.

Remember thisFairness requires examining who receives the benefits, burdens and errors, not relying on one average performance score.

This term is contested

The term is genuinely contested because different fields use different fairness definitions and metrics, and some of those definitions cannot all be satisfied at once in the same system.

Understand it

Why it matters

Fairness matters because AI systems can quietly distribute opportunity, burden, visibility, and risk. A tool can save time overall and still disadvantage certain groups, languages, regions, or people with disabilities. If decision-makers look only at average performance, they may miss unequal error rates, exclusion, or barriers to appeal. That can create legal risk, damage trust, and conflict with human-rights responsibilities. For non-specialists, the key point is practical: fairness affects whether a system should be used, how it should be limited, what evidence is needed before deployment, and when a human must stay in the loop. It also shapes procurement, monitoring, complaint handling, and whether affected people have a meaningful way to challenge harmful outcomes.

How fairness is shaped across an AI system

After viewing this, the reader should be able to explain that fairness depends on choices made before, during, and after model output, not just on one score.

01
1. Frame the task
Decide what the system should do, who it affects, and what harms matter. A badly framed task can be unfair even before any model is built.
Draw the first rectangle labelled 'Frame the task'. Under it, add smaller text: 'Who benefits? Who may be harmed? Should this be automated at all?' Use an icon of a clipboard or question mark.
02
2. Choose data and labels
Select examples, categories, and quality checks. Missing groups, poor labels, or proxy variables can create unfair patterns.
Draw the second rectangle labelled 'Data and labels'. Add subtext: 'Who is represented? Are labels reliable? Are some features acting as proxies?' Use a stacked-dataset icon.
03
3. Set model rules
Training choices, thresholds, and optimisation targets influence who gets errors and who gets access. High overall accuracy can hide unequal performance.
Draw the third rectangle labelled 'Model and thresholds'. Add subtext: 'What is being optimised? For whom do errors rise?' Use slider controls or gauge icon.
04
4. Deploy with safeguards
Interface design, human review, and appeal routes affect real outcomes. The same model can be more or less fair depending on how it is used.
Draw the fourth rectangle labelled 'Deployment context'. Add subtext: 'Human oversight? Explanations? Appeal path?' Use screen-plus-person icon.
05
5. Monitor and remedy
Check outcomes over time, across groups and settings, then correct problems. Fairness is an ongoing governance activity, not a one-time test.
Draw the fifth rectangle labelled 'Monitoring and remedy'. Add subtext: 'Audit results, hear complaints, update or stop use.' Use loop arrow or checklist icon.

A left-to-right flow with five boxes connected by arrows. Each box names a stage where fairness can change: problem framing, data and labels, model and thresholds, deployment context, and monitoring and remedy. Small notes under each box show typical questions to ask. Read from left to right to see how early choices affect later outcomes.

Hover or tap each box to reveal one example question and one example risk for that stage; keyboard focus should reveal the same content in a side panel.

Describe this without the diagram

The diagram is a five-stage horizontal flow showing where fairness can change in an AI system. Stage one is framing the task: ask who benefits, who may be harmed, and whether automation is appropriate. Stage two is data and labels: check representation, label quality, and proxy features. Stage three is model and thresholds: examine what is optimised and whether errors differ by group. Stage four is deployment context: add human oversight, explanations, and appeal routes. Stage five is monitoring and remedy: review outcomes over time, listen to complaints, and update, restrict, or stop the system if needed. The main message is that fairness depends on the whole pipeline, not only the model.

With reduced motion: Show the same five boxes and arrows as a static diagram, with all example questions visible beneath each stage and no animations.

Fairness in AI means asking whether a system treats people and groups in a justifiable way, and whether its benefits and harms are shared reasonably. The focus is not only the final answer from the system. It also includes who was represented in the data, who helped set the rules, who can challenge a result, and who carries the risk if the system is wrong.

What goes in includes data, labels, design choices, success measures, and the social context. A label is the category attached to an example, such as "approved" or "rejected". A success measure is the number a team uses to judge performance, such as overall accuracy. What happens is that people choose what to predict, what data to use, how to test results, and what differences between groups are acceptable or unacceptable. Those choices shape the output.

What comes out may be recommendations, rankings, flags, or generated text. A system can look accurate overall but still work worse for some groups, languages, regions, or disability needs. It can also seem fair by one measure and unfair by another. Different fairness measures can conflict. For example, matching error rates may not align with matching positive prediction rates or calibration.

Fairness is often confused with algorithmic bias and non-discrimination, but they are not identical. Algorithmic bias refers to systematic unfair patterns in data, rules, or outputs. Non-discrimination prevents exclusion, burdens, or denied access linked to protected traits or close proxies. Fairness is the broader question of what differences are acceptable, who decides that, and how to prevent or remedy harm.

Because these trade-offs are context-dependent, fairness is not something a model can prove by itself. It needs human judgement, testing with affected groups, documentation, routes for appeal, and ongoing review after deployment.

How “Adjusting an automatic door” maps onto the real thing

In the story	In the AI system
practice examination questions	the model task and its prompts or decision rules
students from different backgrounds and access needs	different user groups affected by the system
reviewing scores by subgroup	fairness testing across groups, not only overall performance
rewriting unclear or inaccessible questions	changing data, design choices, interfaces, or policies to reduce unfair outcomes

A worked example

An information management officer at a regional hub is testing machine translation before a multilingual situation report workflow is cleared.

Starting from: A set of previously published passages in several languages, with translations reviewed by qualified language staff.

The officer compares overall translation quality, then examines omissions, mistranslations and changes in meaning separately for each language.
The review finds that the system handles common administrative phrases well but more often drops qualifiers and negation in one less-represented language.
The workflow is changed so that passages in that language are routed for human review. A reviewed terminology list is added, and the system is tested again on fresh published material.
The officer records the remaining limitations in the monitoring report and sets a review point after deployment rather than treating the initial test as final.

Result: The tool is not assumed to be fair because its average score is acceptable. Different error patterns are identified, the workflow is adjusted, and language-specific performance remains subject to monitoring.

In a United Nations workplace

A statistician in an evaluation office is completing the technical scoring sheet for three proposals for an AI-assisted survey-coding service before the procurement panel reconvenes at close of business. Each bidder reports strong overall accuracy, but only one provides results broken down by language, response format and accessibility needs.

An overall score can conceal uneven error rates. The statistician asks bidders to show representation in their test data, false positive and false negative rates for relevant groups, language coverage and arrangements for human review. Fairness becomes part of the technical assessment rather than a promise left for deployment.

Before you try thisBid material and evaluation records are procurement-sensitive. They must not be entered into an external AI tool unless that tool has been approved for this use and the handling is consistent with the applicable access controls. Testing should use authorised or synthetic data, not identifiable respondent records.

Where you may hear it

While a procurement panel questions a vendor about subgroup error rates
During clearance of an evaluation report on an automated service
At an ICT governance board reviewing access and accessibility
In the informals on language covering non-discrimination and due diligence

Avoid confusion

Often confused with

Common misconceptions

Common misconception: If a model has high accuracy, it is fair enough.
Correction: High overall accuracy can hide very different error patterns across groups. A system may perform well on average while failing more often for people in a certain language, region, age band, or accessibility category. Fairness asks who bears the mistakes, not just how many mistakes exist in total.
Common misconception: Fairness means every group must get identical outcomes.
Correction: Not necessarily. Equal outcomes can be one goal, but fairness can also involve equal opportunity, equal treatment, justified accommodation, protection from harm, or access to appeal. The right approach depends on the context, the rights at stake, and whether differences are justified or avoidable.
Common misconception: Fairness is purely a technical metric that engineers can solve alone.
Correction: Metrics help, but they do not settle the hard questions. Teams still need policy, legal, operational, and human-rights judgement about acceptable trade-offs, affected groups, oversight, and remedy. People using the system shape fairness as much as the model does.

Quick check

Which statement best captures fairness in AI?

Go deeper

The simple version

Fairness in AI means checking whether a system works reasonably well for different people and groups, and changing it if it creates avoidable or unjust disadvantages.

The technical version

Technically, fairness refers to normative and measurable ways of evaluating whether an AI system’s design, outputs, and real-world effects distribute errors, benefits, burdens, and access in justifiable ways across individuals or groups. Assessment may look at representation in training and evaluation data, label quality, proxy variables, model objectives, decision thresholds, calibration, false positive and false negative rates, ranking behaviour, accessibility, language coverage, and post-deployment outcomes. Different fairness criteria can conflict. For example, matching error rates across groups may not align with matching positive prediction rates or calibration. Because of that, fairness cannot usually be reduced to one universal metric. It also extends beyond model statistics to governance questions: whether the task should be automated, whether affected people can contest outcomes, whether oversight is meaningful, and whether impacts are consistent with human rights and non-discrimination obligations.

Reviewed external resources

Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
The UNESCO Recommendation gives a policy and governance framing for AI ethics and suits readers looking for an official intergovernmental perspective.
Intermediatelink checked 2026-07-31
Primary sourceB-Tech Project (opens in a new tab)UN Human Rights Office (OHCHR)
The OHCHR B-Tech Project offers human-rights-oriented material on technology governance and suits readers focused on business responsibility and rights impacts.
Intermediatelink checked 2026-07-31
EncyclopaediaFairness (machine learning) (opens in a new tab)Wikipedia
Wikipedia gives a broad overview of fairness in machine learning and suits readers who want a starting map of the topic and its different definitions.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Fairness (opens in a new tab)YouTube search
The YouTube search results page can help readers find different video explainers on fairness, and it suits people who prefer to compare several teaching styles before choosing one.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Fairness (opens in a new tab)Reddit search
The Reddit search results page shows public discussions and questions about fairness, and it suits readers who want to see common points of confusion and debate rather than a settled reference.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

Fairness requires examining who receives the benefits, burdens and errors, not relying on one average performance score.