Anthropic Shadow Catastrophe Bias Calculator

What this calculator is actually estimating

Anthropic shadow is a selection effect. If a class of catastrophes sometimes wipes out all observers, then the histories in which those worst outcomes occurred are precisely the histories that leave nobody behind to write them down. That means the catastrophe record seen by surviving observers can look deceptively quiet. A low observed count does not automatically mean the world was objectively safe; it can also mean many severe branches ended observation altogether. This calculator gives you a simple way to translate that idea into numbers.

To do that, the model asks for four inputs. First is the number of observed catastrophes, written as n. Second is the historical observation window, written as T, in years. Third is the survival probability per event, written as s, which is the chance that a catastrophe of the type you are modeling still leaves observers around to notice and record it. Fourth is the future horizon, written as H, again in years. From those values the calculator estimates a corrected underlying event rate and the chance of at least one lethal event in the future horizon.

The key intuition is simple. If only half of the relevant events leave survivors, then the raw count that survivors observe should be roughly half the true number of events. In that setting, every observed catastrophe points to a larger hidden base rate. This is why the page reports a shadow factor of 1 divided by s. A survival probability of 0.5 implies a shadow factor of 2. A survival probability of 0.2 implies a shadow factor of 5. The lower survival becomes, the larger the gap between the visible record and the true rate the model infers.

How to choose each input in a way that makes sense

The most important input is the observed count n. Enter the number of catastrophes of the relevant type that were actually seen during the observation window. The word observed matters. Do not enter the true number you suspect exists in reality; that is exactly what the calculator is trying to infer. If your source data include ambiguous incidents, it is often better to run more than one scenario, such as a strict count and a generous count, rather than pretending that a debatable event classification is perfectly known.

The observation window T should match that count. If you observed one catastrophe across 100 years, then the observation window is 100 years. If you observed three catastrophes across 250 years, use 250. This model treats the rate as an average per year across the whole interval, so the years field is doing real work. A mismatch between the event count and the period it came from is one of the easiest ways to get a misleading answer.

The survival probability per event s is where the anthropic correction enters. This value must be greater than 0 and at most 1. It represents the fraction of events of this class that still allow observers to exist afterward. A value near 1 means most events are survivable and the observational bias is small. A value much lower than 1 means many events remove observers, which makes the visible record far more selective. In practical use, this is often the hardest quantity to estimate, so it is worth testing a range instead of relying on a single point estimate.

The future horizon H is simply the period over which you want to evaluate forward risk. If you care about the next 50 years, enter 50. If you want a century-scale estimate, enter 100. The resulting extinction probability is not the chance of every bad outcome. It is the model's estimate of the chance that at least one lethal event of the type implied by the corrected rate occurs during that future interval.

If you are uncertain about any input, a good workflow is to run three cases. Start with a cautious survival estimate, then a middle estimate, then a more optimistic estimate. That small sensitivity exercise often tells you more than a single neat-looking output. In anthropic-shadow problems, uncertainty about survival probability usually dominates the result.

Formula used by the calculator

The correction begins with an estimated true event rate, usually written as the Greek letter lambda. The observed count is divided by both the survival probability and the observation window. That step inflates the visible count to reflect the branches that likely produced no surviving observers.

Once the corrected rate is estimated, the model focuses on the lethal share of those events, which is the part multiplied by 1 - s. Assuming a constant-rate Poisson process, the probability of at least one lethal event over the future horizon H is:

Those are the domain-specific formulas the button uses. Under the hood, they still fit the same general calculator pattern used in many modeling tools: gather inputs, define a function, and turn the result into a readable estimate. The broader mathematical framing below is preserved because it is still a useful way to think about calculators in general.

The calculator's result R can be represented as a function of the inputs x₁ … x_n:

A very common special case is a total that sums contributions from multiple components, sometimes after scaling each component by a factor:

In this specific calculator, the scaling factor that matters most is the survival probability. Because it appears in the denominator, the estimate becomes very sensitive when survival is low. That sensitivity is not a bug. It is the core claim of anthropic shadow: records taken only from surviving observer branches can dramatically understate the frequency of catastrophes overall.

Worked example with the default inputs

Suppose you keep the default values already loaded in the form: 1 observed catastrophe over 100 years, survival probability 0.5 per event, and a future horizon of 100 years. The shadow factor is 1 ÷ 0.5 = 2. So the model says that the observed record should be scaled up by a factor of two to estimate the true underlying event process.

Next, compute the corrected rate. Using λ = n ÷ (sT), you get 1 ÷ (0.5 × 100) = 0.02 events per year. Interpreted plainly, that means the model estimates one relevant catastrophe every 50 years on average, even though surviving observers only recorded one over the whole century. The difference exists because only half the events are assumed to leave observers alive to count them.

Now isolate the lethal component. If survival is 0.5, then the lethal share is also 0.5. Multiply the corrected rate by that lethal share and the 100-year horizon: 0.02 × 0.5 × 100 = 1. Under the Poisson assumption, the probability of at least one lethal event over the next century is therefore 1 - e^-1, which is about 63.21%. That result is intentionally stark. It shows how quickly risk can rise once the visible record is corrected for observer selection.

The example also shows how each output should be read. The shadow factor tells you how much the observed record may be understating the true event process. The true rate converts that correction into a usable annual hazard estimate. The extinction probability then maps that rate into a forward-looking question about the next H years. If you only remember one interpretation rule, remember this: the calculator is not saying that the past was necessarily calm just because the historical count was small.

How sensitive the result is to survival probability

The table below keeps the observed count at 1 and the observation window and future horizon at 100 years. Only the survival probability changes. This is usually the most revealing scenario test because the shadow effect grows quickly as survival falls.

This is why survival probability deserves careful thought. Even when the observed count stays fixed, changing s can transform the interpretation from mild correction to dramatic undercounting. If your use case depends heavily on a guessed value for survival, report a range of answers rather than a single headline number.

How to interpret the result panel after you click Estimate Bias

The first number in the result panel is the shadow factor. This is simply 1 ÷ s. It tells you how many times larger the hidden event process may be than the observed one. The second number is the corrected rate in events per year. Because it is shown in scientific notation, very small annual rates remain readable. The third number is the future extinction probability over the horizon you entered. That value is often the most intuitive summary for decision-making because it answers a direct forward-looking question.

When checking whether an output is reasonable, use three quick tests. First, confirm that the direction makes sense: lowering survival should raise both the shadow factor and the true rate. Second, confirm that the units make sense: the rate is per year because both the observation window and horizon are entered in years. Third, run a nearby scenario and see if the result moves smoothly. If tiny changes in a shaky survival estimate cause huge swings in the answer, that is a sign to communicate uncertainty explicitly rather than hiding it.

This calculator is most useful as a structured thought tool, not as a proof machine. It helps you separate what you observed from what your observation process may have filtered out. That is exactly the kind of distinction that gets lost in informal discussions of existential risk, rare disasters, and survivor-biased historical records.

Assumptions, limitations, and edge cases

The model assumes a constant catastrophe rate through time. In reality, many risks change with technology, population, surveillance quality, and environment. The formula also assumes that the survival probability per event is stable across the event class you care about. Real catastrophes may have a wide severity distribution, meaning some are nearly always survivable while others are almost never survivable. If that variation is large, a single survival probability is only a shorthand.

The extinction probability is based on a Poisson-style independence assumption. That means events are treated as arriving randomly with a stable average rate, rather than in tightly linked cascades. If catastrophes cluster or one event changes the chance of another, this simple expression can misstate the forward risk. The tool is still useful for first-pass scenario analysis, but it is not a full dynamic risk simulator.

There is also an important zero-count limitation. If you enter zero observed catastrophes, the current formula returns a zero true rate and therefore a zero future lethal-event probability. Mathematically that is consistent with the narrow estimator being used here, but it can be too optimistic in real inference problems. If the historical count is zero, many analysts would prefer a Bayesian prior or a confidence-interval approach rather than treating the true rate as exactly zero. In other words, absence of observed catastrophes does not necessarily imply absence of hazard.

Finally, remember that anthropic-shadow reasoning is sensitive to how you define the event class. If you mix together events with very different severities or different observer-survival consequences, the survival parameter becomes hard to interpret. The cleanest use case is a clearly defined catastrophe type, a known observation window, and a survival probability estimate that you can defend or at least bracket with sensitivity analysis.