Levels of AGI Ontology

• Levels of AGI ontology is a multi-dimensional framework that defines intelligent systems by categorizing them along axes such as performance, generality, and philosophical depth.
• It employs measurable benchmarks and structured criteria—from functional capabilities to quantum versus classical substrates—to gauge developmental progress.
• The framework integrates technical, safety, and geopolitical considerations to guide risk assessment, deployment strategies, and alignment in AGI research.

AGI ontologies provide formal, multi-level conceptual frameworks that organize the structure, progression, capabilities, contexts, and theoretical boundaries of intelligent systems. These ontologies serve as reference points for research, risk analysis, evaluation, deployment, and philosophical interpretation. Across the literature, ontological levels range from fine-grained functional hierarchies (performance, breadth, autonomy) to deep philosophical criteria (generativity, coordination, sustaining identity), technical substrata (classical vs quantum), embodiment, and geopolitical partitioning.

1. Core Functional and Capability-Based Ontologies

A dominant strand defines AGI levels according to observable capabilities across multiple axes. The “Levels of AGI” framework partitions capabilities along two principal dimensions: depth (performance relative to skilled human percentiles) and breadth (generality across diverse benchmarks), producing a six-row (emerging to superhuman) by two-column (narrow to general) matrix (Morris et al., 2023). Each cell specifies performance percentiles and task coverage:

Level	Performance (Depth)	Breadth of Generality	Typical Examples
0	No AI	None	Calculators, paint programs
1	Emerging	Narrow / General	Rule-based systems / ChatGPT, Llama 2
2	Competent	Narrow / General	Smart speakers / AGI (not yet achieved)
3	Expert	Narrow / General	DALL·E 2 / AGI (not yet achieved)
4	Virtuoso	Narrow / General	AlphaGo / AGI (not yet achieved)
5	Superhuman	Narrow / General	AlphaFold / Artificial Superintelligence

Autonomy is layered on top, from tool-like operation (human-in-the-loop) to agentic behavior (fully autonomous AGI), with risks compounding at higher autonomy levels. Benchmarks are envisioned as modular, extensible, domain-spanning, emphasizing both cognitive and metacognitive tasks. The framework explicitly sidesteps subjective, process-centric, or metaphysical attributes, focusing instead on measurable, ecologically valid criteria (Morris et al., 2023).

A closely related ontology structures AGI progression into three discrete levels—Embryonic, Superhuman, and Ultimate—mapped onto four capability axes: internal cognition, interface, system, and alignment (Feng et al., 2024). Each level imposes stricter thresholds, for instance, requiring creativity and self-reflection only at Level 3 (Ultimate AGI). The framework associates each level with a multi-benchmark suite and specifies that advancement requires not just raw scale but the integration of alignment, multi-agent collaboration, metacognition, and fully automated self-improvement.

2. Structural-Generative and Philosophical Ontologies

Beyond external behavior, the “Structural-Generative Ontology of Intelligence” delineates three nested depth conditions as prerequisites for true intelligence (Wang et al., 2 Sep 2025):

1. Generativity: The system actively constructs novel categories, distinctions, or rules not implicit in prior knowledge, and provides explanatory advancement. Standard next-token prediction or pattern recombination fails this test; only categorical innovation counts.
2. Coordination: The system reconciles conflicts among generated structures by producing normatively integrated, reasoned wholes. It must deliver coherent justifications that resolve tension between rules or sub-structures under adversarial interrogation.
3. Sustaining Identity: The system maintains coherent narrative continuity through time, explaining belief revisions and remaining accountable to its epistemic history—formalized as mapping time-stamped states through a coordinated, generative narrative function.

These depth conditions spiral: generativity provokes conflict (necessitating coordination), coordinated commitments require narrative sustaining, and identity over time triggers new generativity. This ontology sharply distinguishes genuine intelligence from simulacra, such as infinite retrievers (“oracle of the library”) or rote memorizers (“the memorizing scholar”), and claims that only satisfaction of all three depth conditions grants AGI the status of a “Second Being” (Wang et al., 2 Sep 2025).

3. Substrate and Physical Ontologies: Quantum vs Classical

A parallel thread formalizes AGI ontologies in terms of computational substrate and allowed channels. The “Quantum AGI Ontology” introduces three principal ontological levels (Perrier et al., 16 Jun 2025):

• Level I (Classical AGI): Classical software on classical hardware, with state spaces over finite strings, Shannon entropy, and non-contextual, definite states.
• Level II (Hybrid): Quantum-simulating software on classical or quantum hardware; quantum properties are only emulated or instrumental and do not constitute a quantum ontology.
• Level III (Quantum-native AGI): Native quantum software on quantum hardware with registers as density operators, Hilbert space of dimension ≥2, genuine superposition, entanglement, contextuality (Kochen–Specker), non-locality (Bell’s theorem violation), and the no-cloning constraint. Identity becomes resource-theoretic rather than persistent, learning involves channel tomography and coherent unitary updates, and agent-environment interaction fundamentally departs from classical locality and realism.

Only at Level III does the ontology force a redefinition of agency, self-inspection, and learning, due to measurement-induced collapse, contextual value assignments, and the impossibility of universal state cloning (Perrier et al., 16 Jun 2025).

4. Safety, Trust, and Alignment-Focused Ontologies

Recent taxonomies center on stratifying AGI according to trustworthiness and safety properties. The “Causal Ladder of Trustworthy AGI” hierarchy defines five progressive levels (Yang et al., 2024):

1. Perception Trustworthiness: Accuracy, robustness, and calibration at the sensor or input layer.
2. Reasoning Trustworthiness: Traceable, verifiable, and robust intermediate inferences, with error diagnosis.
3. Decision-making Trustworthiness: Value-aligned actions, human-in-the-loop intervenability, and justification traces.
4. Autonomy Trustworthiness: Self-regulating, self-reflective policy adjustment and long-term constraint adherence.
5. Collaboration Trustworthiness: Multi-agent, threat-model-verified consensus building, conflict resolution, and trust recalibration in joint environments.

This stratification mirrors Judea Pearl’s Causal Ladder and tightly couples safety criteria to increasing levels of counterfactual and reflective reasoning, from raw data association to deep, multi-agent counterfactual scenario planning (Yang et al., 2024).

5. Embodiment and Task Integration

The embodied AI literature advances taxonomies that capture the physical, multimodal, and closed-loop attributes required for embodied AGI (Wang et al., 20 May 2025). The five-level taxonomy (L1–L5) distinguishes agents by their:

• Modalities: from partial (single sensory domain) to full (vision, audition, haptics, text, etc.)
• Real-time responsiveness: from offline or batch to fully real-time closed-loop control
• Generalization: from single-task to robust open-task execution
• Humanoid cognition: from none (reactive reflexes) to full (theory of mind, goal tracking)
• Body requirements: from robust task-specific platforms to fail-safe, all-purpose humanoid bodies

Transitions between levels require advances in modular skill networks, omnimodal foundation models, continual learning, world model internalization, social cognition, memory reconsolidation, and integrated safety mechanisms (Wang et al., 20 May 2025).

6. Geopolitical and Sociotechnical Ontologies

AGI ontologies extend beyond purely technical dimensions to address the social, political, and institutional lineage of systems. The “High vs Low AGI” framework formalizes a bipartition based on the system’s locus within geopolitical architectures (Max, 6 Oct 2025):

• Low-AGI: Architecturally probabilistic, transparent, market-driven, multi-stakeholder governed, and framed concretely for civilian/commercial use (e.g., foundation models).
• High-AGI: Closed, composable, adversarially robust, secrecy-governed, concentrated in sovereign or security domains, and framed abstractly for existential or regime survival contexts.

A formal partition is imposed via profile functions over actor structure, psychological distance, discursive framing, and spillover dynamics. The framework enables layered regulation, risk vector mapping, and anticipatory governance depending on the system’s ontological register within the sociotechnical ecosystem (Max, 6 Oct 2025).

7. Existence, Perception, and Meaning—Foundational Ontologies

Certain frameworks return to the epistemic ground of AGI, positing that the minimal ontology for meaning-capable intelligence requires two recursively-defined layers: Existence (set-theoretic membership, including three-valued/recursively nested predicates $E(x,S)$ for ambiguous existence) and Perception (relations $R(x,y)$, realized as concrete datum via $P_o(R,x)$). Higher constructs—meaning and knowledge—arise from layering relations atop existence and storing verified events in a knowledge base. This structure supports formal abduction, induction, and deduction in a logic engine, with safety features provided by explicit “unknown” values (Senkevich, 2022).

8. Containment, Security, and Cyber-Physical Ontologies

Five-level containment ontologies decompose the universe into object-agents (human, AGI, cyberworld), classes (individual, society, swarm), physical attributes (composition, hardware/software, locality), abstract attributes (security, intelligence, autonomy), and relationships (policy, attack, defense) (Pittman et al., 2018). Security procedures (e.g., attack/defend events), policies, and system state transitions are structured over these layered attributes, enabling logic-based containment strategies and risk assessment.

---

Synthesizing across these ontologies, the multidimensional structure of AGI levels spans functional capabilities, depth/generality benchmarks, physical substrate, philosophical depth conditions, safety and trust layers, embodiment, sociopolitical and institutional context, foundational existence/perceptual layers, and security interactions. The selection of an ontology is determined by discipline, evaluative requirement, and the theoretical commitments of the researcher or policymaker.