AGI (artificial general intelligence)

Why it matters

AGI is one of the most contested ideas in the AI conversation, which is exactly why it pays to understand it. When headlines promise “human-level AI” or warn of its dangers, they are talking about AGI. For a decision-maker the key skill is telling apart what AI can genuinely do today from what AGI refers to: the latter is still hypothetical, and there is no consensus on its precise definition or arrival date. Getting that distinction right guards against both overhype and unwarranted fear.

It also matters commercially, because AGI’s assumed arrival drives large investment and regulation decisions. When you weigh what AI means for your organisation, it helps to know that today’s tools are narrow AI that already delivers practical value, and that AGI is a separate, open question rather than a product feature.

What is AGI?

AGI, or artificial general intelligence, is a hypothetical AI that could match or beat people at essentially any mental task, not just narrow ones. Wikipedia frames it as an AI that “matches or surpasses human capabilities across virtually all cognitive tasks”. The decisive difference from today’s AI is generality: an AGI could generalise what it learns, transfer skills from one domain to another and solve novel problems without being programmed for each task.

The definition is contested, and that is worth stating plainly. Wikipedia notes there is no single agreed-upon definition of intelligence as applied to computers. Some researchers hold that AGI would require consciousness, goal-setting or emotions; others define it purely through measurable capabilities. As a result the same word means different things to different people.

How is AGI different from narrow AI and superintelligence?

AGI sits between two other ideas. Narrow AI (also ANI) is all the AI we use today: its competence is confined to well-defined tasks such as translation, image recognition or text generation. It can be superhuman within its narrow lane and helpless outside it. AGI would be general: one system handling a wide range of tasks the way a person does.

Superintelligence (ASI, artificial superintelligence) would go further still. Wikipedia describes it as outperforming “the best human abilities across every domain by a wide margin”. So the order is narrow AI (now) → AGI (human level, hypothetical) → superintelligence (beyond human, hypothetical). AGI and superintelligence are both still unrealised; they are subjects of research and debate, not features of current products.

Does AGI already exist, and how is progress measured?

No, but the boundary is blurry, which is why researchers now talk about levels of AGI rather than a single threshold. Google DeepMind researchers (Morris et al., 2023) proposed a framework that rates AGI on two axes: depth of performance and breadth, or generality, of capability. They distinguish five performance levels: Emerging, Competent, Expert, Virtuoso and Superhuman. In the same work they classify current frontier language models as “Emerging AGI”, which they describe as comparable to an unskilled human. A central principle of the framework is to judge what a system can do, not how it works internally.

Practical measurement leans on benchmarks. According to Stanford HAI’s AI Index 2025, AI improved its scores on demanding new tests within a year: MMMU by 18.8, GPQA by 48.9 and SWE-bench by 67.3 percentage points, and language-model agents outperformed humans in time-limited programming tasks. These are still narrow measures, though: a strong score on one test is not the same as general intelligence.

When might AGI arrive, if ever?

Nobody knows, and predictions vary widely, but the largest recent data point is Grace et al. (2024), who in 2023 surveyed 2,778 researchers who had published in top-tier AI venues. It found that machines would outperform humans at every possible task with 50% probability by 2047, and with 10% probability by 2027; the 2047 figure had moved 13 years earlier than in the equivalent 2022 survey. (The survey’s milestone is a task-based “all tasks” threshold, closely related to but not identical with AGI.) Predictions still deserve caution: a Wikipedia-collected meta-analysis of 95 forecasts made between 1950 and 2012 found a strong tendency to place AGI 15 to 25 years from the moment of predicting. In other words, “about 20 years away” has recurred decade after decade.

For that reason this glossary does not offer a single figure. Estimates range from a few years to several decades or never, depending on who you ask and how AGI is defined.

What can today’s AI do, and what can’t it?

Today’s AI is impressive in a narrow lane but brittle at general reasoning, and that contrast is why AGI remains distant. Stanford’s AI Index 2025 gives both sides. The strength: models excel at problems like the Mathematical Olympiad and improve benchmark scores fast. The weakness: the same models stumble on complex planning and reasoning, such as the PlanBench test, where they fail logic tasks even when a provably correct solution exists. Generalisation, reliable multi-step reasoning and handling genuinely new situations are precisely the capabilities that would separate AGI from today’s AI, and they are still missing.

Is AGI even possible?

It is an open question, and experts genuinely disagree. Some treat AGI as an inevitable matter of time; others doubt whether current methods even point in the right direction. The dispute is tied to the definition: if AGI required consciousness or human-like understanding, the question moves into philosophy and cannot be settled by benchmarks alone. Several illustrative yardsticks have been proposed, such as the coffee test (can a machine make coffee in an unfamiliar home), the Ikea test (can it assemble furniture from instructions) or Suleyman’s test (can an AI turn $100,000 into $1 million). None has become the accepted arbiter, which shows how open the question still is.

AGI risks and safety: why is it debated?

The AGI debate runs hot because, if it were realised, the stakes would be large in both directions. On one hand, a general-purpose AI could accelerate science and the economy. On the other, a system acting more broadly and autonomously than a person raises questions of control and alignment: how do you ensure a capable system does what a human actually intended. This is the core territory of AI safety research. It is worth keeping two things separate, though: the concrete harms of today’s narrow AI (such as errors, bias and misuse) and the hypothetical longer-term risks tied to AGI. Both are legitimate, but they are not the same conversation.

What does AGI mean for business and decision-makers?

The practical takeaway is measured: base decisions on what AI can do today, not on AGI arriving in a particular year. That is the ground on which to invest. AGI’s role in decision-making is mainly to watch the direction and keep some flexibility, not to buy promises. When a vendor promises “AGI” or “human-level AI”, treat it as a prompt for sharper questions, not a mature feature. A balanced, source-based understanding here helps separate real value from promises.

Frequently asked questions

Is AGI the same as ChatGPT or other current AI tools?

No. Today’s tools are narrow AI: skilled at bounded tasks but without the general, transferable capability of AGI. In Google DeepMind’s framework (Morris et al., 2023), frontier language models are classified as “Emerging AGI”, described as comparable to an unskilled human, not full AGI. AGI is still hypothetical.

What is the difference between AGI and superintelligence (ASI)?

AGI would mean human-level general capability. Superintelligence (ASI) would, per Wikipedia, outperform the best human abilities across every domain by a wide margin. The order is narrow AI (now) → AGI (human level) → superintelligence (beyond human). AGI and superintelligence are both still unrealised.

When might AGI arrive?

There is no reliable single year. In the largest recent survey (Grace et al., 2024; 2,778 AI researchers in 2023), machines were forecast to outperform humans at all tasks with 50% probability by 2047, 13 years earlier than the previous year’s survey. The spread is still wide, and predictions have a historical tendency to land about 15–25 years out. Treat any single date with caution.

Is AGI definitely coming?

There is no certainty. Experts disagree on whether AGI will be reached with current methods, decades away, or at all. Some see it as a matter of time; others doubt the whole direction. The uncertainty is itself an essential part of the concept.

Should businesses prepare for AGI now?

Watch the field, but base decisions on the real capabilities of today’s AI, not an assumed AGI timeline. AGI is an open question to keep an eye on while staying flexible, not to buy as a promise.

Sources

• Morris et al., Google DeepMind (2023): Levels of AGI for Operationalizing Progress on the Path to AGI – the levels of AGI, the measurement axes and the “Emerging AGI” classification of current models.
• Wikipedia: Artificial general intelligence – the contested definition, narrow AI vs AGI vs superintelligence, the tests and the historical bias in predictions.
• Grace et al. (2024): Thousands of AI Authors on the Future of AI – a 2023 survey of 2,778 AI researchers and its timeline forecasts (machines outperforming humans at all tasks with 50% probability by 2047).
• Stanford HAI: AI Index 2025 – benchmark jumps and the limits of today’s AI reasoning.