Defining AGI

Artificial General Intelligence (AGI) is a term for hypothetical computational systems with broad, transferable cognitive competence that matches or exceeds human capabilities across a wide range of intellectual domains. There is no agreed definition; Morris et al. (2023) review several competing ones. The wording above is this encyclopedia's working definition, a synthesis rather than a quotation from any single source; related proposals appear in Goertzel and Pennachin (2007) and Legg and Hutter (2007). Narrow AI systems, by contrast, are built for bounded problem spaces; AGI is usually taken to require flexible learning and adaptation to unfamiliar tasks, though which properties are necessary is itself disputed (see AGI-DEF-001).

The concept of AGI represents both the founding aspiration of artificial intelligence and the central inflection point in contemporary technological forecasting. Whereas narrow AI achieves superhuman performance in insular domains (e.g., chess, protein folding, image segmentation) by exploiting structured constraints, such systems often generalize poorly beyond their training distribution.

If AGI were achieved, proponents argue it could have far-reaching consequences; these are scenarios, not established outcomes:

1. Economic Discontinuity: Proponents argue that broad cognitive substitutability across knowledge work could decouple economic productivity from human demographic constraints; the extent is uncertain and debated.
2. Scientific Acceleration: Automated hypothesis formulation and experimental iteration could in principle speed research; the magnitude (including claims of compressing decades into years) is speculative.
3. Control and Alignment Challenge: Creating an agent possessing broad strategic intelligence is argued by Russell (2019) to require that the agent's objectives be designed to remain beneficial to humans; whether alignment is solvable, and whether any guarantee is attainable, are open questions.

In the 1950s, early pioneers did not distinguish between "narrow" and "general" AI; the initial objective was the complete emulation of human intelligence. Herbert Simon is widely quoted as predicting in 1965 that machines would within twenty years be capable of doing any work a man can do [VERIFICATION NEEDED: exact wording and page in Simon, The Shape of Automation for Men and Management (1965), not checked].

As early symbolic systems confronted combinatorial explosion and the brittleness of heuristic search, research fragmented into specialized subfields (computer vision, natural language processing, robotics). To reassert the pursuit of holistic, domain-general cognitive architectures, the term Artificial General Intelligence was used by Mark Gubrud in 1997 [VERIFICATION NEEDED: Gubrud (1997) source not checked] and popularized by Ben Goertzel and Cassio Pennachin (2007).

Marcus Hutter's AIXI agent is a theoretical construct from his earlier work on universal artificial intelligence [VERIFICATION NEEDED: Hutter (2005), Universal Artificial Intelligence, not checked]. Shane Legg and Marcus Hutter (2007) proposed a universal intelligence measure grounded in Kolmogorov complexity and relate it to theories of universal optimal learning agents (Legg & Hutter, 2007). AGI-oriented laboratories emerged in the 2010s, including DeepMind (founded 2010) and OpenAI (founded 2015) [VERIFICATION NEEDED: founding years not checked against primary sources]; earlier AGI-focused groups also existed.

The emergence of compute-scaled autoregressive transformers (2020 onward) was reported to show substantial task generalization across language, software engineering, and multi-step reasoning, reigniting mainstream scientific and geopolitical debates regarding the proximity and feasibility of AGI.

#4.1 Formal Mathematical Formulation: Universal Intelligence

Legg and Hutter (2007) proposed a measure $\Upsilon(\pi)$ of an agent's general intelligence as its expected performance across a class of computable environments $\mathcal{E}$, weighted by their algorithmic simplicity. The following is a schematic form; [VERIFICATION NEEDED: the exact definition of the value function $V_\mu^\pi$ (reward normalization and any discounting) in the original paper was not checked against its text]:

where:

• $\pi$ denotes the agent's policy.
• $\mu$ represents an environment drawn from the class of computable distributions $\mathcal{E}$.
• $K(\mu)$ is the Kolmogorov complexity of environment $\mu$ (penalizing overly complex, contrived worlds).
• $V_\mu^\pi = \mathbb{E} \left[ \sum_{t=1}^{\infty} \gamma^t R_t \right]$ is the expected cumulative reward achieved by policy $\pi$ in environment $\mu$.

Kolmogorov complexity is incomputable, so this measure cannot be computed exactly and the related AIXI agent is a theoretical construct. The mathematics is well defined, but whether the measure is an adequate definition of intelligence is contested; for example, the result depends on the choice of reference machine. It is one proposed definition among several.

#4.2 Core Cognitive Requirements for AGI

To achieve practical general intelligence, an architecture must satisfy several interdependent cognitive requirements:

1. Causal Reasoning & Counterfactual Modeling: Transitioning beyond observational correlations ($P(y \mid x)$) to structural causal models capable of evaluating interventions ($P(y \mid \text{do}(x))$) and counterfactuals (Pearl, 2009).
2. Sample-Efficient Transfer Learning: Utilizing inductive abstractions acquired in one domain to master an entirely distinct discipline with minimal empirical demonstrations.
3. Persistent Memory & World Modeling: Maintaining an updated, non-volatile internal representation of spatial, temporal, physical, and psychological states over extended operational horizons.
4. Metacognition & Deliberative Search: Evaluating the validity of its own internal reasoning paths, detecting uncertainty, and executing test-time tree search (System 2 deliberation) rather than unconstrained forward intuition (System 1).

#5.1 The Levels of AGI Taxonomy

Morris et al. (2023), a Google DeepMind preprint, propose a matrix classifying systems along two axes: Performance (depth) and Generality (breadth, narrow versus general). The performance levels are Emerging, Competent, Expert, Virtuoso (renamed "Exceptional" in later arXiv versions) and Superhuman, anchored to percentiles of skilled adult humans [VERIFICATION NEEDED: the exact percentile thresholds and the per-cell example systems of the paper's Table 1 could not be read verbatim during correction; consult arXiv:2311.02462 before quoting them]. An earlier version of this article reproduced a table whose narrow-AI examples (placement of Stockfish and AlphaFold, and a system named "AlphaEvo") did not match the source and has been removed. This framework is a proposal and has not been independently validated or broadly adopted.

#5.2 Proposed Architectural Pathways to AGI

• Path 1: Foundation Model Scaling: Scaling multimodal autoregressive transformers with increased compute, parameter count, synthetic dataset curation, and extended context windows.
• Path 2: Test-Time Compute & Deliberative Search: Proposals to augment neural generators with search, self-critique verification, and formal theorem provers (e.g., Lean, Isabelle); the contribution of such methods to generality is unsettled.
• Path 3: Neurosymbolic Systems: Explicitly binding continuous neural embeddings with discrete symbolic knowledge graphs and formal logic solvers.
• Path 4: Embodied World Models: Training agents in physics simulations and learned video world models; LeCun (2022) proposes a world-model-based architecture, which is a position paper and not an empirical demonstration that this path reaches AGI.

Because true AGI does not currently exist, historical and present milestones illustrate progress along the spectrum toward generality: Specific benchmark results are fast-moving and are not asserted here; see the Living Reference collections (MODEL-001, FACT-001). Earlier drafts of this article cited, without sources, 90th-percentile bar-exam and licensing-exam scores, Codeforces ratings and autonomous lab assistants. Exam figures originate largely in developer reports and have been disputed by independent re-analysis, so they are omitted pending sourcing.

• [VERIFICATION NEEDED: cite primary reports (model, date, settings, comparison population) for any exam or Codeforces claim before reinstating it.]
• [VERIFICATION NEEDED: cite sources for any "autonomous lab assistant" claim; reported systems vary widely in autonomy.]

#Potential Transformative Advantages

• Potentially very large scalability of cognitive labor: Proponents argue that parallel deployment of many AI systems could address biomedical, materials, and climate problems; this is speculative.
• Possible reduction in the marginal cost of expertise: Proponents argue for wider access to diagnosis, legal advice, and education; realized benefits would depend on reliability, deployment and policy.

#Critical Bottlenecks & Failure Modes

• Evaluation Saturation & Goodhart's Law: When benchmarks become training targets, high scores reflect dataset memorization or metric gaming rather than genuine cognitive generality.
• Lack of Physical Grounding: Systems trained solely on digital text and static imagery lack physical world models, causing catastrophic failures in intuitive real-world physics and spatial manipulation.
• Instrumental convergence (Hypothetical): Under the instrumental convergence thesis (Omohundro, 2008; Bostrom, 2014), sufficiently capable goal-directed agents would be expected to pursue subgoals such as resource acquisition, self-preservation and goal-content integrity, which could conflict with human interests. This is a theoretical argument, contested in the literature, not an observed property of existing systems.

#Open Research Questions

1. Is language sufficient? Can a model acquire complete causal world understanding purely by predicting next tokens in text and media streams without active physical agency?
2. How can AGI be conclusively tested? What evaluation battery can definitively separate genuine generalization from sophisticated interpolation across a multi-trillion-token training set?
3. Is recursive self-improvement physically constrained? Recursive self-improvement is a hypothetical scenario (see ASI-EXP-001). Will an AGI's ability to redesign its own software lead to an explosive capability runaway or hit diminishing returns in physical hardware and data collection?

Morris et al. (2023), a Google DeepMind preprint, classify frontier language models as Level 1 (Emerging) AGI. This is the authors' own assessment of 2023-era systems (ChatGPT, Bard, Llama 2 per the audit; later versions may add others); no survey of researchers establishing a consensus is cited, and many researchers reject the AGI framing. Such systems display remarkable cross-domain fluency, code synthesis, and multi-step reasoning, yet fail inconsistently on basic planning tasks, suffer from unpredictable hallucinations, and are reported to be vulnerable to distribution shifts (general characterization; not individually sourced).

The field is actively debating whether scaling current autoregressive transformer paradigms will suffice to reach Level 2 (Competent AGI) or whether fundamental breakthroughs in memory persistence, continual learning, and causal world models are strictly required.

Expert survey evidence: in the 2023 AI Impacts survey (n=2,778 researchers who published at top AI venues), the aggregate forecast gave a 10% chance of machines outperforming humans on every task (high-level machine intelligence) by 2027 and a 50% chance by 2047, assuming science continues undisrupted (Grace et al., 2024). These are expert opinions with wide disagreement and sensitivity to question framing, not calibrated predictions. Registered forecasts on AGI timing are kept in AGI-TIME-001. The earlier near-, medium- and long-term bullet list was unattributed and has been replaced by the single record below; the claim that AGI leads by default to ASI is not asserted.

• Foundational Concepts:
  • FOUND-001 Artificial Intelligence: Theoretical definitions of intelligence and rational agency.
  • TECH-ML-001 Machine Learning: Generalization bounds and empirical risk minimization.
• Related Methods & Technologies:
  • TECH-ARCH-001 The Transformer Architecture: The primary architectural substrate for frontier models.
  • CAP-GENAI-001 Generative Artificial Intelligence: Large-scale generative foundation models.
• Advanced Implications & Future Trajectories:
  • ASI-001 Artificial Superintelligence: Capability trajectories beyond human-level cognitive performance.
  • SAFE-ALIGN-001 The AI Alignment Problem: Technical verification and control of advanced autonomous agents.

This article surveys the competing definitions of artificial general intelligence (AGI) and the operational criteria that have been proposed to decide when a system satisfies one of them. There is no single accepted definition: published definitions differ in whether they specify an economic outcome, a distribution over tasks, a human reference class, or a formal mathematical measure, and these choices change which systems qualify and when forecasters expect the threshold to be met.

The term AGI is used in research papers, corporate mission statements, contracts, safety policies and public forecasting. AGI-001 introduces the concept; this article is concerned with a narrower question, namely what it would mean to say that a given system is an AGI, and how that claim could be checked.

The question has practical weight. A forecast of "AGI by year X" is a forecast about a particular definition, so two forecasts that use different definitions are not comparable. Survey results illustrate the dependence: the 2023 survey of Grace et al. (2025, arXiv version 2024) reported that the aggregate forecast gave a 10% chance by 2027 and a 50% chance by 2047 of "unaided machines outperforming humans in every possible task", whereas the aggregate forecast for "all human occupations becoming fully automatable" was 10% by 2037 and 50% by 2116 (survey conducted in 2023; 2,778 respondents). The two questions describe different thresholds, and the figures are 69 years apart at the 50% level. Definitions also appear in governance: safety frameworks and corporate agreements that refer to AGI inherit whatever ambiguity the underlying definition carries.

The phrase "general intelligence" predates computing. Goertzel and Pennachin (2007) collected research programmes under the label AGI to distinguish work on broad, flexible intelligence from the specialised systems that dominated applied AI. Legg and Hutter (2007a) collected informal definitions of intelligence from psychology and AI and distilled them into a single informal statement; in a companion paper (Legg & Hutter, 2007b) they gave a formal definition, discussed in Section 4.

Informal operational proposals preceded both. Turing (1950) proposed replacing the question whether machines can think with an imitation game, a behavioural criterion that is still discussed as an early operational test (see AGI-TEST-001). Later definitions reflect different priorities: formal definitions (Legg & Hutter, 2007b), skill-acquisition definitions (Chollet, 2019), a taxonomy-based definition (Morris et al., 2024) and a cognitive-profile definition (Hendrycks et al., 2025). Claims that early systems showed signs of general intelligence (Bubeck et al., 2023) prompted renewed debate about what evidence such claims require.

#4.1 Families of definition

The definitions in current use fall into five groups. The groups overlap, and a given author may combine several.

1. Economic or outcome definitions. These define AGI by what it can do in the economy. The OpenAI Charter describes AGI as "highly autonomous systems that outperform humans at most economically valuable work" (OpenAI, n.d.). The definition is operational in principle, since it can be checked against labour-market tasks, but it depends on how "most", "outperform" and "economically valuable" are measured, and on the cost at which the work is done.
2. Formal, environment-distribution definitions. Legg and Hutter (2007b) define the universal intelligence of an agent $\pi$ as its expected performance over all computable environments, weighted by simplicity:

where $E$ is the set of computable reward-bearing environments, $K(\mu)$ is the Kolmogorov complexity of environment $\mu$, and $V_\mu^\pi$ is the expected total reward of $\pi$ in $\mu$. This is a mathematical definition and is Established as such. It is not directly usable as a test: $K$ is not computable, and the result depends on the choice of reference machine, so it does not by itself identify a threshold for "general". 3. Skill-acquisition definitions. Chollet (2019) argues that intelligence should be defined as skill-acquisition efficiency over a scope of tasks, relative to priors and experience, and that measuring skill alone conflates intelligence with prior knowledge or training data. On this view a benchmark score is evidence of generality only when the system could not have been prepared for the tasks. 4. Tiered capability taxonomies. Morris et al. (2024) propose a matrix with two axes, performance and generality, instead of a single threshold. Their performance levels are Emerging (equal to or somewhat better than an unskilled human), Competent (at least the 50th percentile of skilled adults), Expert (90th), Exceptional (99th; called "Virtuoso" in earlier versions of the preprint) and Superhuman (outperforms 100% of humans). The paper states six principles for a definition: focus on capabilities rather than processes; on generality and performance; on cognitive and metacognitive rather than physical tasks; on potential rather than deployment; on ecological validity; and on the path to AGI rather than a single endpoint. As of the paper's assessment, dated September 2023 in the paper, the authors classified then-current frontier language models as Emerging general AI. That classification is the authors' own judgement, not a measurement protocol applied by an independent party. 5. Cognitive-profile definitions. Hendrycks et al. (2025) define AGI as "matching the cognitive versatility and proficiency of a well-educated adult", split general intelligence into ten core cognitive domains, and report domain-level scores for named models (the abstract gives GPT-4 at 27% and GPT-5 at 57%). The scores come from the authors' own scoring scheme and are Preliminary: they await independent replication, and the framework's weights and test selection are choices that other authors might make differently.

#4.2 What an operational criterion must fix

A definition becomes operational only when it specifies the following. Where a published definition leaves an item open, claims made under it cannot be falsified.

• Reference class: unskilled humans, skilled adults, domain experts, or the best human at each task.
• Task distribution: finite fixed battery, open-ended sampling, or the economy's tasks. Fixed batteries are vulnerable to contamination and to optimisation against the benchmark (see CAP-BENCH-001).
• Aggregation: whether superhuman performance on some tasks compensates for failure on others. Morris et al. (2024) require breadth and depth jointly; the economic definition aggregates by value.
• Cost and autonomy: whether the system must act without human help and at what cost, which matters for the "autonomous" element in the Charter wording.
• Access conditions: tools, retrieval, scaffolding, and time allowed per task.
• Prior exposure: whether tasks may appear in training data. This is the core of the Chollet (2019) critique.

#4.3 Relationship between definitions

A system can satisfy one definition and fail another. A system that outperforms most humans on a fixed academic battery may not perform most paid work autonomously, and a system able to do paid digital work may not satisfy a physical-task criterion. Because definitions differ in strength, the term "AGI" without a stated definition should be read as underspecified.

The table classifies representative definitions by their operational character. It records the form of each definition, not a judgement of its merit.

Two further distinctions recur. Strong versus weak AI, in the philosophical sense, concerns whether a system has a mind; it is distinct from the capability-based senses above and is treated in FOUND-005. Human-level versus human-like contrasts matching performance with matching mechanism, a distinction Morris et al. (2024) make explicit by focusing on capabilities rather than processes.

Definitions are used in several places, and each use applies its own criterion.

• Forecasting surveys. Grace et al. (2025) asked about two differently worded thresholds, as described in Section 2; the difference between the two answers is a direct measure of how much the wording matters. See AGI-TIME-001.
• Evaluation design. ARC-AGI-2 (Chollet et al., 2025) was released as a successor to a benchmark introduced in 2019 because the authors judged that recent progress required "finer-grained evaluation at higher levels of cognitive complexity." This illustrates a definition (skill acquisition on novel tasks) driving a concrete instrument. The details of such instruments are covered in AGI-TEST-001.
• Organisational policy. Frontier developers refer to AGI in mission statements and in capability thresholds for safety frameworks. The wording of such documents changes over time and is outside the scope of a stable article; current versions should be read directly.
• Case studies of claimed generality. Bubeck et al. (2023) described early experiments with GPT-4 as showing "sparks" of AGI. Whether it counts as evidence of generality depends on the criterion adopted, which makes it an example of a claim whose status changes with the definition.

Advantages of operational definitions. They make claims testable, allow forecasts to be compared and resolved, and separate disagreement about the facts from disagreement about words.

Limitations.

• Outcome definitions embed economic and institutional factors (regulation, cost, deployment) in what is nominally a capability question. Morris et al. (2024) explicitly avoid this by focusing on potential rather than deployment.
• Formal definitions, such as Legg and Hutter's, are precise but not computable, and the weighting by simplicity is itself a contested assumption.
• Human-relative thresholds depend on the chosen human population. "Skilled adult" is not a fixed standard across tasks.
• Fixed benchmarks can be saturated or contaminated without the system having acquired general skill, which is why Chollet (2019) separates skill from skill acquisition.
• Single-number scoring schemes, such as the percentage scores in Hendrycks et al. (2025), can hide the structure of failure: an agent with a critical deficit in one domain, such as long-term memory storage (which the authors identify), may score well on average.

Open questions.

1. Whether any definition can be both formally precise and practically measurable.
2. Whether "general" should be defined over tasks humans do, or over a wider space of tasks. The first is a statement about human-relevant capability; the second is closer to Legg and Hutter.
3. Whether physical competence belongs in the definition. Morris et al. (2024) exclude it from the core criterion; others include embodied tasks.
4. Whether a threshold concept is appropriate at all, or whether a profile of capabilities is more informative. This is Contested.

As of 2026-10-15 there is no consensus definition of AGI. No source reviewed for this article shows agreement among researchers on a single operational criterion. The most visible definitions are of the five kinds in Section 4.1, and they are used in parallel. The claim that any existing system meets a given definition is itself definition-dependent: under the Morris et al. (2024) matrix the authors assessed frontier language models as Emerging general AI in 2023; under the Hendrycks et al. (2025) scheme a partial score is reported; under the Charter's wording the answer depends on how economic value is aggregated. Evaluations of this kind are Preliminary. For current model assessments, see the Living Reference collection (MODEL-001, FACT-001, TIME-001) rather than this article.

This section is descriptive; it does not register a forecast.

• Near term (0 to 3 years). Definitions are likely to continue to proliferate and to be supplemented by domain-specific capability thresholds (for example, autonomy in software tasks) that avoid the word AGI. The METR time-horizon work (Kwa et al., 2025) is an example of a measurement that sidesteps the definitional question by reporting a single, task-length-based metric.
• Medium term (3 to 10 years). Convergence on one definition is not assured. A plausible alternative is convergence on a small set of measurable profiles, but no source reviewed shows evidence for this, and the point is speculative.
• Long term (10+ years). If systems exceed human performance on most measured tasks, the weakest definitions will be unambiguously met and discussion will shift towards ASI definitions (ASI-DEF-001).

No dated forecasts are made here. Registered forecasts are collected in the Forecast Register (FCST-REG-001).

• Foundational Concepts:
  • AGI-001 Artificial General Intelligence: The overview article; this entry covers definitional detail.
  • FOUND-002 Definitions of Intelligence: Definitions of intelligence in general, on which AGI definitions build.
  • FOUND-005 Philosophy of Mind & The Turing Test: The imitation game and the strong/weak AI distinction.
• Related Methods & Technologies:
  • AGI-REQ-001 Requirements for General Intelligence: Which capabilities a definition may require.
  • AGI-TEST-001 Testing Claims of General Intelligence: How operational criteria are tested.
  • CAP-BENCH-001 AI Evaluation, Benchmarking & Contamination: Benchmark validity.
• Advanced Implications & Future Trajectories:
  • AGI-TIME-001 AGI Forecasting & Timeline Methodology: Why definitions determine timelines.
  • AGI-PATH-001 Pathways to AGI: Routes to systems meeting such definitions.
  • ASI-DEF-001 Defining and Measuring Superintelligence: The corresponding problem above the human level.