Learn AI /AI Foundations /Narrow AI
Narrow AI
- Starter
- AI Foundations
- Policy officers
Also called: weak AI, task-specific AI.
Back to the milestone you came from
In five seconds
What it is
An AI system built for a limited task or related tasks, rather than broad intellectual work across domains.
In plain words
Narrow AI is AI software for specific tasks, rather than broad intellectual work across domains.
An everyday example
At an airport, one desk only checks boarding passes and passports. The staff there can do that task quickly and accurately for thousands of travellers. But if a traveller asks about lost baggage, visa policy, or how to repair a broken phone charger, that desk cannot handle it and sends them elsewhere. It works well for its assigned task, not for every problem in the airport.
A desk worker is a person with broad real-world abilities, even if their role is narrow. Narrow AI does not secretly have general ability behind the scenes. Also, many AI systems do not follow one simple checklist like the desk does; some use learned statistical patterns from large datasets, which makes their behaviour less transparent than a desk procedure.
Remember thisNarrow AI can perform impressively within its defined task and expected inputs, while remaining unreliable or simply incapable outside that scope.
This term is contested
The general idea is widely used, but the boundary is not fully settled. Newer systems can perform many related tasks, so people may disagree about how narrow they still are and where narrow AI ends and more general capability begins.
Understand it
Understand it
Why it matters
The term helps people ask a basic but important question: what exactly can this system do, and what can it not do? That matters when buying tools, writing policy, setting safeguards, or deciding whether a result needs human review. If people assume a narrow system has broad capability, they may trust it too much, use it on the wrong task, or expose people to avoidable harm. If they recognise its limits, they can match the tool to the job, test it properly, protect sensitive information, and keep accountability with people and organisations rather than with the software.
Narrow AI: strong in one lane, limited outside it
Show why a system can perform well on a defined task without having broad, human-like capability across unrelated tasks.
01
Designed task
The system is built or trained for a specific job, such as translating short text or choosing the next word in a sentence.
Top row spanning both columns: a box labelled 'One narrow AI system'. Under it, add a note: 'Built for a defined task or related task family'.
02
Input fits the task
When the input matches that job, the system can often perform well.
Left column middle: arrows from icons for text, image, or numbers into the system box. Label them 'expected inputs'.
03
Useful output
The result may be a prediction, label, ranking, or generated text.
Left column bottom: arrows out to boxes labelled 'translation', 'spam score', 'next word', or 'detected object'. Use a green accent.
04
Input does not fit
When asked to do something outside its design or training, performance can drop sharply.
Right column middle: arrows from mismatched icons into the same system box. Label them 'unrelated or out-of-scope inputs'.
05
Weak or unreliable result
The system may fail, guess, or produce a confident-looking but unsuitable answer.
Right column bottom: arrows out to boxes labelled 'error', 'poor guess', or 'unreliable output'. Use an amber or red accent.
A two-column comparison. The left column shows one AI system receiving inputs that match its designed task and producing useful outputs. The right column shows the same system receiving unrelated inputs and producing poor, uncertain, or no results. Read top to bottom in each column, then compare the outcomes.
Hover or tap each bottom output box to reveal a short note explaining why it is in-scope or out-of-scope for the same system.
Describe this without the diagram
This comparison diagram explains narrow AI by showing one system in two situations. At the top is a box labelled 'One narrow AI system', with a note that it was built for a defined task or a family of related tasks. In the left column, expected inputs such as text, images, or numbers go into the system. The outputs are useful results, such as a translation, a spam score, a next-word suggestion, or an object label. In the right column, unrelated or out-of-scope inputs go into the same system. The outputs are marked as error, poor guess, or unreliable output. The main point is that the same system can work well inside its intended scope and fail outside it.
With reduced motion: A static side-by-side comparison with no animation. The left side shows matching inputs leading to useful outputs. The right side shows mismatched inputs leading to weak or unreliable outputs.
Narrow AI means an AI system designed for a particular kind of job. That job might be translating text, spotting objects in images, recommending songs, or choosing the next word in a sentence. It can produce useful results in that area, but that does not mean it can handle unrelated tasks reliably.
What goes in depends on the system. Inputs might be text, images, sound, numbers, or clicks from users. The system then applies rules or a trained model. A model is a mathematical system adjusted from examples so it can produce outputs for new inputs. The output could be a label, a score, a prediction, or generated text.
Many current AI tools are narrow AI, even when they seem broad at first glance. A chatbot may answer questions, summarise, translate, and draft messages, but it still works within the kinds of patterns it was built and trained for. If the task changes too much, the system may fail, guess, or produce convincing but wrong output.
People often confuse narrow AI with human-like intelligence. They are not the same. Narrow AI can be very strong at one task and still lack the flexible, general capability people usually imagine in science fiction. The boundary is not perfectly agreed, especially for newer systems that can do many related tasks, so the term is useful but not always precise.
How “The airport desk with one job” maps onto the real thing
| In the story | In the AI system |
|---|---|
| passport and boarding-pass desk | a system built for one defined task or a small set of related tasks |
| travellers' documents | the input data the system receives |
| checked and stamped documents | the output, such as a label, prediction, or generated response |
| sending other questions to another desk | failure outside the system's intended scope |
A worked example
A phone keyboard suggests the next word after someone types: “The meeting starts at 10, please bring the”.
Starting from: Text typed so far: “The meeting starts at 10, please bring the”
- The system reads the sequence of words already typed.
- Using patterns learned from training examples, it assigns scores to possible next words such as “report”, “documents”, “files” and “train”.
- It ranks the candidates and displays the highest-scoring suggestions.
Result: The keyboard suggests “report”, “documents” and “files”. It is performing one defined task, predicting likely next words. It does not understand the purpose of the meeting or acquire a broader ability to organise it.
In a United Nations workplace
In a Resident Coordinator's Office, the records officer has until close of business to file the note for the file after a bilateral. The approved records system examines the title and metadata, then recommends one of the office's predefined filing categories.
The classifier may be very accurate at matching records to known categories, but that capability does not extend to judging the substance of the bilateral, drafting follow-up or deciding who should have access. Its competence is narrow and tied to the records task and examples on which it was trained.
Before you try thisThe officer must not upload the note or its contents to an external or unapproved tool. Bilateral records may contain confidential or sensitive information. Only a system approved for the document's classification may process it, and the officer must verify the category and access controls before filing.
Where you may hear it
- A vendor demo for an electronic records system
- An ICT governance board reviewing a document-classification tool
- A procurement unit's technical evaluation of bids
- An internal audit discussion on automated routing and human oversight
Avoid confusion
Avoid confusion
Often confused with
Artificial general intelligence
A debated label for an AI system with broad capability across many intellectual tasks and domains, often compared with human general ability.
Automation
Automation uses a machine or software to carry out set steps with little or no human action after setup.
Software
A set of coded instructions that tells a computer how to take input, process it and produce output.
Anthropomorphism
Attributing human-like traits, feelings, intentions or understanding to an AI system when its observable output does not establish those inner states.
Common misconceptions
Common misconception: If a system performs very well on one task, it must be intelligent in a general human-like way.
Correction: High performance in one area does not mean broad capability. A system may classify images accurately or generate fluent text while still failing at unrelated tasks, lacking context, or producing errors outside its training and design scope.
Common misconception: Narrow AI only means old, simple systems, not modern chatbots or image tools.
Correction: Many modern AI systems are still considered narrow because they operate within particular kinds of tasks and data patterns, even if those tasks are numerous. Being impressive or large does not automatically make a system generally intelligent.
Common misconception: Narrow AI always uses machine learning.
Correction: Not always. Some narrow systems are rule-based, meaning they follow explicit instructions written by people. Others use machine learning, where a model is adjusted from examples. Both can be narrow if their scope is limited.
Quick check
Which example best fits narrow AI?
An AI tool that classifies emails into predefined categories performs a limited, specific task. The other options describe fixed automation, storage software, or broad general capability, not narrow AI.
Go deeper
Go deeper
The simple version
Narrow AI is AI software for specific tasks, rather than broad intellectual work across domains.
The technical version
Narrow AI refers to systems whose capability is limited to a defined task, domain, or set of related functions. Inputs are processed through rules, statistical models, or learned parameters to produce outputs such as classifications, rankings, predictions, recommendations, or generated content. Performance may be strong within the target distribution, meaning the kinds of examples the system was built or trained for, but it often degrades outside that scope. The term is commonly used in contrast with artificial general intelligence, though the boundary can be blurry for systems that support many related tasks through a shared model.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation gives policy and ethics context for AI governance and suits policy readers, managers, and anyone interested in human rights, oversight, and responsible use.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Narrow AI (opens in a new tab)Wikipedia
This search results page helps readers scan several Wikipedia entries or related pages on the topic, which suits beginners who want a quick starting map rather than a single settled source.
search results, not one pagelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This glossary provides short definitions for machine learning terms and suits readers who want plain technical vocabulary around models, data, and related AI concepts.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Narrow AI (opens in a new tab)YouTube search
This search results page offers a range of video explainers with different teaching styles, which suits visual learners who want introductory overviews rather than one authoritative lesson.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Narrow AI (opens in a new tab)Reddit search
This search results page shows community discussions and questions about the term, which suits readers who want to see common confusions and informal explanations, but it should not be treated as an authoritative source.
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
Narrow AI can perform impressively within its defined task and expected inputs, while remaining unreliable or simply incapable outside that scope.