What Is Artificial General Intelligence (AGI)? Definition, Examples and Where We Actually Are

Artificial general intelligence is the point at which a machine can handle essentially any cognitive task a person can, rather than the one narrow job it was built for. It does not exist. Every system you can use today, including the best ones, is a specialist wearing a very convincing generalist costume. That much is agreed on by IBM, Stanford HAI, Amazon and Google alike, and it has been the standard answer for roughly a decade.

What changed is that AGI used to be an argument you could not settle. For years the honest answer to “how close are we?” was a shrug dressed up in citations. Since late 2025 that is no longer true. Two serious research groups have published frameworks that turn the question into a measurement, and the resulting numbers are stranger and more useful than either the hype or the dismissal suggests. This article covers what AGI means, why the definition is contested, and what the first real scores actually show.

The Key Takeaways

• The definition: AGI is a hypothetical AI that matches human performance across essentially all cognitive tasks, transferring what it learns to unfamiliar problems without retraining. No AGI exists today.
• There is no agreed bar: researchers, labs and contracts use meaningfully different definitions, which is why arguments about whether AGI has arrived almost never resolve.
• It now has a score: a psychometric definition published in October 2025 rates GPT-4 at 27% and GPT-5 at 57% of a well-educated adult’s cognitive range.
• The shape matters more than the number: today’s models are described as “jagged”, strong on knowledge and reasoning, badly weak on long-term memory and learning from experience.
• The most valuable AGI definition ever written is gone: the word appeared eight times in OpenAI’s October 2025 Microsoft agreement and zero times in the April 2026 amendment that replaced it.

What Is Artificial General Intelligence?

Artificial general intelligence (AGI) is a hypothetical AI system that can match human performance across essentially any cognitive task, rather than one narrow job. Unlike today’s AI, an AGI would transfer what it learns to unfamiliar problems without being retrained. No AGI exists, and researchers do not agree on a single definition of what would count.

What “general” actually means

The load-bearing word is general, and it is doing more work than most coverage admits. It does not mean “very good”. A chess engine that beats every human alive is not close to AGI, because it does exactly one thing. Generality is about range and transfer: the ability to take a skill learned in one place and apply it somewhere it was never trained, the way a person who learns to drive a car has a real head start on driving a van.

That is the capability current systems most conspicuously lack. A large language model can pass a bar exam and then fail at keeping track of what you told it forty minutes ago. It is not that the model is weak. It is that its strengths and weaknesses are distributed in a shape nothing in nature has, which turns out to be the most interesting fact in this whole field.

AGI in AI is not AGI on your tax return

Worth clearing up quickly, because it sends a lot of people to the wrong page. In artificial intelligence, AGI means artificial general intelligence. In finance it means adjusted gross income, which is a line on a tax return and has nothing to do with any of this. Both acronyms are common enough that search engines routinely mix them up.

AGI vs Narrow AI vs Superintelligence

Almost every explainer describes this distinction in prose, which makes it harder to hold than it needs to be. The three tiers are easier side by side.

The row people find surprising is the last one under narrow AI. ChatGPT belongs in that column. It is astonishingly broad narrow AI, and breadth of training data is not the same thing as generality of mind, but the category is not in serious dispute among researchers. If you want the longer version of this ladder, we walk through every rung in our breakdown of the seven levels of AI.

Why Nobody Agrees What AGI Means

Here is the part that explains most of the noise. There is no single definition of AGI, and the competing ones are not minor variations. They disagree about what kind of thing intelligence even is.

The economic definition

OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” This is a labour market test. It says nothing about understanding, consciousness or common sense. A system qualifies by out-earning people, whatever is happening inside it. It is a coherent definition and a considerably easier one to satisfy than what most people picture.

The cognitive definition

The competing family of definitions asks whether the system has the underlying faculties: memory, reasoning, learning, perception, the machinery a mind runs on. Under this view a system that earns money while being unable to remember yesterday is not general, it is just useful. This is the definition both 2026 measurement efforts adopted, and it is much harder to pass.

The definitions that quietly expired

The Turing Test, for decades the popular shorthand for machine intelligence, no longer appears in serious frameworks. Modern chatbots pass casual versions of it routinely without anyone believing they are general. Definitions built on consciousness or self-awareness have faded for a different reason: nobody can measure consciousness in humans either, so a bar built on it can never be checked. Both still appear in older explainers, which is a decent way to date a page.

Are There Any Artificial General Intelligence Examples?

No. This deserves a blunt answer because a lot of pages fudge it, and some of the highest-ranking explainers on this exact question list ChatGPT, IBM Watson or self-driving cars as “examples of AGI”. They are not. They are examples of narrow AI, which is precisely the category AGI is defined against. Listing them as AGI examples inverts the concept.

What can be offered honestly are examples of the capabilities AGI would need, each of which current systems partly show:

• Transfer: solving a problem type never seen in training, by analogy to something that was.
• Continual learning: getting permanently better from experience, without a retraining run.
• Durable memory: carrying context across days and tasks rather than a single session.
• Self-assessment: knowing when it does not know, and saying so.
• Autonomous goals: breaking a vague objective into steps and adapting when they fail.

Fictional AGI is easier to name than real: HAL 9000, Samantha from Her, Data from Star Trek. That those remain the clearest reference points is itself informative about where the technology sits.

Can We Actually Measure Artificial General Intelligence?

Until recently the answer was no, and that answer is now out of date. This is the single biggest gap between what the top search results say and what the research actually shows.

The definition that produced a score

In October 2025 a large group of researchers including Dan Hendrycks, Yoshua Bengio, Max Tegmark, Gary Marcus and Eric Schmidt published a paper titled “A Definition of AGI”. It defines AGI as “matching the cognitive versatility and proficiency of a well-educated adult”, then grounds that in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition, and adapts human psychometric batteries to test AI systems across ten cognitive domains.

The result is a percentage. GPT-4 scores 27%. GPT-5 scores 57%. Those two numbers do more work than any amount of debate: they show real, fast progress and a large remaining gap at the same time, which is exactly the thing both the hype and the backlash get wrong.

DeepMind’s ten cognitive faculties

In March 2026 Google DeepMind published a cognitive framework for measuring progress toward AGI, authored by a team including DeepMind co-founder Shane Legg. It breaks general intelligence into ten faculties: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving and social cognition. Rather than a single verdict, a system gets a cognitive profile showing where it is strong and where it is empty.

Two independent groups, working from different traditions, landed on the same structure: ten dimensions, measured separately, no single pass or fail. That convergence is the most substantive thing to happen to the AGI question in years.

The jagged profile

Both efforts found the same shape, which the Hendrycks paper calls a “highly jagged cognitive profile”. Current models are proficient in knowledge-intensive domains and have critical deficits in foundational cognitive machinery, particularly long-term memory storage.

Put plainly: the hard parts turned out to be easy and the easy parts turned out to be hard. A model can hold expert-level knowledge across most of human science and still be unable to reliably remember what you asked it last Tuesday. No human mind is shaped like this, which is why our intuitions about what a system “must” be able to do keep failing.

What ARC-AGI shows

The other well-known yardstick is ARC-AGI, a benchmark built from puzzles designed to resist memorisation, so a model has to work out the rule rather than recall it. On the current ARC Prize leaderboard, the top verified result on the second-generation semi-private evaluation is 92.5%, from GPT-5.6 Sol at its highest reasoning setting, against a human panel that scores 100%. The first-generation version is effectively saturated, with top models now scoring above the human panel.

Note the two qualifiers, because they are routinely dropped: the score belongs to a specific reasoning setting and a specific evaluation set. Third-party leaderboards that omit them disagree wildly with each other. A third generation of the benchmark, testing whether agents can adapt inside novel interactive environments, is already running, which tells you how quickly the previous bar stopped being informative.

The AGI Definition That Was Worth Billions

If you want evidence that the definition of AGI is unsettled, the strongest is not academic. It is contractual.

For years the OpenAI and Microsoft partnership contained a clause under which Microsoft’s rights would change once OpenAI declared it had achieved AGI. Reporting by The Information and TechCrunch in December 2024 indicated the working threshold was financial, tied to roughly 100 billion dollars in profits, though OpenAI has never published that figure itself.

In its 28 October 2025 partnership announcement, OpenAI stated that “once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel”, with research IP rights lasting “until either the expert panel verifies AGI or through 2030, whichever is first”. A company had written a definitional dispute into a contract and then hired referees for it.

Six months later it was gone. The 27 April 2026 amendment says revenue share continues through 2030 “independent of OpenAI’s technology progress”. Worth being precise here, because much of the coverage was not: OpenAI never announced that it was removing an AGI clause. The trigger simply is not in the new terms. The word AGI appears eight times in the October 2025 announcement and not once in the April 2026 one that replaced it. The most commercially consequential definition of AGI ever written was not settled or met. It was quietly dropped, because it could not be made to mean anything reliable enough to hang billions on.

So How Close Are We to AGI?

The honest answer is that the people best positioned to know disagree by decades, and their incentives are not neutral.

Sam Altman told TIME in August 2026 that OpenAI expects an internal system it would call AGI by the end of the year, a claim we take apart in detail in our analysis of what Altman actually promised. Anthropic’s Dario Amodei has pointed at 2027, and DeepMind co-founder Shane Legg has long pointed at 2028. Aggregated forecasts from AI researchers outside the labs sit decades later, and prediction markets land between the two camps. We keep the full board, including the dates that have already expired, in our tracker of every AGI prediction.

Notice the pattern: the closest predictions come from the people raising capital against them, and each is using a definition of their own choosing. That does not make them wrong. It does mean a date is only meaningful attached to a bar, and almost nobody quoting these dates states the bar. If the risk side of this interests you, we cover how researchers estimate worst cases in our explainer on p(doom).

What Would Actually Have to Change

Both 2026 frameworks point at the same missing pieces, which makes them the most reliable thing to watch. Continual learning, so a system improves from experience instead of being frozen at training time. Durable long-term memory, named specifically as the deepest current deficit. Metacognition, meaning a reliable sense of its own uncertainty. And transfer into genuinely novel situations, which is what each new generation of ARC-AGI is built to test once the last one is beaten.

Our own read: the interesting question stopped being “when does AGI arrive” some time ago. Systems this jagged do not arrive, they fill in, unevenly, one faculty at a time, and the label gets applied retrospectively by whoever benefits from applying it. Watch the memory and continual-learning numbers rather than the announcements. If you want to see where each current model actually lands today, our comparison of the best AI models tracks the real capabilities rather than the promised ones.

Frequently Asked Questions