AGI vs ASI: what's the difference?
Side by side
| AGI | ASI | |
|---|---|---|
| Stands for | Artificial general intelligence | Artificial superintelligence |
| Benchmark | Roughly a capable human across most cognitive tasks | Far beyond the best humans in nearly every domain |
| DeepMind's 2026 framing | At least median human performance across a broad range of tasks | Outperforms organisations of thousands of experts working for years |
| Economic picture | Can do most remote knowledge work | Can do things no human institution can |
| Exists today? | Disputed. Some say partly | No |
| Typical forecast | Late 2020s to early 2030s (see forecasts) | Very uncertain. Lab leaders say late 2020s, others say decades |
| Main risk debate | Jobs, misuse, concentration of power | Loss of human control (the control problem) |
The ladder of machine intelligence
Google DeepMind's 2023 "Levels of AGI" paper splits the space into finer steps: emerging, competent, expert, virtuoso and superhuman. The last of these overlaps with ASI. The point is that capability comes in degrees. A system can be superhuman in coding while still behind humans in robotics.
How AGI could become ASI
The link most people point to is recursive self-improvement. AGI-level systems could work as AI researchers. Run millions of copies at machine speed, and progress that would take humans decades could be squeezed into a much shorter time. That is why labs watch for automated AI research. OpenAI said in October 2025 that it was aiming for an automated AI researcher by 2028. In July 2026 Sam Altman claimed AI had started improving itself. That claim hasn't been independently verified.
The bottlenecks are real, though. Compute and energy are finite. Many discoveries need slow physical experiments. And as DeepMind's report stresses, checking that a smarter-than-human result is correct gets harder as capability grows.
Why the distinction matters
Many policy proposals and public statements now target ASI specifically. The Future of Life Institute's 2025 statement calls for a prohibition on superintelligence, not on AI in general. Meanwhile US federal agencies now use "Super Intelligence" to mean all AI. That's a third meaning, explained in What is SI? Watch which meaning a headline uses.
Frequently asked questions
Which comes first, AGI or ASI?
AGI, by definition. Most researchers expect ASI to grow out of AGI-level systems, through scaling, new methods, recursive self-improvement or many agents working together. The open question is the gap: months, years or decades.
How long between AGI and ASI?
Nobody knows. 'Fast takeoff' views, such as the AI 2027 scenario, imagine months once AI automates its own research. Slower views expect years or decades, limited by compute, energy, real-world experiments and verification. DeepMind's 2026 report describes the change as a continuum rather than a single jump.
Is ChatGPT or Claude AGI?
There's no agreed answer. Frontier models match or beat experts on many benchmarks but still struggle with reliability, long-horizon autonomy and physical tasks. Some lab leaders say AGI is effectively here. Forecasting communities using strict definitions put it around 2030 to 2033.
Is ASI the same as the singularity?
They're related but different. ASI is a kind of system. The technological singularity is a hypothesised point where technological change becomes so fast and deep that the future is hard to predict, which is often assumed to be triggered by ASI.