Artificial General Intelligence (AGI)

Artificial general intelligence (AGI) is a hypothetical form of artificial intelligence (AI) that matches or exceeds human ability across virtually all cognitive tasks and learns new tasks without being built or retrained for each one. By most definitions, every AI system in use today is artificial narrow intelligence (ANI), competent only within the scope it was trained for, although large language models (LLMs) that write, code, and reason across many domains have made that boundary a subject of active debate. In modern usage, strong AI is a synonym for AGI. Mark Gubrud used the term in 1997, and researchers Shane Legg and Ben Goertzel popularized it around 2002.

Defining AGI#

AGI has no single agreed definition, and the main proposals measure different things:

• Economic output: OpenAI's Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work."
• Performance and generality: A Google DeepMind team proposed Levels of AGI, which rates a system on two axes: how well it performs compared with skilled adults, from Emerging (equal to or somewhat better than an unskilled human) to Superhuman (better than all humans), and whether it performs that well on one task or across a wide range of non-physical tasks, including learning new skills.
• Learning efficiency: François Chollet's "On the Measure of Intelligence" (2019) defines intelligence as skill-acquisition efficiency: how quickly a system masters tasks it has never seen, given limited data and prior knowledge, rather than how much skill it has already accumulated.
• Mind and understanding: The philosophical question is whether a machine would genuinely understand and have a mind, the original sense of strong AI in John Searle's 1980 Chinese Room argument.

The Levels of AGI framework sorts a system with up to three questions, generality first:

The authors placed 2023 chatbots such as ChatGPT, Bard, and Llama 2 at Emerging AGI and listed every higher general level as not yet achieved, while narrow systems already span the scale up to Superhuman, where they put AlphaFold, AlphaZero, and Stockfish.

AGI vs. Narrow AI and Superintelligence#

• ANI: Excels within a defined scope and needs new data and retraining to do anything outside it. An Ultralytics YOLO26 detector trained on COCO sits in the narrow column of the Levels framework: it finds and labels 80 object classes in milliseconds per image, and it cannot learn a new class, answer a question, or plan an action without retraining or other systems around it.
• AGI: Matches skilled people across most cognitive tasks and carries what it learns into domains it was never trained on, the ability transfer learning provides today only between closely related tasks.
• Artificial superintelligence (ASI): Surpasses the best humans in virtually every domain. A system that improves its own intelligence faster than people can follow is the scenario behind the technological singularity.
• Strong AI: A synonym for AGI in modern usage. In Searle's original sense, it is a machine that literally has a mind, a claim about understanding that capability tests do not address.

Frontier LLMs strain these categories. One model can draft contracts, solve competition math, and describe images, yet its abilities are uneven, a pattern called jagged intelligence, and its weights stop learning when training ends.

Open Problems#

• Continual learning: Deployed models adapt only within their context window, and training them further on new data risks catastrophic forgetting of earlier skills.
• Reliability: Models that pass expert exams still produce confident errors, such as hallucinations, and fail unpredictably on simple variations of tasks they solve.
• The physical world: Embodied AI systems must perceive, plan, and act in real environments, and world models that predict how scenes change are one proposed route to the common sense this requires.
• Alignment and control: AI safety research, including superalignment, asks how to keep a system that may exceed its overseers pursuing the goals people intended.

Measuring Progress Toward AGI#

Because the definitions differ, so do the tests. The ARC-AGI benchmarks follow Chollet's definition with novel puzzles that people solve easily, scored on hidden test sets so that answers cannot be memorized. ARC-AGI-3, launched in March 2026, replaced static grid puzzles with hundreds of interactive game environments that come with no instructions or stated goals, so an agent must explore, infer the rules, and plan; at launch, human testers scored 100% and frontier AI systems 0.51%. In September 2026, the ARC Prize Foundation reported that OpenAI's GPT-6 Astra reached 62.7% under its standard test harness and 99.9% under a harness that keeps the model's reasoning state between moves, and stated that saturating the benchmark would not prove AGI. Knowledge-heavy tests such as Humanity's Last Exam measure expert-level question answering instead.

Forecasts of when AGI might arrive vary widely. In a survey of 2,778 researchers who had published at top AI venues, published in 2024, the aggregate forecast gave a 50% chance that unaided machines would outperform humans at every possible task by 2047, 13 years earlier than the same team's survey one year before.