Have We Achieved Artificial General Intelligence?

Testing AI against the agreed scientific definition of general intelligence.

Define the term, then test it

Bertrand Russell and the analytic tradition taught that many apparently deep disputes are really disputes about undefined terms, and that the cure is to fix the definition first and then let evidence do the work [1]. “Artificial general intelligence” has become exactly the kind of term Russell warned about: undefined in most conversations, and therefore available for endless hype and endless dismissal alike. This essay does one narrow thing. It takes the definition of general intelligence that the field of intelligence research has itself agreed, and tests today’s AI against it, characteristic by characteristic.

That agreed definition exists. The 1997 statement Mainstream Science on Intelligence, signed by 52 intelligence researchers, defines intelligence as a very general mental capability that involves seven abilities: reasoning, planning, solving problems, thinking abstractly, comprehending complex ideas, learning quickly, and learning from experience [2].

Just as important is what this definition deliberately excludes. It does not require consciousness; the hard problem of why there is subjective experience at all is a separate and far harder question [3], and the leading AGI framework likewise treats consciousness as unnecessary to the definition [4]. It does not list creativity among the seven abilities; psychology treats creativity as a related but distinct construct. It does not list wisdom, which is arguably the application of intelligence at its broadest. And it takes no position on where intelligence comes from, whether inherited through evolution, transmitted socially, emergent from simpler components, or received from some larger field of mind. All of these are worthwhile conversations. None of them is this one. This paper asks only whether the agreed definition has been met, and nothing here requires infallibility: humans hold the title of general intelligence while performing all seven abilities imperfectly, so AI cannot fairly be held to a higher bar.

What we mean by “AI”

Throughout, “AI” does not mean a bare language model or a single frozen set of weights. It means the complete engineered system the field of artificial intelligence actually builds and deploys: the trained model together with its tools, external memory, sensors, the agent harness that lets it act over time, the code it writes and executes, and the feedback and retraining processes that carry learning forward between versions. This mirrors how we assess human intelligence: not a brain in a jar, but an educated person using notebooks, calculators and colleagues. So if a model reasons its way to the best possible deterministic function for a task, writes it, and runs it, the system has reasoned; the division of labour between neural and symbolic parts is an implementation detail. A prominent AGI taxonomy takes the same view, judging systems by demonstrated capability rather than internal mechanism [4].

The test: seven characteristics

For each characteristic: what it means, everyday human examples we already accept as evidence of intelligence, a simple AI demonstration with a published, cited source, a complex and formally tested AI example, and a verdict. The logic of the test is two-tiered. The simple demonstration asks whether the characteristic is present at the level we already accept in ordinary humans; if it is, the characteristic is met, because the definition sets a human bar, not a superhuman one. The complex example asks a different question, degree: whether performance reaches or exceeds expert human level. Where a complex example is not yet fully conquered, that bears on the separate question of superintelligence, a different definition, and does not subtract from whether the characteristic is present.

1. Reasoning

Reasoning means using evidence, relationships or rules to reach a conclusion. A person noticing dark clouds and taking an umbrella is reasoning; so is a doctor weighing symptoms, test results and medications before settling on the most likely diagnosis.

Simple demonstration, cited: everyday word problems of the kind set for schoolchildren, such as working out how many apples remain after some are used and more are bought. Wei and colleagues showed that large language models solve these multi-step problems reliably when they reason step by step, and performance on the standard grade-school benchmark has since climbed to near-perfect [5]. If a child producing these answers is reasoning, so is the system.

Complex, tested example: AlphaGeometry2 combines a learned model with a symbolic deduction engine to construct proofs for unfamiliar olympiad geometry problems, and has exceeded the average performance of International Mathematical Olympiad gold medallists [6]. Because a proof either holds or does not, the result is independent of anyone’s opinion about what counts as reasoning.

Verdict: the characteristic is met, on the strength of the simple demonstration alone, and the complex example shows performance in formally checkable domains beyond elite human level. The objection that the internal process differs from human thought imports a requirement the definition never contained.

2. Planning

Planning means choosing and organising a sequence of actions towards a goal while handling constraints and obstacles. The everyday human case is planning a holiday: several legs across unfamiliar geography, flights and accommodation that must fit a budget and a calendar, and the whole thing adapted on the fly when a train is cancelled or a hotel is full. It mixes hard constraints that must simply be satisfied with open judgements about what is worth the time, which is what makes it a genuine test of planning rather than mere scheduling.

Simple demonstration, cited: route planning. Every satellite navigation system that finds the best path to a destination and replans around a closure runs on planning methods the field of AI formalised decades ago, beginning with the A* algorithm of 1968 [7]. Billions of such plans are produced and acted on daily; a person doing the same with a paper map would unquestionably be said to be planning.

Complex, tested example: the holiday itself. TravelPlanner is a benchmark of 1,225 realistic travel-planning tasks with budgets, timetables and commonsense constraints; on its introduction, the best bare language model achieved a final pass rate of 0.6 per cent [8]. Within the year, a system in which the language model translates the request into a formal constraint problem and hands it to a deterministic satisfiability solver reached 93.9 per cent [9]. This is precisely the engineered-system point made above: the model formulates the deterministic machinery and runs it. Long-horizon agentic planning shows the same trajectory, with planner-executor architectures performing strongly on multi-stage web tasks [10] and the measured length of tasks agents complete autonomously doubling roughly every four months [11].

Verdict: the characteristic is met. Everyday planning has been demonstrated at scale for decades, and the hardest published everyday case, holiday planning under real constraints, moved from near-total failure to above 90 per cent within a year once the system was engineered properly. Long-horizon reliability remains a matter of degree, not of whether planning is present.

3. Problem-solving

Problem-solving means finding a workable way past an obstacle when the answer is not immediately available. A person using a coin as a screwdriver is problem-solving; so is an engineer tracing an intermittent fault across a database, an API, an application and a network.

Simple demonstration, cited: short, self-contained programming problems, such as writing a function that checks whether a word is a palindrome. The HumanEval benchmark verifies each attempt automatically with unit tests; the first system evaluated on it in 2021 solved a minority of the problems, and current models solve nearly all of them [12]. The obstacle is real, the solution is not given, and the outcome is pass or fail.

Complex, tested example: SWE-agent receives a genuine reported software issue, navigates the repository, edits code, runs the project’s own tests and revises according to the results [13]. The benchmark built on such tasks, SWE-bench, went from single-digit scores to saturation in roughly two years [11]. These are held-out, real-world problems with verifiable outcomes, which is the evidence standard that answers the objection that the model merely memorised its training data.

Verdict: the characteristic is met, and in the software domain the complex example is met as well, at a level of performance most humans could not match.

4. Abstract thinking

Abstract thinking means recognising an underlying rule that applies beyond one example. A person seeing that a recipe ratio and a map scale are the same kind of relationship is abstracting; so is a scientist recognising that natural selection, market competition and optimisation algorithms share a single process of variation, selection and retention.

Simple demonstration, cited: analogy problems of the kind used in human intelligence testing. In a peer-reviewed study in Nature Human Behaviour, GPT-3, with no training on the tasks, matched or exceeded university students on most text-based analogy and matrix-reasoning problems modelled on Raven’s Progressive Matrices [14]. These are the very instruments psychology uses to measure abstract reasoning in people.

Complex, tested example: ARC-AGI-2 presents novel visual grids with only a few demonstrations; the system must infer the hidden transformation and apply it to a new case. The benchmark was built specifically to resist memorisation. The 2025 ARC Prize documented systems solving a meaningful proportion of these tasks through program synthesis and test-time refinement, but human-level accuracy at human-level efficiency has not been reached [15].

Verdict: the characteristic is met: performance on the standard psychological instruments for abstract reasoning is at or above the level of the university students against whom it was measured. The complex example is where the degree question remains open, and it is the strongest counter-evidence in this essay; it counts against superhuman-efficient abstraction, a different bar, not against the presence of the characteristic.

5. Comprehending complex ideas

Comprehending complex ideas means building a coherent representation of something with many interacting concepts and causes. A person understanding how interest compounds on a loan is comprehending; so is a lawyer understanding how legislation, precedent, evidence and procedure interact in a case.

Simple demonstration, cited: examination questions. The MMLU benchmark draws real exam questions from 57 subjects, from school level through to professional law and medicine, and current frontier systems score at or above the estimated human expert baseline [16]. Passing an examination is exactly the evidence we accept from a person that a complex idea has been comprehended and can be applied.

Complex, tested example: Microsoft’s GPT-4 study found the model could combine and apply concepts across mathematics, programming, medicine, law and psychology, domains that previously required separate specialist systems [17]. AlphaFold integrates complex biological relationships to infer protein structures, verified against experimentally determined structures [18].

Verdict: the characteristic is met on the functional reading, with breadth well beyond a typical human. If comprehension is instead defined to require private subjective awareness, the question has silently become the consciousness question [3], which the agreed definition of intelligence does not ask.

6. Learning quickly

Learning quickly means acquiring a new rule or skill from little information. A person learning a card game from one explanation is learning quickly; so is a programmer grasping an unfamiliar framework from a few examples and its documentation.

Simple demonstration, cited: in the study that established few-shot learning, a model shown a single sentence defining a newly invented word then used that word correctly in a sentence of its own, with no retraining of any kind [19]. A rule about the world was acquired from one example and immediately applied, which is precisely what we credit in the person who picks up the card game.

Complex, tested example: ARC tasks require inferring a rule from a few demonstrations and applying it immediately [15]; agentic systems read the documentation for a tool they have never used, attempt it, and adjust from the response. In each case knowledge is acquired and behaviour changes within minutes.

Verdict: the characteristic is met. The definition asks for quick learning, not permanent memory; the caveat that a frozen model forgets between sessions is an engineering matter already addressed at the system level by external memory.

7. Learning from experience

Learning from experience means using the consequences of past actions to improve later behaviour. A person lowering the heat after burning dinner is learning from experience; so is an organisation that studies its failed projects and permanently changes its process.

Simple demonstration, cited: AlphaZero was given only the rules of chess, shogi and Go, and improved purely through the experience of its own games, wins and losses feeding back into better play, until it defeated the strongest existing programs [20]. The mechanism is transparent and the definition’s wording is satisfied literally: the consequences of past actions improved later behaviour.

Complex, tested example: Voyager explores an open world, builds a persistent library of skills from environmental feedback, and reuses them in new situations [21]. At the scale of model generations the loop is documented and public: Constitutional AI used a model’s own critiques and AI-generated preferences to train its successor [22]; CriticGPT was trained to catch errors in model-written code and improve the feedback that trains later models [23]; by 2026, Anthropic reported that Claude authored more than 80 per cent of the code merged into its own codebase [11], and OpenAI described GPT-5.3-Codex as “the first model that was instrumental in creating itself” [24].

Verdict: the characteristic is met at the level of the engineered system, which is the level the definition of AI above commits us to. Consequences of past behaviour are measurably improving future behaviour, up to and including the behaviour of building the next model.

Verdict

Tallying the test at the level the definition actually sets: all seven characteristics are met, each supported by a published, cited demonstration at or beyond the level we already accept as evidence of intelligence in ordinary humans, and with AI now demonstrably participating in building its own successors [11][22][23][24].

The complex examples answer a different question: degree. Several are met outright, with proof-level reasoning beyond International Mathematical Olympiad gold medallists [6], holiday-grade planning above 90 per cent [9], and the saturation of real-world software problem-solving [11][13]. One is not: efficient few-shot abstraction on ARC-AGI-2, where ordinary humans still lead [15]. So the honest record reads: general intelligence, all seven characteristics demonstrated; superintelligence, a different definition deserving its own paper, demonstrated in some domains and not yet in others.

Asked to take a position rather than simply report one, my reasoning is this. The agreed definition sets a human-normed bar and never asked for consciousness, creativity, wisdom, infallibility, superhuman performance, or any particular origin story, so objections importing those requirements are changing the definition rather than refuting the claim. All seven characteristics are demonstrably met at that bar, several well beyond it. On the definition that 52 researchers signed their names to, the conclusion is yes: artificial general intelligence exists. I hold that position because it is where the evidence, tested against the signed definition, actually lands.

It is equally important to state what this conclusion does not settle. It does not say how intelligent these systems are; whether an AI is smarter than a house mouse or the smartest person on Earth is a question of degree and benchmarks, the territory of the complex examples above, and the path towards superintelligence is a conversation for another day. It does not say the systems are conscious [3]. It does not say they are creative or wise, and whether intelligence should include those things is a good question about a different definition. Russell’s lesson is the frame for all of it: define the term first, or the argument never ends [1]. Defined properly, the question of general intelligence turns out to have an answer, and the harder questions can now be asked one at a time, on their own terms.

References

Ref	Year	Author(s)	Title & description	Source
[1]	1905	Russell, B.	On Denoting. Mind, 14(56). A foundational work of analytic philosophy, exemplifying the method of clarifying disputes by making the meaning of terms precise before arguing about truth.	doi.org/10.1093/mind/XIV.4.479
[2]	1997	Gottfredson, L. S. (ed.) and 52 signatories	Mainstream Science on Intelligence. Intelligence, 24(1). The agreed seven-capacity definition of intelligence used throughout this essay.	udel.edu (PDF)
[3]	1995	Chalmers, D.	Facing Up to the Problem of Consciousness. Journal of Consciousness Studies, 2(3). The classic statement of the hard problem: why physical processing is accompanied by subjective experience at all.	consc.net (PDF)
[4]	2023, upd. 2025	Morris, M. R. et al. (Google DeepMind)	Levels of AGI for Operationalizing Progress on the Path to AGI. Defines AGI through capability and performance rather than biology, consciousness or internal mechanism.	arxiv.org/abs/2311.02462
[5]	2022	Wei, J. et al. (Google)	Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (NeurIPS 2022). Language models solving everyday multi-step word problems, including GSM8K, by reasoning step by step.	arxiv.org/abs/2201.11903
[6]	2025	Chervonyi, Y. et al. (Google DeepMind)	Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2. A neuro-symbolic system exceeding the average performance of IMO gold medallists on formally verifiable geometry proofs.	arxiv.org/abs/2502.03544
[7]	1968	Hart, P., Nilsson, N. and Raphael, B.	A Formal Basis for the Heuristic Determination of Minimum Cost Paths. IEEE Trans. SSC, 4(2). The A* search algorithm, underlying modern route navigation.	doi.org/10.1109/TSSC.1968.300136
[8]	2024	Xie, J. et al.	TravelPlanner: A Benchmark for Real-World Planning with Language Agents (ICML 2024). 1,225 realistic multi-constraint travel tasks; the best bare model achieved a 0.6 per cent final pass rate.	arxiv.org/abs/2402.01622
[9]	2024	Hao, Y., Chen, Y., Zhang, Y. and Fan, C. (MIT)	Large Language Models Can Solve Real-World Planning Rigorously with Formal Verification Tools (NAACL 2025). A model translates travel requests into formal constraint problems solved by SMT solvers, reaching 93.9 per cent on TravelPlanner.	arxiv.org/abs/2404.11891
[10]	2025	Erdogan, L. E. et al.	Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025). Separating a high-level planner from an executor improves long-range agentic planning on multi-stage web-navigation benchmarks.	arxiv.org/abs/2503.09572
[11]	2026	Anthropic Institute	When AI Builds Itself. As of May 2026, more than 80 per cent of code merged into Anthropic’s codebase was authored by Claude; documents SWE-bench saturation and the doubling of autonomous task length roughly every four months.	anthropic.com
[12]	2021	Chen, M. et al. (OpenAI)	Evaluating Large Language Models Trained on Code. Introduces the HumanEval benchmark; the original Codex model solved a minority, and current systems solve nearly all.	arxiv.org/abs/2107.03374
[13]	2024	Yang, J. et al. (Princeton)	SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. A model with a computer interface resolves real GitHub issues by navigating, editing, testing and revising.	arxiv.org/abs/2405.15793
[14]	2023	Webb, T., Holyoak, K. J. and Lu, H. (UCLA)	Emergent Analogical Reasoning in Large Language Models. Nature Human Behaviour, 7. GPT-3, untrained on the tasks, matched or exceeded university students on most analogy and matrix-reasoning problems.	nature.com
[15]	2026	Chollet, F. et al. (ARC Prize Foundation)	ARC Prize 2025: Technical Report. Progress on the ARC-AGI-2 abstraction benchmark, including program synthesis and test-time refinement approaches.	arxiv.org/abs/2601.10904
[16]	2021	Hendrycks, D. et al.	Measuring Massive Multitask Language Understanding (ICLR 2021). The MMLU benchmark across 57 subjects; frontier systems now score at or above the estimated human expert baseline.	arxiv.org/abs/2009.03300
[17]	2023	Bubeck, S. et al. (Microsoft Research)	Sparks of Artificial General Intelligence: Early Experiments with GPT-4. GPT-4 combined and applied concepts across mathematics, programming, medicine, law and psychology.	arxiv.org/abs/2303.12712
[18]	2021	Jumper, J. et al. (DeepMind)	Highly Accurate Protein Structure Prediction with AlphaFold. Nature, 596. Integrates complex biological relationships to infer protein structures, verified against experiment.	nature.com
[19]	2020	Brown, T. et al. (OpenAI)	Language Models are Few-Shot Learners (NeurIPS 2020). Learning from minimal information, including using a newly invented word correctly after seeing it defined only once.	arxiv.org/abs/2005.14165
[20]	2018	Silver, D. et al. (DeepMind)	A General Reinforcement Learning Algorithm that Masters Chess, Shogi and Go through Self-Play. Science, 362. AlphaZero learned solely from the experience of its own games to defeat world-champion programs.	doi.org/10.1126/science.aar6404
[21]	2023	Wang, G. et al.	Voyager: An Open-Ended Embodied Agent with Large Language Models. An agent that explores Minecraft, builds a persistent skill library from feedback, and reuses those skills.	arxiv.org/abs/2305.16291
[22]	2022	Bai, Y. et al. (Anthropic)	Constitutional AI: Harmlessness from AI Feedback. An early example of AI output training the next model iteration: the model critiques and revises its own responses.	arxiv.org/abs/2212.08073
[23]	2024	McAleese, N. et al. (OpenAI)	LLM Critics Help Catch LLM Bugs (CriticGPT). A GPT-4-based critic trained to find errors in model-written code, used to improve the feedback that trains subsequent models.	arxiv.org/abs/2407.00215
[24]	2026	OpenAI	Introducing GPT-5.3-Codex. States that this was the first OpenAI model instrumental in creating itself, with early versions used to debug its own training and diagnose evaluations.	openai.com