What does AGI actually mean?
The term "artificial general intelligence" is used to describe an AI system capable of learning and performing any intellectual task that a human being can. The key word is general. Current AI systems — including the large language models behind tools like ChatGPT and Gemini — are narrow. They are exceptional at specific tasks (generating text, translating languages, identifying patterns in images) but cannot transfer that ability to an entirely new domain without extensive retraining.
A plausible AGI would be different. Ask it to write a poem, then ask it to debug software, then ask it to plan a business strategy — and it would handle all three, learning from each experience, the way a human generalist does. It would apply reasoning flexibly across contexts rather than pattern-matching within a single trained domain.
The definition, however, is contested. Some researchers draw a sharp line: AGI requires human-level performance across all cognitive tasks. Others use a looser framing — a system that can perform most economically valuable tasks better than most humans. OpenAI has reportedly adopted a version of the latter. The differences matter enormously, because the bar you set determines how close you think we are.
Where are we today?
Current large language models demonstrate capabilities that would have seemed impossible a decade ago. They can pass professional exams, write coherent code, explain complex concepts, and engage in sophisticated reasoning about abstract topics. This has led some prominent figures — most notably Sam Altman and Demis Hassabis — to suggest AGI could arrive within years rather than decades.
The sceptics push back on several grounds. First, impressive performance on benchmarks does not necessarily reflect genuine understanding. LLMs remain brittle in specific ways: they hallucinate confidently, struggle with novel logic puzzles that fall outside their training distribution, and lack any persistent memory or model of the world that updates in real time.
Second, general intelligence likely involves more than language. Embodiment, sensorimotor grounding, causal reasoning, and social understanding may be foundational to intelligence in ways that purely text-based models cannot replicate. A system that can write fluently about riding a bicycle has no functional model of what balance, momentum, and muscle coordination actually feel like.
Narrow AI
Excels at one specific task. Cannot generalise. Every current commercial AI system.
AGI
Human-level performance across all or most cognitive domains. Does not yet exist.
ASI
Artificial Superintelligence: hypothetical AI that far surpasses all human intelligence. Theoretical.
Why the debate matters
The disagreement about AGI is not merely academic. If transformative AGI is decades away, the appropriate policy response is different from a scenario where it arrives in five years. Training compute, safety research timelines, regulatory frameworks, and investment decisions all depend on these estimates.
The alignment problem — how to ensure a highly capable AI system pursues goals aligned with human values rather than misaligned ones — becomes vastly more urgent as systems become more capable. A narrow AI that generates misleading text is a nuisance. A general-purpose AI operating autonomously in high-stakes environments is a different category of concern entirely.
It is worth noting that AI researchers themselves are divided. A 2022 survey by AI Impacts found that the median expert estimated a 50% chance of human-level machine intelligence by around 2059, but with enormous variance — some experts estimated it could happen within a decade, others said it was over a century away, and a significant minority considered it fundamentally impossible.
The 60-second version
- AGI means AI that can do any intellectual task a human can — not just one specific task.
- We do not have AGI. Current AI systems are powerful but narrow and brittle in specific ways.
- Experts sharply disagree on whether AGI is years, decades, or centuries away — or possible at all.
- The definition itself is contested: what counts as "general" is genuinely unclear.
- The debate matters because safety, regulation, and policy timelines depend on how close we think we are.
What to read next
If you want to go deeper, the most grounded technical discussion remains Stuart Russell's Human Compatible (2019), which argues that the alignment problem is the defining challenge of our era. For a more sceptical perspective, Gary Marcus has written extensively about the limitations of current LLM-based approaches. And for a wide survey of expert opinion, the 2022 AI Impacts survey is available freely online.
The honest answer about AGI is that nobody knows — and anyone who speaks with certainty in either direction is probably overconfident. The appropriate response is to take both the possibility and its implications seriously, while remaining clear-eyed about what current systems can and cannot do.