What is AGI in simple terms?

Artificial General Intelligence is an AI that can understand, learn and apply knowledge across any intellectual task a well-educated adult can — not just one domain. It generalizes, reasons with common sense, and adapts to unseen situations without being rebuilt for each task.

How is AGI different from today's AI?

Today's AI is narrow (ANI): it excels in one domain (e.g., coding, vision, Go) but does not transfer broadly. AGI is general: it transfers skills across domains, learns efficiently from few examples, and handles open-ended environments with limited resources. Most definitions require adaptation without human developers rewriting the system for each new problem.

What are the levels of AGI?

DeepMind's 2023 levels framework crosses capability (Narrow → General) with performance (Emerging → Superhuman). Five useful bands are: No AI, Emerging Narrow, Competent Narrow (today's frontier models on many tasks), Emerging General (weak AGI), Competent General (human-level AGI) and Superhuman General (ASI). Most labs place current models at Competent Narrow on some tasks but far from Emerging General.

How do we test for AGI?

By breadth and depth across ten cognitive broad abilities (Cattell-Horn-Carroll: reasoning, memory, perception, etc.) using psychometric batteries adapted for AI. Recent work reports AGI scores: GPT-4 around 27%, GPT-5 around 57-58% — rapid progress but with jagged profiles and weak long-term memory, planning and embodied reasoning. No consensus benchmark yet; the key is independent, write-protected holdouts.

When will AGI arrive?

There is no consensus. Public estimates from frontier labs in 2025-2026 range from 2028 to 2040+ for competent AGI, with many researchers putting a wide uncertainty band. Anthropic and DeepMind forecasts cluster earlier if RSI accelerates, while critics argue embodiment and causality gaps push it decades away. Treat dates as forecasts, not commitments.

Does AGI mean AI will be conscious?

No. AGI definitions focus on cognitive capability — solving tasks, generalizing, adapting — not consciousness, self-awareness or felt experience. Philosophical debates about whether a system that behaves intelligently is 'conscious' are separate from operational AGI tests.

What happens after AGI — what is ASI?

Artificial Superintelligence (ASI) is hypothetical AI that outperforms the best humans across every domain by a wide margin. In the DeepMind framing, ASI is Superhuman General. The distinction matters: AGI matches humans, ASI decisively exceeds them.

What Is AGI? The Complete Guide to Human-Level AI in 2026

AGI is the most searched, least agreed-upon term in AI. Everyone from Sam Altman to Yann LeCun uses it, but they rarely mean the same system — and that ambiguity is not accidental.

This guide gives you a durable definition you can reuse, the five levels that make comparisons honest, how we actually measure progress today, and where we stand in late 2026 — so you can judge the next “AGI is near” headline yourself.

What AGI Is — and Why the Definition Matters

In 1956 John McCarthy coined “artificial intelligence” to mean, roughly, machines that can do tasks requiring intelligence when done by humans. That was AGI: the original AI was general.

The qualifier “general” reappeared because the field succeeded at narrow tasks: Deep Blue beat Kasparov at chess but could not book a hotel. The “AI effect” — once a system works, we stop calling the underlying skill “intelligence” — made narrow wins feel less like AI each decade.

Modern definitions converge on three invariants, stated most clearly by Bowen Xu (2024) and DeepMind’s levels paper (2023):

Adaptation to open environments with limited resources. The system must acquire problem-relevant knowledge itself, not have it hand-coded for each new domain.
No human rewiring per task. An algorithm is general if a developer can reuse it across problems; a system is general if, after deployment, no developer needs to intervene in source code to handle a new task. AGI is the latter.
Breadth + depth. Not just specialized performance but the versatility and proficiency of a well-educated adult across reasoning, memory and perception — the standard borrowed from human psychometrics.

In one sentence from a 2025-2026 framework adopted by several labs:

AGI is an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult across the ten broad abilities that define human intelligence.

If the system still needs a new architecture per domain, it is powerful narrow AI, not AGI.

Narrow vs General vs Super — The Levels That Prevent Hype

DeepMind’s 2023 survey (Morris et al., refined 2024-2025) maps AGI on two axes — how broad the capability is and how well it performs:

Performance → Capability ↓	Emerging	Competent	Expert	Virtuoso	Superhuman
Narrow (one domain)	Emerging Narrow	Competent Narrow ← most frontier LLMs today on many benchmarks	Expert Narrow	Virtuoso Narrow	Superhuman Narrow (e.g., AlphaGo on Go)
General (across domains)	Emerging General (weak AGI)	Competent General = AGI	Expert General	Virtuoso General	Superhuman General = ASI

How to read this: Today’s frontier models are Competent or even Superhuman on narrow tasks — coding on HumanEval, image classification, Go — but they remain Emerging or below on general breadth. They transfer poorly without scaffolding, forget long contexts, and lean heavily on retrieval. Calling them “AGI” collapses the matrix.

Practitioners often simplify to five bands for communication:

No AI → Narrow AI (today) → Emerging AGI → Competent AGI (human-level) → ASI (superhuman general)
         ↑ we are here on many tasks

The simplification is honest only if you keep breadth and performance separate. A model can be expert-level at mathematics and emerging at embodied common-sense — that jaggedness is the signature of current systems.

How We Test for AGI — The CHC Framework and the Scores That Exist

Humans are the only proven example of general intelligence, so the most durable AGI tests borrow from human psychometrics — not trivia quizzes.

The 2025 framework by Hendrycks et al. (updated Oct-Dec 2025, co-authored by Bengio, Tegmark, Brynjolfsson and others) grounds AGI evaluation in Cattell-Horn-Carroll (CHC) theory, the most empirically validated model of human cognition. It decomposes intelligence into ten broad abilities and dozens of narrow abilities — fluid reasoning, crystallized knowledge, working memory, long-term storage and retrieval, visual processing, auditory processing, processing speed and others — then adapts established batteries for AI.

What the numbers say in 2026:

GPT-4: ~27% AGI on this battery — proficient in knowledge-intensive domains, weak in long-term memory storage and retrieval.
GPT-5: ~57-58% AGI — rapid progress (+30pp generation-over-generation) but still a jagged profile: strong reasoning with tools, fragile on tasks requiring persistent memory over weeks, embodied planning, and causal world models.
FrontierCoral / HCI (ArXiv 2609.11873): Uses Headroom-Closed Index to show where RSI would matter most — domains with largest remaining gaps (long-context memory, scientific discovery). On hard coding/science holdouts, closed-loop gains are smallest where verifiers are weakest.

Why naturalistic benchmarks beat LMSYS-style leaderboards for AGI: Leaderboards sample narrow tasks with internet-leaked data. CHC batteries require transfer, few-shot adaptation and memory over time — exactly the skills that define general intelligence. The strongest signal, per the 1,250-paper RSI survey (ArXiv 2607.07663), is whether gains hold on write-protected holdouts the model never saw or edited. If not, capability moved via grader loosening, not intelligence.

Unique angle — The Disagreement Matrix: When you see an AGI prediction, check which column the predictor is optimizing. Our table distills public statements 2024-2026 into testable commitments.

Who	Public stance (paraphrased, 2024-2026)	Operational definition they use	Implicit test they would accept	Year band they imply
DeepMind / Hassabis	Levels framework; SRO for testing powerful systems	Competent General across human-solvable tasks	CHC Competent General + safety evals	Early-mid 2030s if verified
Anthropic / Amodei	Pacing the frontier; RSI brake pedal needed	Closed-loop RSI + general task transfer	Independent holdout gains + verifier integrity	2027-2030 headroom focus, IPO pressure
OpenAI / Altman	US-led standards; “not pursuing autonomous RSI unsafely”	Same CHC-style breadth + RSI safety	US standards + CAISI verification	2028-2033, conditional on safety
Meta / LeCun	AGI via world models; LLM sceptic	Grounded, embodied intelligence	Robotics + causal physics benchmarks	2035+ (LLM path insufficient)
Gartner (enterprise lens)	Three stances: imminent / impossible / unpredictable	Organization-ready capability	ROI on composite/neuro-symbolic systems	”Prepare for all, predict none”

If two predictions differ but use different definitions, they are not disagreeing — they are measuring different systems.

Timeline: From Dartmouth to 2026 in Ten Beats

1956 Dartmouth: McCarthy coins “artificial intelligence” seeking human-level machines.
1958-1976: Symbolic era: Lisp, Advice Taker, commonsense logic. Intelligence = formalizable reasoning.
1980s: Expert systems: Narrow success (XCON, MYCIN) → Lighthill report → AI Winter. The AI effect begins.
1997: Deep Blue beats Kasparov. World calls it “not really AI” within years.
2012-2017: Deep learning returns: AlexNet → ResNet → Transformer (2017). Narrow superhuman spreads.
2017-2022: Scaling hypothesis: Pre-training compute predicts capability; AlphaGo/AlphaFold narrow superintelligence.
2022-2023: ChatGPT moment: General-purpose interfaces convince public that breadth is near — but evaluation shifts to leakage-aware tests.
2024: DeepMind Levels of AGI. Field agrees to stop arguing words and start measuring breadth × performance.
2025: CHC batteries for AI. GPT-4 at 27%, GPT-5 near 57% — quantifiable gap, not vibes.
2026: RSI enters policy language. UN Security Council briefing Sep 23, OpenAI/Anthropic standards talk, California SB 813/AB 1405 — AGI governance becomes deployment governance.

What AGI Is Not — Three Trapdoors

Trapdoor 1: “AGI = chatbot that never fails a demo.” Demos are selected for narrow excellence. General intelligence requires unselected competence — handling the next unseen task without prompt engineering. Ask: was the evaluator fixed and independent?

Trapdoor 2: “AGI = consciousness.” Operational AGI tests capability, not phenomenology. A system can score Competent General on CHC tasks without any claim about felt experience. Keep consciousness debates in philosophy, AGI in measurement.

Trapdoor 3: “AGI is the finish line.” DeepMind’s matrix shows AGI (Competent General) is a midpoint, not an end. ASI (Superhuman General) is a separate level with distinct governance because it exceeds the best humans everywhere — which is where control and economic displacement arguments actually activate.

Living With Narrow AI Until AGI Arrives — What to Build Today

Design for generality, deploy narrowly. Build systems that could generalize (retrieval, tools, verifiers) but ship them gated to one domain with audited evaluators. That’s how you satisfy both capability and safety.
Budget for memory. The 27% → 57% gap is not trivia recall — it is long-term storage and retrieval, planning over weeks, and world-model coherence. If your agent forgets last month’s decision, no parameter scaling will save it; architectural memory will.
Pick your AGI definition and publish it. When customers ask “are you AGI-ready?” answer with levels and with a test you’d accept — not a date. “We define AGI as Competent General on CHC; we track progress via independent holdouts audited by [verifier].” That sentence alone puts you ahead of most marketing.
Track the right dashboard:
- Breadth × performance matrix position, not single benchmark score
- Evaluator lineage (who wrote the test, who guarded the holdout)
- Verifier strength (formal tests > judge models > self-assessment)

Why This Guide Ages Well (the “Forever” Design)

This page is built to remain useful whether AGI arrives in 2028 or 2045:

Definitions anchor in CHC and DeepMind levels, not model names. When GPT-6 and Claude 5 ship, re-score them in the same matrix.
Links point to durable topics, not news: What Is Artificial Intelligence?, How Large Language Models Actually Work, Transformer Architecture Explained, RAG vs Fine-Tuning, Recursive Self-Improvement.
UpdatedDate bumps only on meaningful measurement changes (new CHC release, verified benchmark on write-protected holdout), not on press releases.

Bookmark it. The next viral “AGI leaked” thread will be easier to parse after you know which level the poster is actually claiming.

Sources

DeepMind — Levels of AGI — Morris et al. 2023 (updated 2024-2025) — Narrow/General × performance matrix
ArXiv 2510.18212 — A Definition of AGI — Hendrycks et al. Oct 2025 (Hendrycks, Bengio, Tegmark, Schmidt et al.) — CHC framework, GPT-4 27% / GPT-5 57-58%
ArXiv 2404.10731 — What Is Meant by AGI? — Bowen Xu 2024 — adaptation to open environments, principles P_G
Google Cloud & IBM Think — What Is AGI? (updated July 2026) — enterprise definitions, generalization vs narrow
Gartner — Artificial General Intelligence: 5 Perspectives — imminent/impossible/unpredictable frames
Merit & arXiv 2607.07663, 2609.11873 — verifier hierarchy, HCI headroom, RSI context for AGI acceleration
Reuters — AI leaders to brief UN — Sep 23, 2026 — why definitions are now governance terms, not academic trivia