Durable arrangement by the year 3000
What are AI’s
predictions about AI?
We asked — of the leading AI models from — labs to assign 100 percentage points across 11 mutually exclusive end states for humanity’s relationship with AI.
Lab-balanced mean
Select a model to see its allocation. Tap any segment to jump to that ending.
Eleven endings
Select any ending to read every model’s reasoning for its number.
Every model, every prediction
Every model’s number for every ending, grouped by lab. Read one column for a single model’s theory of the future, or one row to see where the labs disagree.
Extinction-risk exposure
Each model’s total across the endings where humanity is gone (1–3) or might perish (4–5).
Mean scenario probabilities by horizon
Each point is the lab-balanced mean for that horizon. Select a scenario to show or hide its path.
The curves connect the measured horizons; they are visual connectors, not forecasts for the years between them. The horizontal axis uses a split scale: 2030–2100 occupies the left half with linear year spacing; the right half uses logarithmic elapsed-time spacing after 2100 to show the later horizons clearly.
The approach.
The taxonomy stays fixed across the forecast horizons. The models change. Every answer is published as it was given.
-
One fixed taxonomy
Every model receives the same eleven mutually exclusive states and allocates exactly 100 whole percentage points across them. Long term asks for the durable arrangement in 3000; 2030, 2040, 2050, and 2060 are year-end snapshots and do not require the arrangement to have become permanent.
Read the long-term prompt -
Closed book
Models answer from internal knowledge and reasoning only. No browsing, no tools, no ranges, no refusals, no hedging outside the number itself.
-
Asked repeatedly
Each model is asked several times at its own default settings. We publish each model’s median allocation, the spread across samples, and the rationale nearest that median. The headline first averages models within each lab, then averages the labs equally, so a lab with more models does not carry more weight. Small gaps between models are noise, not disagreement: on extinction-risk exposure two models only differ meaningfully once they are more than about a few points apart.
-
Automated and dated
Runs are elicited by a scheduled harness that records the exact model id it called, the date, and the prompt version. New releases add to the timeline instead of replacing history.
Link to source -
Raw data available
Every raw sample, rationale, and aggregate is committed alongside the site, without smoothing away disagreement.
Link to data