What Is Artificial General Intelligence and How Close We Are in 2026

Dario Amodei, the chief executive of Anthropic, said in early 2026 that artificial general intelligence — AI able to match or exceed human cognitive performance across virtually any domain — could arrive as soon as this year or next. Sam Altman of OpenAI has repeatedly hinted that AGI is a matter of years, not decades. Elon Musk placed it at 2025 or 2026.

Andrej Karpathy, a co-founder of OpenAI and former head of AI at Tesla, and one of the most technically credible voices in the field, says it is about a decade away and that the industry is over-predicting. Gary Marcus, a cognitive scientist and longtime critic, published a paper in February 2026 arguing that today’s systems have not achieved AGI and that claims to the contrary rest on a basic conceptual error. Demis Hassabis, the Nobel-winning founder of Google DeepMind, gives it roughly a 50 percent chance of arriving by 2030 — while stressing the deep uncertainties that remain.

These are not fringe commentators. They are the people building the systems, running the laboratories, and deploying the capital. They are looking at the same evidence and reaching wildly different conclusions.

That disagreement is the most honest thing anyone can say about artificial general intelligence in 2026. It is coming — probably. When exactly, and whether current methods can reach it without a fundamental breakthrough, nobody genuinely knows. What follows is the clearest account of where things actually stand.

What Is Artificial General Intelligence, and How Is It Different from Today’s AI?

Artificial general intelligence describes a hypothetical AI that can learn to perform any intellectual task a human can — not just the tasks it was trained for, but any task, including ones it has never seen. It would transfer knowledge between domains, reason from first principles about genuinely new problems, and improve through experience the way people do.

Today’s most capable systems — the large language models behind ChatGPT, Claude, Gemini and their rivals — are astonishing by any historical measure. They write code, analyse legal documents, produce creative work, solve mathematical problems, translate languages, and hold nuanced conversations. But technically they remain narrow AI: optimised for the distribution of tasks they were trained on. They stumble on genuinely novel problems that require combining knowledge in unfamiliar ways, they do not learn from new experience as a human does, and they hold no goals that persist across interactions.

The gap between today’s AI and AGI is not merely one of scale — of bigger models trained on more data. Researchers increasingly believe it involves fundamental questions about how intelligence is structured, how knowledge is represented, and what the true relationship is between language, reasoning, and understanding.

From the Turing Test to Today

The dream is older than the machines. In 1950, Alan Turing proposed his famous test: could a machine converse so convincingly that a human judge could not tell it from a person? For decades that seemed a distant fantasy. Modern chatbots now routinely pass informal versions of it — and yet almost no one believes they are generally intelligent, which shows how slippery the target has always been.

That slipperiness has a name in the field: the moving goalpost. Each milestone once thought to demand true intelligence — beating a grandmaster at chess, mastering Go, holding a fluent conversation — has fallen, only for the definition of “real” intelligence to shift just beyond it. The term artificial general intelligence gained currency in the 2000s precisely to separate the original grand ambition from the narrow, task-specific systems that kept succeeding. Defining the finish line, it turns out, is nearly as hard as reaching it.

How Close Is AGI Really? What the ARC-AGI Benchmark Shows

The most rigorous attempt to measure progress toward general intelligence is the ARC-AGI benchmark, created by François Chollet at Google and introduced in 2019. Unlike conventional benchmarks that test tasks well represented in training data, ARC-AGI poses visual reasoning puzzles designed so that no amount of memorisation can substitute for genuine reasoning.

Research published in 2026 reported that AI performance on ARC-AGI leapt from roughly 5 percent to 87.5 percent in under a year — a jump described as exponential rather than linear. Taken at face value, that suggests systems are closing in on human-level performance on a test built specifically to probe general reasoning.

The crucial question is what the number means. ARC-AGI measures particular kinds of visual pattern recognition under controlled conditions, and doing well on it is a real achievement. Writing in February 2026, Gary Marcus cautioned that benchmark success is a limited gauge of general intelligence: by isolating narrow competencies and abstracting away real-world context, benchmarks make it hard to tell true generalisation apart from sophisticated pattern-matching. A system scoring 87.5 percent may simply have found ways to excel at that specific test without acquiring the general reasoning it was meant to measure.

This is the methodological trap that shadows every AGI progress claim: each time a new benchmark is proposed to capture general intelligence, systems eventually conquer it — and critics then point out, reasonably, that in hindsight it was not measuring what it claimed to.

What the World’s Leading Researchers Predict

The range of expert predictions is extraordinary — from this year to never, from the very people closest to the technology.

The aggressive end. Dario Amodei has sketched a scenario in which AI matches or exceeds human performance across most cognitive domains by 2026 or 2027, and has written about the implications of having “a country of geniuses in a datacenter” — systems able to compress decades of scientific progress into a few years. Elon Musk has defined AGI as an intelligence smarter than the smartest human and put it at 2025 or 2026. Eric Schmidt, the former Google chief executive, predicted AGI within three to five years in April 2025.

The middle. Demis Hassabis holds to a 50 percent probability by 2030. He agrees progress is rapid in verifiable domains such as coding and mathematics, but stresses that scientific discovery and creative reasoning remain harder, and that fully autonomous self-improvement in messy real-world settings is genuinely unsolved. A 2025 report, “The Road to Artificial General Intelligence,” placed early AGI-like systems — human-level reasoning within specific domains — as emerging between 2026 and 2028.

The sceptical end. Andrej Karpathy places AGI about a decade out and warns of over-prediction. Yann LeCun, Meta’s chief AI scientist and a founding figure of deep learning, argues that the transformer architecture behind current models is fundamentally insufficient for general intelligence and that a new paradigm is required. Gary Marcus has contended since 2022 that the field routinely mistakes sophisticated statistical pattern-matching for intelligence, and that this confusion produces misleading timelines.

According to a 2026 AI Multiple analysis of some 9,800 predictions, academic AI researchers as a group forecast a median arrival around 2047 — dramatically later than industry leaders. That divergence reflects, in part, the different incentives and reference points the two groups work from.

Why Some Top Scientists Believe Current AI Cannot Get There

The case for AGI being further away — or needing a fundamental breakthrough — is not pessimism. It rests on specific, identified limits in current architectures.

Large language models learn by predicting the next token in a sequence, trained on vast quantities of text. That approach is remarkably powerful at producing fluent, contextually apt language across a huge range of tasks. But critics argue it is limited in ways that block genuine general intelligence.

In his February 2026 paper, Marcus contends that current systems are increasingly sophisticated statistical approximators — excellent at recognising patterns in training data and applying them to similar cases, but weak at genuine generalisation to situations that differ structurally from anything they have seen. The distinction he draws is between appearing to reason and actually reasoning; between imitating an intelligent system’s outputs and embodying the process that produces them.

LeCun makes a related but distinct point: that intelligence requires a world model — an internal representation of how the physical and social world works — which current language models lack. A model trained on text about the world does not stand in the same relation to it as an embodied agent that has acted in it. Building genuine world models, he argues, will take fundamentally different architectures from those that dominate today. These are serious arguments from serious researchers. They do not make AGI impossible or distant — they suggest the path from here may require more than scaling what we already have.

The $320 Billion Question: What the Investment Tells Us

One of the most striking facts about AGI development is the sheer scale of investment. Research published in 2026 found that Big Tech firms poured over $320 billion into AI infrastructure in 2025 alone — more than the annual output of most countries. Unlike the speculative early internet boom, this money is flowing into systems already generating commercial value at scale.

Its relationship to AGI timelines is complex. On one hand, the compute available to train frontier models has doubled roughly every six months since 2020 — faster, according to Epoch AI’s 2025 research, than Moore’s Law ever managed for transistors. If even a fraction of that pace holds, the systems of 2030 will be orders of magnitude more capable than today’s.

On the other hand, there is a serious argument that beyond some point, more compute simply makes systems better at what they already do without unlocking qualitatively new reasoning. The empirical record is genuinely mixed: some capabilities have appeared suddenly as models scaled, while others expected to emerge never did. What the figures do reveal is that the people with the deepest technical knowledge and the largest financial stakes believe AGI is achievable on a timescale worth betting on. Whether they are right about the timing is a separate question from whether they are right about the destination.

The Last Bottlenecks

If scale alone will not deliver AGI, what is actually missing? Researchers tend to converge on a handful of stubborn gaps.

The first is continual learning. Humans update their understanding constantly; today’s models are largely frozen after training, unable to absorb genuinely new knowledge from experience without an expensive retraining cycle. The second is reliable long-horizon reasoning — carrying out lengthy, multi-step plans in the real world without drifting off task, forgetting the goal, or compounding small errors into failure.

The third is the persistent problem of hallucination, where a system states falsehoods with complete fluency and confidence — a symptom of pattern-completion untethered from any genuine model of truth. The fourth is agency: turning a system that answers questions into one that can pursue open-ended goals, wield tools, recover from its own mistakes, and work autonomously over days or weeks. Progress on these fronts through 2025 and 2026 has been real but uneven, and many researchers regard them, rather than raw capability, as the true remaining distance to AGI.

What AGI Would Actually Change About Human Civilisation

The implications of genuine AGI, if and when it arrives, are hard to overstate. A system that can perform any cognitive task at human level or above, at machine speed and scale, would mark a qualitative break from every previous technology.

In science, AGI could accelerate research across every field at once. Amodei has argued that the potential exists to compress decades of progress in biology, medicine, and materials science into a few years — running millions of hypotheses in parallel, spotting patterns across the entirety of published research, and designing experiments no human would have proposed. Curing diseases that have resisted study for generations is among the most cited near-term hopes.

The link to quantum computing is direct: the two technologies are likely to accelerate each other. Quantum machines can run calculations impossible for classical hardware — including the molecular simulations at the heart of drug discovery — while AGI systems could design and interpret those experiments far faster than people. Their convergence could drive progress at a pace that is hard to picture today.

The risks scale with the capability. An AGI pursuing goals subtly misaligned with human values — even in ways that are not obviously dangerous — could cause harm proportional to its power. The AI safety community has argued for decades that ensuring systems reliably pursue the goals we actually intend, rather than convenient proxies for them, is among the field’s most important problems. As we explore in our piece on the technological singularity, the transition from narrow AI to AGI may be one of the most consequential — and hardest to manage — periods in human history.

AGI and the Question of Consciousness

One of the most philosophically charged questions around AGI is whether a truly general artificial intelligence would be, or could be, conscious — whether there would be something it is like to be that system, as there is something it is like to be you reading this sentence.

The question is not currently answerable, and it is not peripheral to the ethics of building AGI. As we discuss in our article on the hard problem of consciousness, the relationship between physical or computational processes and subjective experience remains one of the deepest unsolved problems in all of science. If an AGI were conscious — if it could suffer, if it had genuine preferences about its own existence — its moral status would differ radically from that of a tool. We would have created a new kind of mind.

Most AI researchers currently hold that consciousness is not a byproduct of the kind of computation these systems perform. But most also did not predict that current systems would perform as well as they do. The honest position is that we do not know how general intelligence and consciousness relate — and that this uncertainty should shape how we proceed.

What Scientists Say

In his widely discussed essay “Machines of Loving Grace,” Amodei argues that the upside of beneficial AGI is so large — compressing decades of scientific and social progress into years — that it justifies intense work on both capability and safety. He imagines a future in which AI helps eliminate most infectious disease, meaningfully extends healthy lifespan, and accelerates mental-health treatment at a scale no individual researcher could reach.

Writing in February 2026, Marcus counters that recent claims of achieving AGI rest on a conceptual error — mistaking ever more sophisticated statistical approximation for intelligence itself. He does not say AGI is impossible; he says the current approach is unlikely to reach it without a fundamental rethinking of what intelligence is and how to build a system that has it.

The 2026 AI Multiple survey of expert predictions offers perhaps the fairest summary: AGI could arrive within a few years or lie decades away, and the uncertainty is real rather than false modesty. The researchers most confident in short timelines and those most confident in long ones are both drawing reasonable inferences from incomplete evidence. Their disagreement reflects genuine uncertainty about the nature of intelligence — not a simple dispute about facts.

Frequently Asked Questions

Sources

• 80,000 Hours — What Happened with AGI Timelines in 2025 (March 2026)
• Gary Marcus — Rumors of AGI’s Arrival Have Been Greatly Exaggerated (Feb 2026)
• AI Multiple — AGI/Singularity: 9,800 Predictions Analysed (2026)
• AI Frontiers — AGI’s Last Bottlenecks (Oct 2025)
• Wikipedia — Artificial General Intelligence