Work

Civilizational Entropy

Type: book
Abstract: One crisis, not four In January 2026 the United States left the World Health Organization and the Paris Agreement. Binding limits on the world's two largest nuclear arsenals lapsed with no successor. The Doomsday Clock moved closer to midnight than at any time in its history. In the same weeks, capital expenditure on artificial-intelligence infrastructure passed seven hundred billion dollars for the year while the European Union postponed its obligations for high-risk systems; the Biological Weapons Convention passed fifty-one years with no way to verify compliance; and the pandemic treaty adopted in 2025 still lay unsigned. Four expert communities read these as four pieces of bad news. This book shows they are one — and that the thing they share can be measured. The quantity Civilizational entropy is not a metaphor. It is the accumulated physical, ecological and institutional disorder a civilization generates through its ordinary operation — not through its failures, but through its successes. Every act of production dissipates energy, degrades materials and carries institutions further from the condition in which they were designed to work. Three features separate it from the thermodynamic quantity it is named for. It is not monotonic: a civilization is an open system, and disorder can be dissipated by doing work — which is exactly what institutions that detect error and act on it are. It is not one number: the physical, ecological and institutional registers are incommensurable, and are reported here as a profile rather than summed into a single misleading figure. And the stock is not what matters. The margin is. That margin is this book's governing quantity. Written Λ, it is the ratio of corrective capacity to amplification load, and stability requires Λ ≥ 1. A civilization does not fail because it produces entropy; all activity does. It fails when its capacity to dissipate falls below its rate of production, and stays there long enough that the accumulation overwhelms what the system can absorb. Four hazard domains — nuclear weapons, advanced artificial intelligence, biotechnology, and climate and Earth systems — are placed under one diagnostic protocol and assessed against that condition. All four fall below it, though only one of the four can presently be given a number. Six findings Substitution is not contraction. Every correction in the historical record that worked, worked by substitution. Refrigeration grew after CFCs were replaced. Generation grew after sulphur was scrubbed. Military spending continued after the warheads came down. Not once did correction require the amplifying process itself to shrink. Climate is the first case that does — which makes it not a harder version of a problem solved four times, but a different problem with no precedent, and the confidence borrowed from those four successes misplaced. Concealment is the mechanism of catastrophe. Near misses that are disclosed produce structural reform; near misses that are hidden produce more of what caused them. Cuba, public within days, produced the hotline and a test ban within a year. Petrov, secret for a decade, produced nothing. Sverdlovsk, concealed for thirteen years, was followed by the expansion of the programme that caused it. Confirmed independently in two domains, this is the book's best-evidenced result, and its remedy — modelled on aviation's reporting regime — is the cheapest and most effective instrument examined here, and the least discussed. The constraint is use, not capacity. In three of the four domains, what binds is not the corrective capacity a civilization possesses but the fraction it actually exercises. The means of decarbonization largely exist and go undeployed, because deployment means retiring capital that is still profitable. This inverts the standard diagnosis and changes what is worth building: instruments that add nominal capacity will not move the outcome; instruments that shift cost, legitimacy and the visibility of unused capacity will. Competence is the danger. The gravest risk from artificial intelligence in this framework is not a system that behaves badly. It is a system that behaves well, is adopted because it behaves well, and leaves the adopting institution unable to detect its errors — not defeated, but atrophied through disuse. No bad output occurs at any point along the way. Shed load, do not absorb it. Correction whose cost rises with the load it contains will saturate and fail. Correction that diverts or prices load does not. Prices, liability, border adjustments and disclosure hand the corrective work to actors who have their own reason to do it — the principle of an irrigation system built at Dujiangyan around 256 BCE and running still. There are two delays, and they pull opposite ways. Delay in an amplifying loop is stabilizing, because it buys the time in which correction can act. Delay in a correcting loop is destabilizing, because drift accumulates while the response is prepared. One principle follows — slow the amplifier, speed the corrector — and it governs circuit breakers, de-alerting, epidemic response and staged deployment alike. Why it matters The danger here is not that some technology will destroy civilization in an afternoon. It is that a civilization can lose the ability to see that it is going wrong, and lose it while every indicator it consults reports success — because those indicators are its own output. Gross domestic product rises when more is extracted, more consumed, and more spent repairing damage the system itself caused. GDP is not a corrupt measure. It is an accurate measure of the wrong thing. A society whose principal index is its own throughput has no channel through which excessive throughput can register as failure. Such a failure has no dramatic signature. It wears the appearance of competence until the moment it does not, and it reverses only across generations — the time it takes to rebuild expertise and institutions once they are gone. How these claims were tested Six hypotheses and five falsifying observations were registered before a single case chapter was written. The successful cases — Montreal, smallpox, the arms reductions — were designated in advance, so the book could not select only failures. Chapter 24 then audits the book against that registration and publishes what it finds, including the losses: two hypotheses never tested, a model comparison never run, and the author's own cosmological premise refuted by four cases. That refutation produced the first of the six findings. A framework that cannot lose is not a framework, and this one is shown losing. The measurement waiting to be taken In every institution now delegating judgement to an artificial system — courts, clinics, review boards, benefits offices, electoral administration — measure how often human judgement overrides the system's output, against the institution's own error rate, and whether its practitioners can still say why they would depart. An institution whose overrides approach zero has lost the capacity to detect the machine's error, whatever its accuracy figures say. The data already exists. The measurement costs almost nothing. Nobody is taking it.
Language: English
Open Access status: green
OpenAlex ID: W7225843616
Cites: 160
Cited by:
Related to: 10