Other ·
How the Human Extinction Date Is Calculated: Method, Formula, and Limits
A transparent walkthrough of the signal pipeline, heuristic risk index, date mapping, model role, limitations, and calibration roadmap.
The date on Human Extinction Date is a heuristic risk indicator expressed as a calendar date. It is not a scientifically established extinction date, a probability distribution, an expected date of extinction, or a statement of scientific consensus. The precision of the countdown belongs to the display; it does not imply equivalent precision in the evidence.
This page documents methodology version 2026-08-14 as implemented in the site's forecasting pipeline. The aim is to make the transformation from public signals to the displayed date inspectable, including the arbitrary choices and known weaknesses.
What the system observes
The production scheduler is configured to run at 06:00 and 18:00 in Europe/Brussels. Each run loads the enabled entries in the public source configuration, collects new records, and deduplicates them by a source-derived identifier before storing them. The forecast uses signals whose recorded occurrence time falls within the previous seven days.
At this method version, the configuration declares 26 enabled source entries: 24 marked daily and two marked weekly. They span 12 of the 13 risk domains supported by the schema: AI, biological, nuclear, geopolitical, space, geophysical, ecological, cyber, nanotechnology, socioeconomic, demographic, and unknown or novel risks. A thirteenth crowd-forecasting domain exists, but the configured Metaculus question list is currently empty.
The source mix is deliberately broad but uneven. It combines publisher or institutional feeds with search aggregators. Direct examples include IAEA top news, USGS earthquake feeds, and CISA cybersecurity advisories, alongside Google News searches, arXiv, and the AI Incident Database. The configured outbreak entry is a Google News query aimed at WHO and outbreak terms, not a direct WHO Disease Outbreak News feed. WHO itself says its outbreak-news list is selective rather than exhaustive, and USGS describes its summary feed as a subset of event data. Source silence therefore cannot be read as evidence that a risk is absent.
Two slower-moving references, the SIPRI Yearbook and the IPCC Sixth Assessment Report synthesis, are configured as context-only baselines and are not polled as daily signals. The IPCC uses calibrated confidence language for its assessed climate findings; this site's displayed confidence field does not inherit that calibration.
Step 1: headline cues become a capped milestone score
The deterministic scoring stage checks each recent signal title for configured phrases. Milestones have hand-set weights from 2 to 5. A signal can activate a milestone when its title contains one of that milestone's phrases and, where specified, belongs to the required domain.
For each activated milestone, the score is its weight multiplied by at most two distinct matched phrases. Scores are then summed and capped at 8 per run. A capped score below 2 is labelled no material risk signal, 2 to 3 is minor, and 4 to 8 is clear. These labels describe the internal rule; they are not empirical probability bands.
The matching stage reads titles, not the full underlying research or event report. It does not estimate whether an article is correct, novel, independent of another article, or causally connected to human extinction. The weights and phrases were selected as product rules; they were not fitted to a historical extinction dataset.
Step 2: three observables form a risk index
The pipeline combines three normalized quantities. Let S be the capped milestone score, D the number of active domains among the 13 supported domains, and N the number of recent signals. Each term is capped at one. The implemented formula is R = 0.50 x min(S / 8, 1) + 0.30 x min(D / 13, 1) + 0.20 x min(N / 250, 1).
This makes the formula inspectable: milestone matches carry half of the index, breadth across risk domains carries 30%, and signal volume carries 20%. It also exposes an important limitation. More coverage, syndication, or reporting can increase D or N even when the underlying hazard has not changed. Conversely, a serious development can be underweighted when it is described without a configured phrase.
Some prior-looking values exist in configuration, but they do not enter this risk-index calculation. The current deterministic score also has no subtractive term for successful prevention, stronger governance, or risk reduction. Those omissions are part of the present method, not hidden scientific assumptions.
Step 3: the index becomes a date anchor
The index is mapped linearly to an anchor year. R = 0 maps to 2120 and R = 1 maps to 2032, with intermediate values rounded to a year. A month is derived from the index and the anchor is set to the first day of that month. The endpoints are editorial product choices, not estimated physical bounds on human survival.
The pipeline then sends a compact summary, up to six recent headlines, and the deterministic anchor to a language model. The model is instructed to stay within plus or minus five years of the anchor and return a two-sentence rationale. That range is an instruction, not a programmatically enforced constraint in this method version. A higher-capability model is selected for certain high-score, multi-domain rebaselines; otherwise the smaller configured model is used.
If the model call or response parsing fails, the pipeline uses the deterministic anchor and creates a template rationale. If there are no recent signals, if the material-signal hash is unchanged, or if the daily model-token budget has been reached, the current date is held instead of being regenerated.
Every stored forecast currently receives a confidence value of 0.5. That number is a fixed placeholder. It is not a measured 50% probability, a confidence interval, or evidence of calibration.
What the countdown means and does not mean
The displayed date is best read as a compressed visualization of this particular rule set's recent signal environment. Movement can answer a narrow question: did the configured sources and headline rules produce a different index? It cannot answer the much larger question of when humans will actually go extinct.
A calendar date loses information that a defensible forecast would normally expose: a clearly resolved event, a time horizon, a probability distribution, uncertainty intervals, alternative models, and a track record against outcomes. Human extinction is also not repeatedly observable, so direct calibration of an extinction date is impossible. Shorter-horizon proxy forecasts can be resolved, but they would answer narrower questions.
This distinction matters because the underlying institutions make different kinds of claims. WHO outbreak notices report selected acute public-health events; USGS feeds report earthquakes; CISA advisories describe cyber issues; SIPRI summarizes security and weapons developments; and the IPCC assesses climate evidence with defined confidence language. Combining their publications into one product index does not transfer each institution's authority to the final date.
Known limitations
Selection bias. The configured feeds overrepresent events that are published in English, exposed through RSS or search, and described using the selected terms. Coverage is not a representative sample of all global catastrophic risk evidence.
Headline and duplication effects. Core milestone scoring uses substring matches in titles. Source-specific identifiers remove exact re-ingestion from the same source, but separately syndicated accounts of one event can remain separate signals.
Hand-set parameters. Milestone phrases and weights, the score cap, the three formula weights, and the 2032-to-2120 date mapping are design choices. Sensitivity to those choices has not yet been published.
Weak treatment of dependencies. The formula adds score, breadth, and volume. It does not model causal pathways, conditional probabilities, common causes, or interactions among nuclear, biological, AI, climate, and societal risks.
No published calibration. This version has no published hindcast, held-out evaluation, or resolved-question scoring record. Calling it calibrated would therefore be inaccurate. The NIST AI Risk Management Framework recommends performance assessment with uncertainty, benchmarks, reporting, and documentation; those are useful requirements for the next stage, not achievements claimed here.
Model-generated rationale. The language model receives only a compact summary and a handful of headlines. Its two sentences can aid readability but are not a substitute for reading the linked primary material.
A practical calibration roadmap
The next credible step is not to claim a better extinction date. It is to replace or supplement the date with a suite of resolvable forecasts tied to explicit outcomes and horizons: for example, whether a named institution changes a risk status by a fixed date, whether a specified threshold is crossed in an official dataset, or whether a forecast-market question resolves yes.
Before observing outcomes, each question should define its resolution source, close date, missing-data rule, and aggregation method. Probabilistic forecasts can then be evaluated with a proper scoring rule. Metaculus describes such a rule as a function of a prediction and an outcome, designed so honest probabilities maximize expected score.
For this site's heuristic index, the immediate evaluation work is different: version the source list and parameters; retain each input snapshot; measure duplicate and source-failure rates; run sensitivity tests over weights and date endpoints; compare headline-only matches with human-reviewed full-text labels; and publish every change. An independent review would further reduce the risk of internal blind spots.
How to inspect the current output
The site exposes the current forecast, forecast history, recent signals, and configured sources as JSON. Stored forecasts also include the identifiers of the signals supplied to that run.
Those endpoints make the data path inspectable, but transparency is not the same as validity. The honest current description is: a public, reproducible-in-principle heuristic that turns recent media and institutional signals into a provocative date visualization, with substantial uncertainty and no claim to scientific consensus.