presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Foundations /Artificial general intelligence

    Artificial general intelligenceAGI

    Also called: AGI, general AI, strong AI.

    Detail

    In five seconds

    What it is

    A debated label for an AI system with broad capability across many intellectual tasks and domains, often compared with human general ability.

    In plain words

    In simple terms, AGI is the idea of one AI system that could handle many very different mental tasks, instead of just doing one narrow job well.

    An everyday example

    Imagine a busy office in Panama City, Bogotá, and Geneva that all use different kinds of forms. One worker can already match a photograph of a damaged paper form to the right template very well. Another worker can also sort invoices, draft polite replies, spot missing fields, explain the filing rules to a new colleague, and adapt when the form layout changes, even with little extra instruction. People might call the second worker “general-purpose” because the same person can handle many different office tasks, not just one. But the comparison only goes so far, because a human worker brings real-world experience, judgement, and accountability.

    Where the story stops working

    A human office worker has lived experience, social understanding, legal responsibility, and can ask clarifying questions in the real world. AGI is only a label for a possible technical system with broad performance, not proof of human-like understanding, judgement, or consciousness. Also, office work is only one slice of the tasks people imagine; AGI discussions usually reach much wider.

    Remember thisAGI is a contested label for systems expected to transfer capability across many unfamiliar tasks with limited retraining. No single benchmark settles the claim, and the term does not imply consciousness.

    This term is contested

    The meaning of AGI is genuinely disputed. 'AGI', 'general AI', and 'strong AI' are often used interchangeably, but not consistently, and there is no universally accepted test or threshold.

    Understand it

    Understand it

    Why it matters

    The term matters because it shapes expectations, budgets, regulation debates, procurement decisions, and risk discussions. If leaders hear 'AGI' and assume a system can reliably handle almost any knowledge task, they may overtrust it. If they dismiss the term entirely, they may miss real changes in capability. For non-specialists, the key issue is practical: what evidence supports the claim, what tasks were tested, what still fails, and what safeguards remain necessary. In policy and management settings, precise language helps separate current tools from future speculation.

    From narrow AI to claims about AGI

    Show the difference between a system that performs selected tasks well and a system claimed to transfer across many tasks with limited extra training, and why the boundary is disputed.

    1. 01

      Narrow tasks

      Show separate systems or evaluations for translation, coding, image classification, and form-template matching. Each task has its own route from input to output.

      Draw four small input-output cards in the left column, each labelled with one task and one output, using distinct colours and no connecting arrows between tasks.

    2. 02

      Broader input

      Show one candidate system on the right receiving text, images, audio, numbers, and tool access.

      Draw one larger central box in the right column with five arrows entering from icon-labelled inputs: document, photo, waveform, table, and tool.

    3. 03

      Transfer to new tasks

      Show the same candidate system being asked to perform several different tasks, including one not seen before, with only a small note saying 'limited extra training'.

      From the central box, draw arrows to five output cards: explanation, plan, code, classification, and 'new task', with a small badge reading 'limited extra training'.

    4. 04

      Debated boundary

      Show that experts disagree on where 'generally capable' begins. This is not a settled line.

      Add a bottom band spanning both columns with a dashed gradient line labelled 'no agreed threshold for AGI' and small speech bubbles saying 'human-level on most tasks?' and 'broad transfer across domains?'.

    A side-by-side comparison. The left column shows today’s narrow systems handling separate tasks. The right column shows the stronger claim behind AGI: one system receives varied inputs, adapts across domains, and produces many kinds of outputs. A final row shows that no single agreed test marks the boundary, so the right column remains a debated category.

    Hover or tap each row to highlight the contrast between separate task-specific capability and a broader cross-task claim; the dashed boundary band shows a tooltip explaining that AGI has no universally accepted test.

    Describe this without the diagram

    This diagram compares two ideas. On the left are examples of narrow AI: separate systems or separate evaluations for translation, coding, image classification, and matching a photographed form to a template. Each task is shown as its own input-output path. On the right is the stronger idea often called AGI: one candidate system takes many kinds of input, including text, images, audio, numbers, and tools, then produces different kinds of outputs such as explanations, plans, code, classifications, and a response to a new task with limited extra training. Across the bottom, a dashed line says there is no agreed threshold for AGI. The main message is that AGI refers to broad transfer across tasks, but the boundary is disputed.

    With reduced motion: A static two-column comparison with labelled arrows and a dashed note at the bottom stating that no single agreed test defines AGI.

    People use “artificial general intelligence”, “AGI”, and “general AI” for a proposed kind of AI that could handle a wide range of tasks, not just one narrow job. “Strong AI” is often used in the same conversations, but it can mean different things. Some people use it as a synonym for AGI. Others use it for an even stronger claim, such as human-like mental abilities or awareness, which is more controversial.

    What would go in? In principle, many kinds of input: text, images, numbers, speech, software tools, and new task instructions. What would happen? The system would need to adapt what it has learned to unfamiliar problems, move between domains, and keep working with limited extra training. What would come out? Useful performance across many activities such as writing, planning, analysing, coding, explaining, and perhaps controlling tools.

    Today’s widely used systems may support many tasks, but broad usefulness alone does not establish AGI. A model that writes reports, answers questions, or matches a photograph of a form to the right template may be very capable without meeting a proposed threshold for AGI.

    This is where confusion starts. AGI is not the same as generative AI, which produces content such as text or images. It is not the same as a frontier model, which means a leading-edge model by capability, not necessarily a generally capable one. It is also not the same as consciousness. A system may appear fluent or flexible without there being evidence that it has awareness.

    Because no agreed test exists, people disagree about where the boundary lies. For some, AGI means “as good as humans at most cognitive tasks”. For others, it means “able to learn and transfer across domains with little task-specific engineering”. So the term is useful, but it is also contested and should be used carefully.

    How “One office helper for many kinds of paperwork” maps onto the real thing

    In the storyIn the AI system
    worker who only matches form photos to templatesnarrow AI built for a specific task
    worker who can switch among several office tasksthe idea behind AGI: broad capability across tasks
    new form layout with little extra instructionadapting to unfamiliar tasks with limited retraining
    different offices using different paperworkdifferent domains and input types

    A worked example

    Before a vendor demonstration to an ICT governance board, an evaluation officer in a regional bureau tests one assistant on six controlled tasks, using public or synthetic material and no task-specific fine-tuning.

    Starting from: Tasks: summarise a progress report, answer factual questions, write a short script, classify document images, translate brief passages and plan a route using a map tool.

    1. The officer gives the same system natural-language instructions and the material needed for each task, without fine-tuning it separately.
    2. Outputs are scored against pre-agreed rubrics and compared with suitable specialist tools or human-produced reference answers.
    3. The officer then changes an unfamiliar element, such as the map tool’s output format, to test whether the system can transfer its procedures rather than repeat a familiar pattern.
    4. The system performs well on five tasks but produces an invalid route after the format change.
    5. The evaluation records broad cross-task performance alongside the brittle failure and does not treat the run as a definitive test of AGI.

    Result: The system shows useful breadth, but the failed transfer test exposes a limit in generalisation. The evidence may support a claim of broad capability under defined conditions, not a settled claim that the system is AGI.

    In a United Nations workplace

    At a liaison office, a chief of section is reviewing how free-text answers from a needs assessment have been grouped before the monitoring report goes for clearance. A vendor says its assistant is “approaching AGI” because the same product can propose themes, summarise responses, translate short passages and draft a results table.

    Doing several related tasks in one workflow is not, by itself, evidence of AGI. The useful questions are whether the system can transfer what it has learned to genuinely unfamiliar tasks, how much retraining or hand-engineering that requires, how it performs when formats change, and where human review remains necessary.

    Before you try thisRaw needs-assessment answers must not be entered into a tool that has not been approved for them. They may contain personal, protection-sensitive or otherwise confidential details, and a claim of broad capability provides no privacy or security assurance. Even in an approved environment, use only the minimum necessary data and follow the office’s data protection guidance.

    Where you may hear it

    • At an ICT governance board reviewing a general-purpose assistant
    • During a vendor demonstration in a procurement meeting
    • In the margins of a General Assembly discussion on AI governance
    • At a research briefing for a chief of section on frontier models

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • AGI means any AI that seems impressive or can chat fluently.

      Fluency and broad usefulness are not the same as general capability. Many current systems can produce polished language across topics, yet still fail on tasks requiring robust transfer, reliable reasoning, long-horizon planning, or adaptation outside their training patterns. Calling every advanced chatbot AGI lowers the meaning of the term.

    • AGI and consciousness are the same thing.

      They are different claims. AGI usually refers to breadth of task performance. Consciousness refers to subjective experience or awareness, which is a separate and highly contested question. A system could be described as broadly capable without evidence of awareness, and claims of awareness need much stronger justification than impressive output.

    • Once AGI exists, human oversight will no longer be needed.

      Even a very capable system could still be wrong, biased by its training data, poorly aligned with instructions, insecure, or unsuitable for sensitive contexts. Organisational responsibility, legal accountability, and human rights duties do not disappear because a tool performs well on many tasks.

    • AGI is a clearly defined technical milestone with a universally accepted test.

      There is no single global definition or exam that everyone accepts. Some researchers focus on human-level performance across most cognitive tasks. Others focus on transfer, adaptability, or economic usefulness. That is why discussions about whether a system 'counts' as AGI often turn into debates about definitions.

    Quick check

    Which statement best matches careful use of the term AGI?

    Go deeper

    Go deeper

    The simple version

    In simple terms, AGI is the idea of one AI system that could handle many very different mental tasks, instead of just doing one narrow job well.

    The technical version

    Technically, AGI is a contested umbrella term for a hypothetical artificial system with generalisable capability across a wide range of cognitive tasks and domains. The core idea is not just high performance on one benchmark or one product category, but transfer: using learned representations and procedures across unfamiliar tasks with limited task-specific retraining or hand-engineering. Different definitions emphasise different thresholds, such as human-level performance on most economically valuable tasks, broad competence across multimodal inputs and outputs, strong out-of-distribution adaptation, or general problem-solving under changing goals and environments. In current usage, 'general AI' and 'AGI' are often near-synonyms, while 'strong AI' may either be used loosely as another synonym or more strongly to imply human-like mental properties, which makes it especially easy to confuse with machine consciousness. None of these terms guarantees a particular architecture, training method, or benchmark, and none has a universally accepted operational definition.

    Reviewed external resources

    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

    AGI is a contested label for systems expected to transfer capability across many unfamiliar tasks with limited retraining. No single benchmark settles the claim, and the term does not imply consciousness.