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

Inclusion

Starter
Responsible AI, Human Rights and Safety
Human rights practitioners
Programme & data teams

Also called: inclusive design, representation.

Detail

In five seconds

What it is

Inclusion means designing and using AI so different people can access it, be represented in it, and avoid being unfairly left out or harmed.

In plain words

Inclusion means making sure an AI system works for different kinds of people and does not leave some groups out by accident.

An everyday example

You are ordering food for a group. Choosing the most popular dish may still leave out someone with an allergy, someone who does not eat certain ingredients, or someone who cannot read the menu. An inclusive approach asks about needs before ordering, provides clear information, offers workable choices and checks that everyone can actually take part. AI inclusion follows the same logic: do not assume that what works for the majority works for everyone.

Where the story stops working

A translated report is a small document, but AI inclusion applies across the whole system life cycle, including data collection, model training, deployment, monitoring, and complaints processes. Also, making one report easier to use does not by itself show whether the underlying AI performs well across groups or avoids discriminatory effects.

Remember thisInclusion asks who can access the system, who is represented in its design and testing, whose results are worse, and what must change before it is used.

This term is contested

The general idea is widely accepted, but organisations may define the boundaries differently. Some use 'inclusion' mainly for participation and access, while others also fold in representation, fairness, accessibility or broader equity concerns.

Understand it

Why it matters

If inclusion is ignored, an AI system may work smoothly for people who are already easiest to serve and fail others quietly. That can mean missed services, distorted information, extra burdens on staff, and higher risk of unfair treatment. For a non-specialist, this matters because choices about data, language, interface, and review shape who benefits and who is left out. A tool can look successful on average while still failing a specific group. In practical terms, inclusion helps organisations ask better questions before deployment: Who is missing from the data? Who cannot use this interface? Who needs human support instead of automation? Those questions improve quality, reduce harm, and support more accountable decisions.

How inclusion is built into an AI system

Show that inclusion comes from deliberate design and testing choices, not from adding one feature at the end.

01
1. Define who is affected
List the groups who may use the system or be affected by it, including people with less access, different languages, or disabilities.
Draw a box with diverse person icons around it. Add a warning badge labelled 'some users missing?' and a check badge labelled 'include affected groups'.
02
2. Check the data
Review whether the examples, labels, and sources represent those groups well enough for the task.
Draw a dataset box feeding into the process. Show uneven stacks for under-represented groups, then a corrected balanced set beside it.
03
3. Design for real conditions
Build interfaces and workflows that work across devices, literacy levels, languages, and connectivity conditions.
Show phone, laptop, screen-reader, and low-signal icons under a design box. Add small labels: 'language', 'format', 'bandwidth', 'readability'.
04
4. Test with different users
Measure whether performance and usability differ across groups and contexts, then fix problems before wider use.
Draw a test panel with several user groups and a results table. Highlight one lower score in red, then an arrow to a repaired version.
05
5. Monitor and allow feedback
After deployment, track who benefits, who struggles, and how people can report problems and get human review.
Show an output box leading to a feedback loop with alert, comment, and human-review icons.

A left-to-right five-step diagram. Each step shows one point where people can be included or excluded: scoping, data, design, testing, and use. Read from problem definition to deployment and review. Red warning icons mark common exclusion risks; green check icons mark inclusive actions.

Users can click each step to reveal one example of exclusion risk and one practical inclusion action.

Describe this without the diagram

This diagram explains inclusion as a five-step process. First, identify all groups who may use or be affected by the AI system, not just the easiest users to reach. Second, check whether the data represents those groups well enough. Third, design the system for real conditions such as different languages, devices, disabilities, literacy levels, and internet access. Fourth, test performance and usability across groups and repair gaps. Fifth, after deployment, monitor outcomes and provide ways for people to report problems and get human review. The message is that inclusion must be considered from the start and revisited over time.

With reduced motion: A static five-column diagram with numbered boxes, warning markers for exclusion risks, and check markers for inclusive actions. No animation is used; the feedback loop is shown as a simple curved arrow from deployment back to earlier steps.

In AI, inclusion means asking who can use a system, who is represented in the data, and who may be excluded by design choices. It covers language, disability, internet access, literacy, culture, gender, age, and other factors that affect whether a tool works fairly for real people.

What goes in are design choices, training data, testing plans, and decisions about where and how the system will be used. If those inputs mostly reflect one group, one language, or one kind of device, the system may work well for some people and poorly for others.

What happens is not magic. Teams choose examples, labels, interfaces, rules, and review processes. They test the system with different user groups and contexts, then fix barriers they find. A barrier can be technical, such as poor speech recognition for some accents, or practical, such as a tool that only works with fast internet.

What comes out can be broader access and fewer people left behind. But inclusion is not the same as fairness, accessibility, or non-discrimination, even though it overlaps with all three. Inclusion is the wider practice of making sure affected people are considered, represented, and able to participate, while important human decisions still remain under human responsibility.

How “Planning a meal everyone can eat” maps onto the real thing

In the story	In the AI system
The first draft that only suits some readers	An AI system built around a narrow set of users or contexts
Adding alt text and clearer wording	Design changes that improve inclusion for different needs
Checking place names and terms	Testing whether outputs represent people and places accurately
Preparing a low-bandwidth version	Accounting for access limits such as device, connectivity, or cost

A worked example

A conference services focal point in New York must clear an AI assistant before invitations to a side event are circulated. The assistant answers registration questions and helps users complete the online form.

Starting from: A test pack covering keyboard-only navigation, screen readers, low-bandwidth access, several languages, non-standard names and requests phrased in different ways.

Accessibility and language reviewers try common registration tasks and record where the assistant blocks, misdirects or confuses users.
The focal point compares failure patterns across access methods and languages rather than relying on one overall completion measure.
The form is simplified, inaccessible controls are replaced, weak language responses are restricted, and uncertain requests are routed to conference staff.
The revised assistant is retested with the same range of tasks before the registration link is circulated.

Result: The assistant is more inclusive because access needs, language differences and varied user contexts shape its design and evaluation. A human route remains available where the automated support does not work.

In a United Nations workplace

A human resources officer in a country office drafts a note for the file in response to an audit observation on AI-supported applicant screening. Before seeking the chief of section's clearance, the officer checks whether the procurement file and recruitment workflow addressed accessibility, language coverage, varied career histories and differential screening results.

Inclusion is not established by applying the same model to every applicant. It requires accessible candidate processes, appropriate representation in evaluation data, consultation with affected users, testing across relevant groups and a route for candidates to seek human review.

Before you try thisDo not upload applications, disability information or other HR records to a tool that has not been approved for such data. The system must not infer protected characteristics or determine the shortlist without accountable human review. Document limitations and provide a way to challenge errors, because exclusion at this stage can affect access to employment.

Where you may hear it

A vendor demo to an ICT governance board
An accessibility review before a public portal is launched
A procurement evaluation of an AI-enabled service
An inter-agency working group testing a multilingual tool
A response to an audit observation on automated recruitment

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Inclusion just means adding more demographic groups to the training data.
Correction: Data representation matters, but inclusion is broader. It also includes language choice, interface design, device and bandwidth limits, disability access, testing methods, feedback channels, and whether people can get human review when needed.
Common misconception: If a system is accessible, it is automatically inclusive.
Correction: Accessibility focuses on whether people with disabilities can use a product or service. Inclusion is wider. A system might meet accessibility needs yet still exclude people because of language, cost, literacy, geography, or poor representation in the data.
Common misconception: A high overall accuracy score proves the system is inclusive.
Correction: An average score can hide large differences between groups. A model may perform well for the majority and poorly for a minority language, accent, region, or user group. Inclusion requires looking at who benefits and who does not.

Quick check

Which example best shows inclusion in AI?

Go deeper

The simple version

Inclusion means making sure an AI system works for different kinds of people and does not leave some groups out by accident.

The technical version

In AI practice, inclusion refers to designing, developing, evaluating and governing systems so affected populations are meaningfully considered across the life cycle. This includes representational coverage in data, participatory or consultative design, usability across languages and contexts, accessibility, and evaluation of differential performance and impact across groups. It is related to fairness and non-discrimination, but it is not identical to either. Fairness often focuses on how outcomes are distributed or whether treatment differs unjustifiably. Inclusion focuses more broadly on whether people can access, use, influence, and benefit from the system, and whether deployment conditions systematically exclude some populations. Inclusion can therefore involve technical measures, such as subgroup testing or better data coverage, and non-technical measures, such as human support channels, localisation, procurement requirements, and ongoing monitoring.

Reviewed external resources

Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation gives an official multilateral framework for ethical AI, and it suits readers who want policy language on inclusion, rights and governance.
Intermediatelink checked 2026-07-31
Primary sourceB-Tech Project (opens in a new tab)UN Human Rights Office (OHCHR)
This OHCHR B-Tech Project material helps readers connect inclusion to business, human rights and technology practice, and it suits programme, policy and risk-focused audiences.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Inclusion (opens in a new tab)Wikipedia
This Wikipedia search results page helps readers explore different meanings and uses of 'inclusion', and it suits beginners who need orientation rather than one fixed definition.
search results, not one pagelink checked 2026-07-31
VideoVideo explanations: Inclusion (opens in a new tab)YouTube search
This YouTube search results page helps readers find varied video explainers on inclusion, and it suits visual learners who want introductory overviews from different speakers.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Inclusion (opens in a new tab)Reddit search
This Reddit search results page shows how people discuss and question inclusion in practice, and it suits readers who want informal examples and debate but will still need to verify claims elsewhere.
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

Inclusion asks who can access the system, who is represented in its design and testing, whose results are worse, and what must change before it is used.