A system can perform exceptionally well at one task while remaining unable to handle a related task without new software, data, or training. Another system can work with text, images, and code but still make unreliable decisions in a particular business process. Both may be called intelligent. The label alone leaves important differences unexplained.
The familiar distinction between narrow AI, artificial general intelligence, and superintelligence concerns the breadth and level of capability. It provides useful language for research and long-term discussion, but the terms are not a complete specification of a product. Understanding what each term describes makes it easier to ask which capabilities have been demonstrated and what the system is authorised to do with them.
Narrow AI Describes a Bounded Scope of Capability
Narrow AI generally refers to systems developed or evaluated for particular tasks or domains. Examples include a model that classifies defects, a forecasting service for electricity demand, or a system that recommends a route within a defined network. Such systems can be extremely capable within their scope.
Narrow does not mean incapable of generalising. A defect classifier is useful partly because it can recognise relevant patterns in images it did not encounter during training. Its scope is limited by the tasks, inputs, conditions, and performance for which there is evidence. Generalisation within that scope is different from an ability to acquire arbitrary new skills.
Nor does narrow mean simple or harmless. A specialist model can contain many parameters and influence consequential decisions. A narrow credit model or medical classifier still needs evaluation appropriate to its use. Risk depends on consequences, access, and operating arrangements as well as capability breadth.
When considering a product, ask for a description of its supported tasks and evaluated conditions. “Narrow AI” is a starting category; the useful specification identifies what the service predicts or produces, for whom, with what evidence, and under which limitations.
Foundation Models Complicate the Single-Task Picture
A model trained on broad data may support many downstream tasks through prompting, adaptation, or additional training. The foundation-model research report by Bommasani and colleagues uses this idea to describe models that can underpin a range of applications. A broad training base is different from building a separate model from scratch for every task.
Language models can support drafting, summarisation, classification, coding, and other activities through a shared interface. Multimodal systems can process more than one kind of input or output. These capabilities make the simple phrase “does only one thing” a poor description of some systems commonly grouped below AGI.
Their breadth does not establish uniform competence. A model may perform strongly on a familiar coding task and poorly on a specialised document interpretation. Some apparent flexibility depends on surrounding tools, retrieved information, or task-specific adaptation. Evaluate the complete application while identifying which component contributes which capability.
Be careful with claims about unseen tasks. A model’s training history may include related examples or formats even when a demonstration supplies no explicit training step. Evidence of successful performance can still be valuable, but claims about entirely novel learning require information about prior exposure that may not be available.
The AI and toddlers article examines the same accounting problem in few-shot learning. The number of examples shown in a demonstration is not necessarily the system’s complete learning history.
AGI Concerns Broad Competence and Adaptability
Artificial general intelligence usually describes a proposed level of broad cognitive capability, often discussed in relation to human performance or the ability to learn across domains. Definitions differ in which tasks count, how performance is measured, and what adaptability is required. A claim about AGI should therefore state its criteria.
Morris and colleagues’ Levels of AGI framework proposes distinguishing performance and generality, while also considering autonomy separately. It is a research framework for clearer discussion, not a universal certification scheme. Its value is in making the dimensions of a claim more explicit.
Human comparison also needs a reference. An average adult, a trained professional, and the best specialist do not provide the same benchmark. No individual person performs every conceivable task expertly. A definition that compares AI with humans must specify the relevant population, task range, conditions, and opportunity to learn.
This does not make AGI a meaningless concept. It means evidence should match the definition being used. A broad benchmark suite, a successful demonstration, or an economic forecast may each address part of a proposal, while leaving other requirements unresolved. Disagreement about terminology should lead to clearer evaluation rather than automatic acceptance or dismissal.
AGI should not be treated as a synonym for consciousness or moral status. A definition based on performance does not establish subjective experience. Those questions require distinct arguments and evidence. The neural-networks and brain comparison explains why biological vocabulary and impressive behaviour do not settle them on their own.
Superintelligence Describes a Stronger Capability Claim
Superintelligence commonly refers to an intellect that substantially exceeds human cognitive performance across a broad range of important domains. In Nick Bostrom’s discussion of advanced AI, the comparison extends beyond calculation speed to areas such as scientific creativity and social capability. This is a conceptual definition, rather than an ordinary product category with an agreed acceptance test.
Exceptional performance on one task is not the same claim. A system can exceed people at a specialised game or numerical operation without thereby demonstrating broad superiority in planning, learning, or other domains. State whether “superhuman” refers to one evaluated capability or to a much wider hypothesis.
Discussion of superintelligence also includes uncertain scenarios about how capabilities might develop and interact. A proposed path is not a guaranteed timetable. Estimates depend on assumptions about learning, compute, available data, integration, and the performance threshold being considered.
Greater capability does not by definition establish safe objectives, suitable permissions, or reliable control. Those properties need their own analysis. Conversely, present-day risks do not require a system to meet a superintelligence definition: a limited system with inappropriate access can already create serious consequences.
Keep Capability, Autonomy, and Reliability Separate
Capability concerns what a system can do. Autonomy concerns how much of a process it can carry out without further intervention. Reliability concerns how consistently it meets requirements under relevant conditions. These dimensions can vary independently enough that one label should not stand in for all three.
A specialist classifier could automatically reject products on a production line. A broad language model could be restricted to suggesting text that a person reviews. The first has narrower scope but may have more immediate operational authority. The second may be more flexible while holding less power to change external systems.
| Dimension | Question to make explicit |
|---|---|
| Breadth | Which tasks and domains are supported? |
| Performance | How well does it meet the relevant standard? |
| Adaptability | What can it learn or handle beyond familiar cases? |
| Autonomy | Which steps can it execute without another decision? |
| Reliability | Under which conditions does performance remain acceptable? |
| Authority | Which records, services, and actions can it access? |
Determinism is another separate property. A model can estimate probabilities while using a deterministic decision rule, and a generative application can sample outputs. Repeating the same output does not guarantee correctness. Variation in output does not automatically make a system unsuitable; the task and controls determine what variation can be accepted.
Interpretability and integration should be assessed independently too. A system’s breadth does not tell the organisation whether it can reconstruct a decision, inspect relevant evidence, or recover from a failed action. The guardrails guide connects permissions and validation with those operational consequences.
Examine What a Capability Claim Actually Demonstrates
For any claim, identify the task, inputs, evaluation set, allowed assistance, and success condition. A system that solves a problem with access to a search engine or code executor has demonstrated the performance of that arrangement. It should not be described as though every result came from the model alone.
Ask how representative the tests are of intended use. Familiar task formats, possible overlap with training material, and selective demonstrations can limit the conclusion. A held-out evaluation can provide meaningful evidence, but no finite test establishes performance across every conceivable future circumstance.
Assess difficult and unsupported cases as well as successful ones. Can the system recognise missing information, decline an unsuitable task, or request help? Those behaviours need evaluation in context; they cannot be inferred from the generality label or from a reassuring self-description.
The predictive evaluation guide shows how information availability and test design affect conclusions in a bounded domain. Similar discipline applies to broader capability claims, even where the exact evaluation methods differ.
Use the Taxonomy at the Right Level
Narrow AI, AGI, and superintelligence help frame questions about breadth, performance, and possible future systems. They become less useful when a label is used to skip the description of an actual service. Neither calling a model narrow nor describing it as approaching AGI resolves the requirements of a business application.
A decision maker should be able to connect the claim with demonstrated tasks, operating limits, permissions, and consequences. Research can continue to refine the broader categories while organisations evaluate the capabilities they intend to use. That leaves the terminology informative without asking it to answer questions it was never designed to settle.
