Artificial general intelligence is a hypothetical kind of AI that matches or surpasses human capabilities across virtually all cognitive tasks. The word “hypothetical” still applies in November 2026, although it is contested. Several AI leaders say AGI is a few years away, or already arriving in their labs. Many researchers doubt the current approach will get there, and the most demanding public test of general learning still shows a large gap between people and machines.

Most of the confusion comes from one fact: “AGI” has no agreed definition. Whether it has arrived, or when it will, depends on which definition you use.

What artificial general intelligence means

AGI refers to a system with broad, flexible competence rather than skill in one area. It contrasts with narrow AI, which can solve a specific problem but lacks general cognitive abilities. A chess engine or a spam filter is narrow. A system that could learn an unfamiliar job, switch between domains, and handle situations it was never prepared for would be general. 

The idea has other names, including strong AI and human-level AI. The term itself was popularized among computer scientists around 2001 by Shane Legg and Ben Goertzel, who described a machine able to do the cognitive tasks people can typically do. 

Three terms are often mixed up:

Term	What it describes	Status
Narrow AI	Systems built or trained for specific tasks	Exists everywhere
AGI	Human-level breadth and competence across cognitive work	Disputed; not broadly accepted as achieved
Artificial superintelligence	Intelligence that clearly exceeds human ability across the board	Hypothetical

AGI does not require consciousness. DeepMind researchers argue that a definition should focus on what a system can do rather than whether it thinks or understands like a human or has properties such as sentience.

Why nobody agrees on a definition

The definitions in use are not minor variations of one idea. They measure different things.

• Cognitive tasks. The Legg and Goertzel framing asks whether a machine can do what people typically do mentally. Even so, it leaves open which tasks and which people.
• Economic work. OpenAI’s charter describes highly autonomous systems that outperform humans at most economically valuable work. This is measurable in principle, but it depends on what counts as valuable work and how much of it must be covered.
• Levels. Google DeepMind proposed grading systems by performance (depth) and generality (breadth), in the way autonomous driving is graded in levels. Under this framework, today’s leading chatbots sit on the first rung of the ladder. The framework also says a system does not need a physical body to count.
• Contract language. The legal history is revealing. The AGI clause between Microsoft and OpenAI reportedly rested on a profit threshold of at least $10 billion. In October 2025 the clause was changed so that an independent expert panel would verify any OpenAI declaration of AGI. By April 2026 the arrangement was revised again, and commentators concluded that the AGI clause was effectively dead.

The practical consequence is that an AGI claim means little until you know the yardstick. A company can truthfully say it has reached “AGI” by its own definition while researchers using stricter tests say it has not.

Where things stand in late 2026

What the evidence shows. No system has been broadly accepted as AGI. The picture is uneven. Frontier models now do very well on many static tests. On ARC-AGI-2, a puzzle benchmark designed to resist memorization, Gemini 3.1 Pro reached 77.1% roughly a year after the benchmark appeared. The weakness shows up when a task requires learning something new while doing it.

ARC-AGI-3 launched on March 25, 2026 and drops an agent into unfamiliar interactive environments with no instructions. At launch, Gemini scored 0.37%, GPT-5.4 scored 0.26%, and Claude scored 0.25%. Ordinary people solved all 135 environments. These were launch results, and leaderboards change quickly, so check the current standings before relying on any number. Still, the pattern matters: strong performance on familiar kinds of problems did not carry over to novel, interactive ones. Benchmark creator François Chollet has also argued that today’s models perform well mainly when humans build elaborate scaffolding around them.

What the labs say. Sam Altman told TIME on August 26 that OpenAI was not yet at AGI but could have an internal system he would call AGI by the end of 2026. That is a target set against his own definition, not an independently validated result, and an internal system is one outsiders cannot test.

Demis Hassabis told the Wall Street Journal in October that he now expects AGI within two to three years, down from three to five years earlier in 2026. The Journal’s profile described his bar through an “Einstein test,” asking whether a system trained only on knowledge up to 1901 could produce Einstein’s 1905 papers. That is a much harder standard than many companies use, which means his “two to three years” and another leader’s “this year” are not predictions of the same event.

When might AGI arrive?

Forecasts span decades, and they cluster by who is making them.

Source	Estimate	Worth knowing
Large survey of AI researchers (2023)	50% chance of “high-level machine intelligence” by 2047	Down from 2060 in the 2022 version of the survey
Metaculus forecasters (Feb 2026)	25% by 2029, 50% by 2033	Aggregated track-record forecasters, but definition-dependent
Frontier lab leaders (2026)	Roughly 2026 to 2029	Best visibility into upcoming models, but incentives to sound confident
Longitudinal Expert AI Panel (2026)	Panelists’ median year for agreeing AGI exists: 2050	Scientists, more conservative

Two patterns stand out. First, nearly every group has moved its dates earlier over time. Second, the people building the systems give the shortest timelines. A review of expert forecasts notes that company leaders have the best visibility into the next generation of AI and have been right more often recently, but also face selection bias and incentives to hype. Both halves of that sentence matter.

There is a separate technical debate behind the dates. In a 2025 survey by the Association for the Advancement of Artificial Intelligence, 76% of respondents said scaling up current AI approaches is unlikely or very unlikely to yield AGI. Read this carefully. It is not a claim that AGI is impossible; it questions whether more of the same methods will be enough. Respondents were also mostly academics, a group that may weigh these questions differently from industry labs. Yann LeCun, among the best-known sceptics, continues to say superintelligence is not close.

What would still need to work

The ARC results and the research debate point to a similar short list of gaps. These are a synthesis of the discussion rather than an official checklist:

• Learning on the job. Humans pick up a new game or workflow within minutes. Current systems mostly rely on what was learned in training plus whatever fits in a prompt.
• Exploration under uncertainty. Acting to discover the rules, not just generating plausible answers.
• Reliability over long tasks. A system that is brilliant for five minutes but drifts over five hours is not a general worker.
• Independence from scaffolding. If strong results depend on custom tooling built by people, the generality belongs partly to the people.

Why it matters: benefits and risks

If AGI arrived under a demanding definition, the implications would be large. DeepMind’s framework itself notes the potential for widespread labor substitution, along with large economic effects. Researchers also see an upside in science: Hassabis frames accelerated discovery as the nearest-term case. 

The risks are contested but taken seriously. Hassabis warned in October that the industry is heading toward a serious security incident within roughly two years unless governments and companies raise safety standards. In a survey of researchers, a sizable minority gave at least a 10% chance to consequences as grave as human extinction. Those figures reflect how worried experts say they are, not a measured probability, and they remain disputed.

How to judge the next AGI claim

Headlines will keep coming. Five questions separate substance from marketing:

1. Which definition is being used? Economic output, human-level breadth, or a specific test?
2. Who verified it? A company’s own statement is different from independent evaluation.
3. Can outsiders test it? An internal system cannot be checked.
4. Does it hold up on new problems without special scaffolding? This is where current systems struggle.
5. What are the performance and generality levels? “Expert on some tasks” is not “better than most people on nearly everything.”

Frequently asked questions:

Is ChatGPT, or any current chatbot, AGI?

Not by most definitions. DeepMind’s 2023 framework treated leading chatbots as an early stage on the path, not as the destination, and benchmarks that test learning from scratch still show large gaps.

Is AGI the same as superintelligence?

No. AGI is roughly human-level breadth. Superintelligence would go clearly beyond it. Some forecasters expect the second to follow the first quickly, but that is a prediction, not a fact.

Has anyone declared AGI?

Executives have made statements and set targets, but none has been independently verified. The Microsoft and OpenAI expert-panel mechanism introduced in 2025 was later made moot when the contractual AGI trigger was removed, so no formal referee currently exists.

Does AGI need a robot body?

Not under DeepMind’s framework, which focuses on cognitive capability. Other researchers disagree.

The bottom line

AGI is best understood as a moving target with several yardsticks rather than a single finish line. By the strictest tests, no system is close; by looser, work-focused definitions, some leaders say it is months or a few years away. The reasonable stance is to ask what each claim measures, who checked it, and whether the system can learn something it has never seen. Those questions will stay useful whichever forecast turns out to be right. If you’re eager to learn more about technology, visit softscr.com.