The labels are useful, but they are not natural laws
Artificial narrow intelligence, artificial general intelligence, and artificial superintelligence are best understood as regions on several capability dimensions—not three precisely marked boxes. Researchers and institutions use different definitions, and there is no scientific authority that can certify that a system has “become AGI.” A careful reader should therefore ask what a speaker means by the term and what evidence supports the classification.
Artificial narrow intelligence (ANI) traditionally means a system optimized for a bounded task or domain: a chess engine, fraud detector, speech recognizer, or protein-structure predictor. Such a system can exceed every human at its target task while remaining useless elsewhere. But calling every deployed model “narrow” is increasingly misleading. General-purpose language and multimodal models can write, code, analyze images, use tools, and adapt to many tasks from instructions. They are broad in one sense, yet unreliable, dependent on scaffolding, and uneven across domains. “General-purpose AI” or “broadly capable model” is often a more informative description than forcing these systems into the old narrow/general binary.
Artificial general intelligence (AGI) usually denotes a machine with broad competence that transfers to unfamiliar tasks. Beyond that core, definitions diverge. Some require performance comparable to a typical human across most cognitive work; some require the ability to learn efficiently; some emphasize economically valuable labor; and some add autonomy in the physical or digital world. Google DeepMind researchers proposed describing AGI by both breadth—how many tasks it can do—and depth—how well it performs—while treating autonomy as a separate deployment property (DeepMind, Levels of AGI). That framework avoids pretending that one benchmark creates a magical threshold.
Artificial superintelligence (ASI) is a hypothetical system that substantially exceeds the best humans across nearly all cognitively valuable domains. This is not an observed technology category. It is a scenario used to reason about systems that could outperform human institutions in science, strategy, engineering, persuasion, and perhaps AI research itself. Whether such a system is possible, how quickly it might follow AGI, and whether intelligence across every domain can be combined in one system remain disputed questions.
Capability has more than one axis
A useful assessment separates at least five properties:
- Breadth: the variety of tasks and environments in which the system is competent.
- Performance: whether it matches a novice, a typical worker, an expert, or the best human.
- Reliability: how often it succeeds, including under unusual or adversarial conditions.
- Learning and transfer: how readily it handles a genuinely new task without task-specific retraining.
- Autonomy: how long and how safely it can plan, act, use tools, recover from mistakes, and pursue a goal with limited supervision.
These axes can separate dramatically. A model may answer questions across hundreds of subjects while making elementary mistakes. Another may be exceptionally reliable in one scientific workflow but unable to generalize. A conversational model may appear broadly knowledgeable while an agent built around the same model fails because it loses context, misuses a tool, or cannot recognize that its plan is going wrong. The OECD’s definition of an AI system similarly distinguishes generated outputs and influence on environments from varying levels of autonomy and post-deployment adaptiveness (OECD AI principles).
This matters because polished language is not the same as robust agency. Humanlike conversation encourages anthropomorphism, while benchmark scores can hide dependence on carefully structured prompts, clean data, and automated grading. Conversely, a system need not converse like a person to have consequential capabilities. A scientific model, trading agent, or cyber tool may matter because of what it can do, not because it resembles a human mind.
Why no single benchmark proves AGI
Benchmarks sample behavior under specified conditions. They do not directly reveal understanding, consciousness, intentions, or performance in every untested setting. Scores may rise because of genuine generalization, exposure to similar training examples, better prompting, tool access, or engineering around the model. A benchmark can also saturate: once leading systems approach its ceiling, small differences become difficult to interpret.
The same warning applies to economic definitions. “Can automate most jobs” sounds concrete, but jobs are bundles of technical tasks, social negotiation, accountability, physical activity, local knowledge, and responsibility for errors. Success on self-contained tasks does not establish that a system can replace a worker in a messy organization. METR explicitly says its task-completion time horizon is concentrated in software, machine learning, and cybersecurity and should not be read as a general measure of job automation (METR methodology).
AGI should therefore be treated as a claim requiring a profile of evidence. Which domains were tested? How novel were the tasks? What tools and human help were allowed? Was performance replicated independently? What was the success rate? Did the system remain dependable when goals were ambiguous or the environment changed? Without those details, an AGI announcement is branding, forecasting, or personal judgment—not a measurement result.
Why the distinction changes policy
Different capability profiles create different risks. Present systems already cause documented harms through fraud, discrimination, privacy invasion, unsafe advice, labor disruption, and synthetic media. Those problems do not require AGI and should not be deferred until a future threshold. Broadly autonomous systems could add other risks by conducting cyber operations, accelerating dangerous research, or controlling critical processes at a scale that exceeds human review. Hypothetical superintelligence raises still more uncertain questions about whether human institutions could supervise a system strategically superior to them.
The categories also shape proportionate safeguards. A narrow medical classifier may need clinical validation, monitoring, and an appeal process. A general-purpose model may need capability evaluations, abuse detection, incident reporting, and controls around dangerous tools. A highly autonomous system operating critical infrastructure would justify stronger assurance, containment, and human-authorization requirements. Using “AI” as one undifferentiated category conceals these differences.
A better way to read AGI claims
When someone says AGI is here—or impossible—translate the claim into observable questions. Is the system broadly competent or merely broad in interface? Does it transfer to new work? How reliable is it? Can it complete long, real-world projects? Does it require continuous human correction? Does it learn after deployment? Is the evidence public and independently reproducible?
The honest conclusion in 2026 is not that all deployed AI is simply narrow, nor that broad language models have crossed a universally accepted AGI line. Current systems combine striking breadth with serious brittleness. AGI remains a contested concept, and ASI remains hypothetical. The most useful “capability ladder” is therefore a multidimensional map with dated evidence—not a three-rung staircase on which a company can declare itself the winner.