AGI — artificial general intelligence — is the term for an AI that can do anything a human can do, across any domain, without being specifically retrained for each task. It's the threshold that most of the AI industry is now explicitly racing toward. And according to the median expert forecast, it's roughly five years away.
The word "general" is doing all the work in AGI. Current AI systems — GPT-4o, Claude, Gemini — are extraordinarily capable, but they're narrow. Each one was trained on a specific distribution of data to perform well on a specific range of tasks. Push them outside that range and they fail in ways a human wouldn't. Ask them to autonomously learn an entirely new skill without retraining, and they can't.
AGI removes that constraint. An AGI system would be able to learn any new domain a human can learn, perform any cognitive task a human can perform, and do it across the board — not in one specialty, but in all of them simultaneously. The AGI that can write code can also conduct a clinical trial, argue a court case, or compose a film score, switching between domains as easily as a person who spent years learning each one.
What makes this genuinely hard to build isn't raw intelligence in any one area — current AI already exceeds the best humans in specific tasks like chess, protein structure prediction, and certain coding benchmarks. The challenge is the generality: reliable cross-domain performance, autonomous learning, and robust reasoning in entirely novel situations the system has never seen before.
| Capability | Current AI (2026) | AGI |
|---|---|---|
| Domain coverage | Broad but brittle — fails outside training distribution | Any cognitive domain a human can work in |
| Learning new skills | Requires retraining or fine-tuning | Learns autonomously, like a human does |
| Persistent goals | Session-limited; no persistent autonomous intent | Can pursue multi-step goals over time |
| Novel situations | Degrades significantly on truly novel inputs | Handles novelty robustly across domains |
| Professional tasks | Superhuman in narrow tasks; unreliable across broad roles | Human-level or better across all professional domains |
| Cost structure | Cheap per query; still requires significant human oversight | Replaces human-level cognitive labor almost entirely |
The gap between the two columns is what makes the timeline so consequential. Closing it doesn't mean making current AI slightly better. It means crossing a qualitative threshold that changes everything about how cognitive work gets done.
The honest picture: 2026 AI systems are doing things that would have qualified as AGI-level predictions as recently as 2019. PhD-level science benchmarks — once considered a marker of human-expert capability — have been cleared. Top coding competitions, legal reasoning tests, and medical licensing exams all have AI systems scoring at or above the professional human average.
But passing exams isn't AGI. Current systems still fail the generality test. They can't reliably take autonomous multi-step action in novel environments without human checkpoints. They don't accumulate knowledge across sessions in the way a person builds expertise over a career. And they remain brittle in ways that matter for real deployment — confident in wrong answers, inconsistent across runs, unable to recognize the edges of their own competence.
The progress is real. The threshold hasn't been crossed. What the live tracker shows is that the forecast has moved dramatically — from 2060 to 2031 in just four years — driven by benchmark scores that no serious researcher expected to fall this fast. That shift is documented in detail here.
No — and OpenAI, which builds ChatGPT, says so directly. Their published AGI definition requires a system that can outperform humans on most economically valuable cognitive tasks. GPT-4o and its successors don't meet that bar yet. OpenAI has said internally that their current systems are roughly 70% of the way to their internal AGI definition, but that last 30% is the hard part — and the part that requires genuine generality rather than just scaling existing capabilities.
The question people are really asking when they ask "is ChatGPT AGI?" is usually: is this thing smart enough to replace me at work? The honest answer to that question depends heavily on what you do. For a narrowly defined, text-based knowledge role, the gap is closing fast. For work that requires physical presence, sustained novel reasoning, or autonomous multi-step execution in unpredictable environments, the gap is still meaningful. See the Will AI Replace My Job? breakdown for a profession-by-profession analysis.
There is no single agreed definition of AGI — which matters more than it sounds, because different definitions lead to very different conclusions about how close we are. The main camps:
The task-completion definition (OpenAI's working version): AGI is a system that can outperform humans on most economically valuable cognitive tasks. This is concrete and measurable, which is why it's widely used. It's also relatively achievable: you can test whether something outperforms a median human professional on a representative range of tasks.
The autonomy definition: AGI requires a system that can set and pursue its own goals over extended time horizons, learn new skills without human input, and operate without meaningful human oversight. This is a harder bar. No current system comes close.
The consciousness definition: some researchers argue AGI requires something like genuine understanding or consciousness — not just behavioral competence. This bar is almost impossible to measure and is not widely used in the forecasting community.
The tracker at howcloseisagi.com uses the task-completion definition — the most measurable and most widely cited version — when aggregating forecasts. It's worth knowing which definition someone is using when you read any AGI timeline claim.
AGI is not the end of the road. ASI — artificial superintelligence — is the threshold beyond it: a system that doesn't just match human intelligence across all domains, but significantly exceeds the best human performance across all of them. An ASI doesn't just do what a genius human does faster. It reasons in frameworks that don't have human analogues, compresses research timelines from decades to days, and finds solutions in complex systems that no human team could reach.
Most researchers who take the AGI timeline seriously also believe ASI could follow AGI relatively quickly — potentially within years rather than decades, due to recursive self-improvement. The full AGI vs ASI breakdown is here if you want to go deeper on what separates the two thresholds and what the best scenario models say about the gap.
The reason AGI gets so much attention isn't philosophical — it's practical. Once a system can perform any cognitive task at human level, the economics of knowledge work change entirely. Any task that involves reading, writing, reasoning, coding, designing, diagnosing, advising, or planning becomes automatable — not just assisted, but automatable at near-zero marginal cost.
This is why the professions analysis matters. The gap between "current AI" and AGI isn't evenly distributed. Some professions are already effectively AGI-exposed — the tasks that matter most in those roles can already be done well by current systems. Others have a meaningful buffer. But at AGI, that buffer essentially disappears across the board.
It's also why the builder perspective has shifted. If you believe AGI is five years away — not fifty — the rational response to building a company is completely different. You optimize for speed, for AI-native architecture, for the leverage that comes from having the right stack before the threshold hits. That's the thesis behind the one-person unicorn framing — and why it only makes sense if AGI is actually close.
AGI stands for artificial general intelligence — an AI system capable of performing any cognitive task a human can perform, across all domains, at human level or better, without task-specific retraining.
Current AI is narrow — excellent within its training distribution, brittle outside it. AGI is general — it performs reliably across all cognitive domains, learns new skills autonomously, and doesn't require retraining for each new task. The difference is qualitative, not just quantitative. Think of current AI as an extraordinarily fast specialist. AGI is the generalist who can become a specialist in anything.
The median expert forecast as of September 2026 puts AGI arrival around 2031 — roughly five years. That figure has moved dramatically: it was 2060 as recently as 2019. The pace of benchmark progress is the main driver. Some teams at OpenAI believe it could arrive as early as 2027. Full analysis of how the forecasts moved here.
AGI would make essentially all cognitive work automatable at near-zero marginal cost. Which jobs are exposed first depends on how far current AI already goes in those roles — some professions are already heavily affected, others have meaningful buffers that close at AGI. The full profession-by-profession breakdown is here.
ASI — artificial superintelligence. A system that doesn't just match human intelligence across all domains, but significantly exceeds it. Most forecasters who take AGI seriously expect ASI to follow within years, not decades, due to recursive self-improvement dynamics. The full AGI vs ASI breakdown is here.