Methodology

How these numbers are calculated

Every number on this site — the End of Time Index, the risk scores, the probabilities, the predicted date — is produced by an AI model, refreshed automatically every 12 hours, following a fixed and public calculation. This page explains exactly how, so you can judge the result for what it is: a model estimate, not a prophecy.

1 — The pipeline

From AI evaluation to a predicted date

Every 12 hours, an automated job asks an AI model (Gemini) to evaluate each of the 7 risk categories below. The model has real-time web search enabled, so its assessment is based on actual recent events — not just what it learned during training. For each category, the model returns a severity score from 0 (negligible) to 100 (imminent), a short justification, and a handful of notable events from the past 7 days that moved the needle. These 7 scores are then combined into a single weighted index, converted into a probability of civilizational collapse at several time horizons, and into an estimated year — the point at which that probability crosses 50%.

Recalculated every 12 hours. Steps 3–5 can also be re-run for free, without a new AI call, whenever the weights or model parameters change.

2 — The 7 risk categories

What each score covers

Each category's weight controls how much it contributes to the overall index — a higher weight would mean that category moves the index more per point of severity. Today, all 7 weights are equal, so the index is currently a simple average of the 7 scores. The mechanism supports giving categories different weights, and this table will be updated if that changes.

3 — From severity to probability

Turning a score into a probability

The weighted index is not itself a probability — it needs to be translated into one. This is done with a standard cumulative-probability model: if a category implies a constant annual risk rate, the probability of collapse within a given number of years is 1 − (1 − p_annual)^years. This is the same general approach used by researchers such as Toby Ord in The Precipice, cited here as a methodological reference — not as the source of this project's specific parameters, which were set independently. The predicted end date shown on the homepage is simply the year at which this cumulative probability first crosses 50% (and hence becomes more likely), assuming the current risk level stays constant — it is not a forecast of the risk level itself changing over time.

4 — Role of the AI, and its limits

What the model can and can't do

The scores come from a Gemini model, with an automatic fallback chain if a call fails or the model is temporarily unavailable — the site always shows the last successful prediction rather than blank data. Real-time web search is enabled on every attempt, so scores reflect current events. Worth being upfront about: a language model can misjudge a situation, overweight a recent headline, or be inconsistent from one run to the next. This is one model's synthesis of current events, not a scientific consensus or an aggregation of expert forecasts. The 7 categories and their weights are a deliberate simplification of a much more complex risk landscape — they are a lens, not a complete map.

5 — Data freshness

How often this updates

The index and risk scores are recalculated every 12 hours. The 7-day trend shown next to each score compares the current run to the snapshot from exactly 7 days ago at 00:00 UTC. Between two runs, the site keeps showing the last computed values — nothing is ever blank or reset while waiting for the next cycle.

See the FAQ for shorter, specific questions — like why the predicted date can change from one visit to the next, or what happens if the AI call fails.