Quiescence · open research
How much human life is in the things you buy?
HumanToll estimates the deaths and injuries embodied in a product, a material, or €1,000 of spending — and shows the whole chain behind the number.
Everything is made by an economy that costs people their lives: in armed conflict, at work, and through the water, food and air that economic activity provides or withholds. This project traces that harm through global supply chains and attributes it to what is bought. Deaths and injuries are both published, always with their uncertainty, and every figure can be walked back to the conflicts, materials, countries and datasets behind it.
Two headline numbers, never one
A death and an injury are different harms, and both are counted. Nothing is ever published as a bare figure.
Lives
Deaths attributed to the supply chain of a product, a material, or an amount of spending.
Injuries
Non-fatal disabling injury and illness, published with equal prominence — a co-equal metric, not a footnote.
Life-years
Years of life lost and years lived with disability, so that a death at nineteen and a death at seventy-one are not the same entry.
Denominators that mean something
Per unit, per kilogram and per €1,000 — and inverted, because "one death per N units" is a fact a person can hold and a decimal with six leading zeros is not.
Four kinds of harm, not just war
Conflict is where this started. It is not where most of the harm is.
-
Conflict and political violence
Battle deaths, one-sided violence, and the famine and disease that follow them. Built on UCDP, with injuries derived from documented casualty ratios rather than assumed.
-
Work
Fatal and non-fatal occupational injury and disease, by industry. This is the one channel where injuries are directly measured rather than inferred, and it needs no attribution judgement at all.
-
Deprivation
Deaths attributable to undernutrition and to unsafe water, sanitation and hygiene — harm from what an economy fails to provide.
-
Exposure
Airborne particulates, occupational carcinogens, and heavy metals from extraction and processing.
What “excluding natural deaths” actually means here
A death enters the ledger only if it carries a population attributable fraction against a modifiable exposure — something economic activity produces, distributes or withholds. Everything with no such exposure is excluded by construction. Nobody has to decide what counts as natural, because the absence of an exposure–response relationship decides it.
One consequence is worth stating plainly: harm that cannot be attributed is dropped. Every figure this project publishes is a floor, never a total.
How a number is built
Six steps, each of which can be inspected, disagreed with, and traced to a checksummed file.
-
Count the harm
From primary sources, with the uncertainty the publisher reports, and reconciled across the four channels so that a miner killed in an armed attack is not counted twice.
-
Attribute it to economic activity
Each subject gets a share across economic sectors, scored against nine documented criteria with an evidence link per score — never asserted, and never a single number nobody can question.
-
Attribute it to materials
Each sector’s share is split across the raw materials it actually extracts and consumes, from physical production and trade data rather than judgement.
-
Propagate it through the supply chain
A multi-region input–output model carries the harm from the mine to the component to the product, including every loop back through the industries that supply each other.
-
Divide by what was made
Expressed per unit, per kilogram and per €1,000 of spending, so a croissant and a car can be compared at all.
-
Carry the uncertainty all the way
Every stage is sampled thousands of times, and what gets published is the interval and a confidence grade — not the midpoint on its own.
Production today. Use next.
The model currently answers one question: how many people were harmed making this. A second half — harm from using a product, which is road deaths, combustion, firearms, tobacco and alcohol — is specified and will be published as a separate figure alongside the first. Until it ships, every number here says production phase only, because for something like a car the use half is very likely the larger one and implying otherwise would be misleading in a specific direction.
What this is not
- Not a causal claim about any individual death. Attribution is statistical and applies to populations.
- Not a verdict on a purchase. A high figure describes a supply chain, not the person who bought from it — the lever it points at is procurement, regulation and sourcing.
- Not comprehensive. Harm with no attributable pathway is dropped, so every figure is a lower bound.
- Not precision. A figure printed without its range would be a lie of format.
- Not a boycott list. The comparison tool exists to show why two products differ, and the answer is usually one material and one country — which is a fixable fact.
Questions
Is this saying my shopping kills people?
No. It measures harm statistically attributed to the supply chains that produce goods, under a stated set of assumptions. What follows from that is a question for procurement, regulation and sourcing — not for guilt.
Are all wars economic?
The default scenario assumes that organised violence is embedded in economic activity, and quantifies the economic activity associated with conflict rather than the motivation of anyone fighting it. A conflict that began in religious hatred still buys rifles, burns fuel, seizes land and sells what it digs up. Four alternative scenarios ship alongside, including a conservative one that counts only high-confidence links.
Where does the data come from?
Open, citable sources: UCDP for conflict, ILOSTAT and Eurostat for work, FAOSTAT and WHO/UNICEF for deprivation, Eurostat FIGARO and EXIOBASE for supply chains, and the World Bank for prices and population. Every source, its licence and the exact release used are listed and cited.
Can I check a number?
That is the point. Every figure links to the conflicts, materials, countries and datasets behind it, every model run records a manifest of checksummed inputs, and a published result is never edited — a correction is a new run and the old one stays readable.
When will there be numbers?
When the backtest passes. The model has to produce sixteen years of results that move with events that actually happened and change only for reasons that can be written down. Until then this page publishes the method.