Learn AI /Responsible AI, Human Rights and Safety /Explainability
ExplainabilityXAI
- Intermediate
- Responsible AI, Human Rights and Safety
- Policy officers
- Human rights practitioners
Also called: explainable AI, XAI.
Back to the milestone you came from
In five seconds
What it is
Explainability means being able to show, in human-usable terms, why an AI system produced a particular output or pattern of outputs.
In plain words
In simple terms, explainability means giving people a usable reason for an AI result so they can check it instead of just accepting it.
An everyday example
A city office receives 500 comments from a public consultation. An assistant sorts each comment into tabs such as housing, transport, health, or "other". When the manager asks why one comment was placed under transport, the assistant points to the phrases "bus route", "travel time", and "fare increase" and says those phrases drove the choice. The manager can then judge whether the sorting makes sense or whether the comment was misunderstood. Without that explanation, the label alone is much harder to trust or correct.
A human assistant can usually describe a reason in ordinary language because the person followed a conscious sorting process. Many AI models do not have a single neat reason stored inside them. Explanations are often generated after the output, using methods that estimate which inputs mattered. So the analogy helps with the purpose of explainability, but not with the exact mechanism or certainty of the explanation.
Remember thisExplainability makes an output or pattern open to human scrutiny. It can support checking and challenge, but a convincing explanation is not proof that the model is correct or fair.
This term is contested
The exact boundary between explainability and interpretability is not fully settled. Different research and policy communities use the terms differently, especially when discussing whether explanations come from directly understandable models or from post hoc methods.
Understand it
Understand it
Why it matters
If an AI system affects people, resources, or risks, a bare output is often not enough. People may need to ask why one case was flagged, why one message was prioritised, or why one applicant received a higher score. Explainability helps users test whether the output fits the evidence, spot errors or bias, and recognise when the system should not be trusted in that context. It also supports accountability: if no one can give a usable reason, it is harder to challenge the result, correct harm, or show that human rights, fairness, and safety were taken seriously. For non-specialists, this matters because explainability shapes whether AI can be used responsibly, especially where decisions need human oversight rather than automatic acceptance.
From input to output to explanation
Show that explainability adds a human-checkable account of what influenced an AI output, so people can review, challenge, or override it.
01
1. Input enters the system
Text, images, numbers, or other data are provided to the model.
Draw a box labelled 'Input' with small icons for document, image, and chart; add a right-pointing arrow to the next box.
02
2. Model produces an output
The model classifies, scores, ranks, or generates a response.
Draw a box labelled 'Model processing' leading to a box labelled 'Output' with examples like 'label', 'score', and 'summary'.
03
3. Explanation shows what influenced it
A separate panel shows the most influential words, variables, regions of an image, or comparison cases.
From the 'Output' box, draw an arrow to a side box labelled 'Explanation' containing highlighted words, a short feature list, and a note 'why this result'.
04
4. Human reviews and decides next step
A person checks whether the reason is plausible, fair, and sufficient for the context.
Below the flow, add a review lane with a person icon and three buttons: 'accept', 'question', 'escalate/override'; connect the 'Explanation' box downward to this lane.
A left-to-right flow diagram with four boxes: input, model processing, output, and explanation for humans. A review lane below shows a person checking whether the explanation is sufficient and whether action should be taken, questioned, or stopped. Read from left to right, then down to the review lane.
Hovering or tapping each stage reveals one example explanation format: highlighted text, feature importance bars, image heatmap, or similar-case comparison.
Describe this without the diagram
The diagram shows a sequence with human review. First, input data such as a document, image, or numbers enters an AI system. Second, the model processes the input and produces an output like a label, score, ranking, or summary. Third, an explanation panel shows what most influenced that output, such as highlighted words, important variables, marked parts of an image, or similar examples. Fourth, a human reviews both the output and the explanation and then chooses to accept the result, question it, or override and escalate it. The main message is that explainability does not replace human judgement; it supports it.
With reduced motion: A static left-to-right diagram with four labelled boxes and arrows, plus a lower review lane showing a human checking the explanation before accepting, questioning, or overriding the output.
Explainability is about giving people reasons they can inspect and question. In an AI system, the input might be text, images, numbers, or other data. The system processes that input and produces an output, such as a label, score, ranking, summary, or recommendation. Explainability adds information about what most influenced that output.
That extra information can take different forms. It might highlight words in a document that pushed a classification, show which variables mattered most in a risk score, or compare the actual case with similar past cases. In simple terms, it tries to answer: "What parts of the input mattered, and in what way?"
This is often confused with transparency, but they are not the same. Transparency is broader: it can include who built the system, what data was used, what limits are known, and who is accountable. Explainability is narrower: it focuses on making the system's behaviour or output more understandable to people.
It is also often confused with interpretability. Interpretability focuses on inspecting how a model's internal parts contribute to its outputs. Explainability focuses on giving people a human-usable account of an output or pattern of behaviour. The terminology varies, and the practices can overlap. Both matter because people need to check errors and challenge outcomes. Neither one by itself gives a person authority to change or stop an AI-supported action.
How “Sorting public comments with coloured tabs” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the public comments | the input text given to the model |
| the tab chosen for each comment | the model output or classification |
| the highlighted phrases like "bus route" | features or parts of the input that most influenced the output |
| the manager checking the assistant's reason | human oversight of the model's output and explanation |
A worked example
A programme officer in a regional bureau uses an approved text classifier to prepare a briefing note from public feedback before front-office clearance.
Starting from: Comment: "The bus fare has risen again and the evening service is unreliable, so getting to the clinic is harder."
- The classifier assigns scores of 0.78 for transport, 0.15 for health and 0.07 for cost of living.
- A local explanation identifies "bus fare" and "evening service" as the strongest influences on the transport label. "Getting to the clinic" contributes to the health score.
- The programme officer checks the explanation against the full comment. It supports the main label but also shows that a single-label summary would obscure the health-access concern.
- The officer retains transport as the main theme, adds health access as a secondary tag and notes the classifier's single-label limitation in the file.
Result: The explanation made the output reviewable rather than merely plausible. It did not decide the final classification, but helped the programme officer identify what the model had captured and what it had missed.
In a United Nations workplace
At a small field presence with unreliable connectivity, the duty officer is finalising a side-event concept note for clearance by close of business. An approved classifier has grouped public consultation comments by theme, and its explanation view shows which words most influenced each label.
The duty officer can check whether the apparent pattern reflects the comments or merely repeated wording. Local explanations help challenge individual labels, while a global view can reveal whether the classifier is systematically overlooking a theme before figures are used in the concept note.
Before you try thisAn explanation does not prove that an output is correct or reveal the model's full reasoning. The duty officer must review the underlying comments and record material limitations. Personal or protection-sensitive submissions must not be entered into any tool that has not been approved for that information.
Where you may hear it
- A vendor demo before an ICT governance board
- A procurement meeting reviewing an AI evaluation matrix
- An internal audit interview on an automated scoring tool
- A Human Rights Council side event on AI due diligence
- A first reading of guidance on responsible AI
Avoid confusion
Avoid confusion
Often confused with
Transparency
Being open about where an AI system is used, what it does, what information it uses, and who remains responsible for its effects.
Interpretability
Interpretability is the extent to which people can understand how a model works or produces outputs, sometimes by inspecting its internal parts.
Meaningful human control
People with real authority must have enough information and practical ability to shape, limit, change or stop AI-supported actions, especially in high-risk situations.
Side-by-side
Common misconceptions
Common misconception: If a system gives an explanation, the output must be reliable.
Correction: An explanation can be helpful and still be incomplete, misleading, or too simple. Some explanation methods only estimate which inputs mattered. People should treat explanations as evidence to inspect, not proof that the output is correct.
Common misconception: Explainability means revealing the full source code or model internals to everyone.
Correction: Sometimes deeper technical access is useful, but explainability often aims at a different question: can the affected person, reviewer, or operator get a usable reason for an output? In many settings, meaningful explanation can exist without publishing every internal detail, especially where security, privacy, or intellectual property also matter.
Common misconception: Highly accurate systems do not need explainability.
Correction: Even a system with strong average performance can fail on particular groups, unusual cases, or changing conditions. Explanation helps people detect those failures and decide when a human should intervene. Accuracy and explainability address different questions.
Quick check
Which statement best describes explainability in AI?
Explainability is about giving people usable reasons or clues about what influenced an output. That supports review and accountability, but it does not guarantee correctness, does not always require full public disclosure, and does not replace human oversight.
Go deeper
Go deeper
The simple version
In simple terms, explainability means giving people a usable reason for an AI result so they can check it instead of just accepting it.
The technical version
Technically, explainability refers to methods that make an AI system's outputs or behaviour more understandable to humans. Depending on the system, this can include feature importance measures, local explanations for a single prediction, global explanations of overall model behaviour, counterfactual examples showing what changes would alter an output, saliency visualisations, rule extraction, and surrogate models. Interpretability instead focuses on inspecting how internal model parts contribute to outputs, although terminology varies and the practices can overlap. A key technical caution is that many explanation techniques are approximations of behaviour, not guaranteed statements of a model's true internal mechanism. Their usefulness depends on the audience, the decision context, and whether the explanation is faithful, stable, comprehensible, and actionable.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
UNESCO's Recommendation on the Ethics of Artificial Intelligence gives policy-oriented principles and governance context, and suits readers looking for an official international framework.
Intermediatelink checked 2026-07-31
Primary sourceB-Tech Project (opens in a new tab)UN Human Rights Office (OHCHR)
The OHCHR B-Tech Project provides a human rights lens on technology governance and responsibility, and suits readers concerned with impacts on people and accountability.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Explainability (opens in a new tab)Wikipedia
A Wikipedia search results page helps readers find background articles and related terms on explainability, and suits readers who want a broad starting point rather than one authoritative source.
search results, not one pagelink checked 2026-07-31
VideoVideo explanations: Explainability (opens in a new tab)YouTube search
A YouTube search results page helps visual learners find varied video explanations of explainability, and suits readers who prefer short audiovisual introductions.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Explainability (opens in a new tab)Reddit search
A Reddit search results page shows community discussions and practical questions about explainability, and suits readers who want informal debate but should verify claims carefully.
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
Explainability makes an output or pattern open to human scrutiny. It can support checking and challenge, but a convincing explanation is not proof that the model is correct or fair.