What is AGI? The definition behind the timeline fights
Every definition above stops being true on a specific, published condition.
The 62.5/100 above is not an opinion — it recomputes from eight graded predictions, each with a pre-registered flip condition. When one flips, the definitions table on this page changes with it. Subscribe and you hear on the day that happens, not when a headline does.
Tell me when a verdict flips →The working definitions that matter
| Definition (bar) | What it requires | Status, mid-2026 |
|---|---|---|
| Benchmark-level capability | Match skilled humans on scoped professional tasks | Largely here (~83% GDPval, ~80% SWE-Bench Pro) |
| Drop-in remote worker | Do a real job end-to-end, unsupervised, reliably | Not yet — reliability lags benchmarks |
| Automated AI researcher (Aschenbrenner's bar) | Autonomously conduct AI research itself | Undemonstrated |
Almost every public disagreement about “when AGI” is really a disagreement about which row counts. By the first bar, something like AGI is arriving now. By the third — the one Situational Awareness uses, because it triggers an intelligence explosion — it has not arrived, and that is what the AGI-by-2027 verdict tracks.
What AGI is not
AGI is not the same as superintelligence (ASI) — AI far beyond the best humans at essentially everything. In the standard sequence, AGI is the trigger: once AI can do AI research, hundreds of thousands of automated researchers compress progress, and superintelligence follows. Nor is AGI the same as a strong chatbot: fluent conversation was passed years ago without the autonomy that defines the serious bars.
How close is AGI in 2026?
On capability, close: models sit near the top of the skilled-human range on scoped knowledge work and agentic coding. On autonomy, the defining gap remains: no system has run the full research loop — or a full job — reliably without human supervision. That split verdict is why this scorecard grades the capability prediction On track while the headline AGI-by-2027 claim stays Open, resolving by January 1, 2028.
When do forecasters expect it?
Public timelines span Musk's end-of-2026, Aschenbrenner's 2027, Hassabis's ~50% by 2030, Metaculus's 50% by 2033, Karpathy's roughly-a-decade, and the academic survey's 2047 — with expert medians compressing from ~2060 to ~2033 in about six years. The full comparison: when will AGI arrive?
Frequently asked questions
AGI (artificial general intelligence) is AI that can do essentially any cognitive task a skilled human professional can — not just chat or pass tests, but perform real work across domains. The strictest common bar is AI that can autonomously do AI research itself.
AGI matches skilled humans at general cognitive work; ASI (superintelligence) is far beyond the best humans at essentially everything. In most forecasts AGI comes first and, by automating AI research, accelerates the path to ASI.
By the loosest definition (benchmark-level capability on scoped tasks), arguably close — ~83% on GDPval-style knowledge work. By the serious bars — a reliable drop-in worker, or an automated AI researcher — no. That autonomy gap is why the AGI-by-2027 prediction is still Open.
Mostly definitions and weighting of the autonomy gap. Forecasters using capability-centric bars predict 2026–2027; those weighting reliability and autonomy land 2030–2047. Public forecasts currently span Musk (2026) to the academic survey median (2047).
There are none — no deployed system in 2026 meets a serious AGI bar. The closest candidates are frontier models scoring ~83% on GDPval-style knowledge work, which still fail the two bars that matter: working as a reliable drop-in employee, and doing autonomous AI research. Until one exists, the honest examples are hypothetical — which is why this site tracks eight dated predictions (62.5/100) instead of pointing at a product.
The live scorecard updates as models ship and verdicts change.
View the live scorecard →Be first to know when a verdict flips
One email when one of the eight verdicts changes and the AGI-2027 Thesis Tracker moves — the single auditable score no other tracker has. Not a weekly newsletter; nothing arrives until the score actually moves.
Subscribe free →