Levels of AGI for Operationalizing Progress on the Path to AGI

• AGI describes AI systems at least as capable as humans at most tasks and has gained near-term practical relevance as machine-learning models advance.Source: §1 · p. 1
• AGI functions as an aspirational goal for AI research, reflected in historical field-wide ambitions and current research organizations’ mission statements.Source: §1 · p. 1
• AGI-related predictions concern increasing generality, emergent capabilities, labor substitution, economic effects, and geopolitical consequences.Source: §1 · p. 1
• Some researchers associate AGI with extreme risks including deception, manipulation, resource accumulation, goal-directed behavior, human displacement, and recursive self-improvement.Source: §1 · p. 1
• The paper calls for operational measures of performance, generality, and autonomy to support model comparisons, risk mitigation, policymaking, and communication about progress toward AGI.Source: §1 · pp. 1–2

• The authors analyze nine prominent AGI definitions and use their strengths and limitations to motivate a two-dimensional, leveled ontology.Source: §2 · p. 2

Case Studies 1–2

• The Turing Test and consciousness-based strong AI frame AGI through imitation or internal processes, but the paper argues capabilities are more practical to measure.Source: §2 · p. 2

Case Study 3

• Brain-based definitions emphasize human-like processes, whereas transformer performance suggests human-like learning mechanisms are not inherently necessary for AGI.Source: §2 · p. 2

Case Study 4

• Human-level cognitive-task definitions avoid requiring robotic embodiment but leave ambiguity about which tasks and which people determine AGI status.Source: §2 · p. 2

Case Studies 5 and 7

• Learning-based definitions treat metacognition as necessary for generality, while Marcus’s flexible-and-general formulation links reliability, resourcefulness, and learning new skills.Source: §2 · pp. 2–3

Case Study 6

• OpenAI’s economically valuable work definition offers a performance yardstick but may omit non-economic intelligence and conflate capability with deployment.Source: §2 · p. 2

Case Study 7

• The Coffee Test and related task suites add concrete tests of flexibility but may be incomplete, insufficient on their own, or dependent on robotic embodiment.Source: §2 · p. 3

Case Study 8

• ACI emphasizes complex, multi-step, real-world tasks with ecological validity, but its profit-focused Modern Turing Test is narrower and may introduce alignment risks.Source: §2 · p. 3

• The authors derive six criteria for clear, operationalizable AGI definitions from their review of existing formulations.Source: §3 · p. 3
• AGI definitions should focus on capabilities rather than processes, so human-like thought, consciousness, or sentience are not necessary conditions.Source: §3 · p. 3
• AGI should combine generality with performance because broad task coverage without sufficient reliability does not establish capable general intelligence.Source: §3 · p. 3

Focus on Cognitive and Metacognitive, but not Physical, Tasks

• Physical abilities may increase generality, but robotic embodiment should not be required for AGI because many definitions focus on cognitive, non-physical tasks.Source: §3 · p. 3
• Metacognitive abilities such as learning new tasks and knowing when to request human help are treated as prerequisites for generality.Source: §3 · p. 4

Focus on Potential, not Deployment

• AGI should be defined by demonstrated task capability or potential rather than open-world deployment, which introduces legal, social, ethical, and safety hurdles.Source: §3 · p. 4

Focus on Ecological Validity

• AGI benchmarks should use ecologically valid tasks aligned with broadly valued real-world abilities rather than relying only on easy-to-automate traditional metrics.Source: §3 · p. 4

6. Focus on the Path to AGI, not a Single Endpoint.

• Levels of AGI support comparisons among competing formulations while distinguishing Emerging, Competent, and Exceptional AGI thresholds.For example, Aguera y Arcas and Norvig’s definition maps to Emerging AGI, OpenAI’s labor-replacement threshold to Exceptional AGI, and several other formulations to Competent AGI.Source: §6 · p. 4

4. Levels of AGI

• Performance measures capability depth relative to relevant human performance, while generality measures the breadth of tasks reaching a target threshold.Performance above Emerging is compared with adults who possess the relevant skill; generality concerns the range of tasks meeting the threshold.Source: §4 · p. 4
• The matrix requires minimum performance over most tasks for a level, allowing higher performance on subsets; current frontier language models therefore remain Emerging AGI overall.They show Competent performance on some tasks such as short essay writing and simple coding, but Emerging performance on most tasks, including mathematics and factuality.Source: §4 · p. 4
• Progression may be nonlinear, and the order of acquiring capabilities can create safety concerns, especially when powerful technical skills precede ethical reasoning.The framework also notes that learning new skills may accelerate movement toward the next level.Source: §4 · p. 4
• Table 1 presents approximate assignments in a performance-generality matrix, with Competent AGI unachieved by public systems at the time of writing.General systems may perform a narrow subset of tasks at higher levels than their broad overall level.Source: Table 1 · p. 5
• Deployed performance may fall below benchmark-based ratings because interfaces and failure modes can prevent users from eliciting a system’s theoretical capability.The DALLE-2 example illustrates how complex prompting and output errors can lower practical performance from Expert to Competent.Source: §4 · p. 5
• ASI is the highest combined level, describing systems that perform across a wide range of tasks beyond the ability of every human.The framework defines Superhuman performance as outperforming 100% of humans and gives AlphaFold as an example of Superhuman Narrow AI.Source: §4 · pp. 5–6

5. Testing for AGI

• The ontology’s measurement principles emphasize capabilities rather than processes, cognitive and metacognitive tasks, potential rather than deployment, and ecological validity.These principles complement the performance-generality dimensions used to classify systems.Source: §5 · p. 6
• The paper leaves benchmark construction open because generality criteria, task coverage, and required task proportions remain unresolved.It clarifies the ontology and benchmark properties rather than proposing a completed benchmark.Source: §5 · pp. 6–7
• A future benchmark should span linguistic, mathematical, logical, spatial, social, learning, and creative abilities, while evaluating each task for ecological and construct validity.Psychometric categories may inform task selection, but their suitability for computing systems must first be established.Source: §5 · p. 6
• Metacognitive testing should include learning new skills, knowing when to request help, calibration, and modeling users through theory-of-mind abilities.These abilities are linked in the paper to generality, alignment, appropriate interaction, and awareness of model limits.Source: §5 · p. 6
• Benchmark designers must decide whether tools are allowed by considering the ecologically valid counterfactual for each task.For driving safety, comparison with drivers using modern assistance may be more informative than comparison with unaided drivers.Source: §5 · p. 6
• Open-ended and interactive tasks may better approximate real-world capability than traditional automated metrics, despite requiring qualitative evaluation.The paper also proposes a living benchmark because a sufficiently general intelligence’s full task set cannot be enumerated.Source: §5 · p. 6

6.1. Levels of AGI as a Framework for Risk Assessment

• Expert AGI may introduce structural risks such as economic disruption and job displacement, while Exceptional AGI and ASI raise stronger concerns about deception and misalignment.Lower levels more often involve misuse and incorrect-execution risks, while higher levels can alleviate some earlier risks even as they introduce new ones.Source: §6.1 · p. 7
• Below Expert AGI, risks likely stem more from human actions, while rapid progression beyond regulation or diplomacy could create systemic geopolitical risks.The authors identify fuller level-specific risk profiles as important for safety, ethics, and policymaking.Source: §6.1 · p. 7

6.2. Capabilities vs. Autonomy

• Autonomy is correlated with AGI capability but is not determined by it, so lower autonomy may remain preferable for particular tasks and contexts.The proposed six autonomy levels characterize human-AI interaction styles rather than merely the amount of control relinquished to computers.Source: §6.2 · p. 8
• The paper emphasizes that no-autonomy deployment can remain valuable for education, enjoyment, assessment, or safety even when more capable automation is available.It gives driving as an analogy and extends the same contextual choice to AGI systems.Source: §6.2 · pp. 8–9

6.3. Human-AI Interaction and Risk Assessment

• More capable systems unlock interaction paradigms such as full autonomy, but each paradigm introduces new risks that depend on deployment context.The framework therefore combines AGI level, autonomy level, interface, task, scenario, and end-user rather than assessing capabilities alone.Source: §6.2 · pp. 7–8, §6.3 · p. 9
• General systems may unlock higher autonomy for selected tasks despite lower overall levels because their capabilities can be uneven across tasks.Consultant-style interaction is predicted to be possible with either an Expert Narrow AI or an Emerging AGI.Source: §6.3 · p. 9

7. Conclusion

• The authors distill six principles for operationalizing AGI, including capability-focused definitions, performance and generality, cognitive and metacognitive tasks, potential rather than deployment, ecological validity, and progress toward an endpoint.These principles follow an analysis of nine AGI definitions.Source: §7 · p. 9
• The Levels of AGI ontology combines Narrow or General breadth with five performance levels from Emerging to Superhuman.The paper also discusses how current systems and existing AGI definitions fit this framing.Source: §7 · p. 9
• A living, ecologically valid AGI benchmark is presented as difficult but vital for measuring progress against the ontology.The benchmark is intended to reflect the principles developed in the paper.Source: §7 · p. 9
• AGI and autonomy are not synonymous: autonomy levels are unlocked but not determined by AGI progression, and joint analysis can support more nuanced risk assessment.The conclusion underscores investment in human-AI interaction research alongside model improvements.Source: §7 · p. 9
• The framework is intended to remain adaptable as technical advances and societal changes alter how AGI progress, interaction paradigms, and risks are defined.Interpretability improvements and social adaptation are given as examples of factors that may change the ontology.Source: §7 · pp. 9–10

• The ontology is designed to support clear communication about AI systems’ capabilities and associated risks across research, practice, and policymaking.Source: Impact Statement · p. 10