Cobra Effect

The story goes that in colonial India the government wanted to wipe out cobras, so it paid a bounty for every dead snake handed in. Citizens quickly worked something out: the corpses did not have to come from the wild, and breeding snakes yourself was more reliable. When the government discovered the bounty was having no effect and cancelled it, the breeders' snakes were worthless, so they let them go. The result: more snakes than at the start. The German economist Horst Siebert used it as a book title in 2001, and the "cobra effect" has been shorthand for perverse incentives ever since.

Put the honesty boundary up front: this story has no reliable historical source. Even the place name "Delhi" was added by later hands, and Siebert's original book only tells a general bounty legend from British India. The mechanism holds, the story is doubtful. The version with an archive behind it happened in Hanoi, and comes below.

The archived version has more absurd detail, and is more credible. In 1902 the government of French Indochina, guarding against plague, went after the rats in Hanoi's sewers. At first it paid per whole rat, then found that troublesome and switched to paying a bounty on rat tails, one cent each: to cut off a tail you first need a dead rat, which sounds unobjectionable.

Before long, sanitation officials started seeing a strange animal around the city: large numbers of living rats with no tails. The trappers had worked out exactly what the snake breeders in the legend worked out. Tails convert to money, and the rat itself is the capital that produces tails. Cut the tail, release the rat, let it keep breeding to produce the next generation of tails. Dedicated rat farms even appeared in the suburbs. The scheme was abandoned. The historian Michael Vann found the case files marked "destruction of harmful animals: rats" in the French overseas archives in 1995. Every detail above is backed by colonial archival records. He published the account in 2003 and expanded it into a book in 2018. This is the historical case with the highest evidential grade in the whole family of measurement failures.

You can point precisely at where the design failed. The bounty was meant to put a price on "dead rats." What it actually priced was "evidence that a rat died": the tail. Evidence and outcome had been bundled together by one single act, killing the rat: kill the rat and both became true at once. The moment the bounty appeared, that bundle sprang a seam. Hand over a tail and the money was paid, no questions asked about whether the outcome came with it. Cutting the tail and releasing the rat is the shortest path to peeling the evidence off the outcome. Note one detail: the measure was never faked at any point. Every tail was real. Yet the goal was precisely bypassed, and the population under bounty acquired an institutional reason to keep breeding.

Why did a fabricated story circulate for over twenty years without anyone checking? Siebert was then president of the Kiel Institute for the World Economy and a member of the German Council of Economic Experts, and his book was subtitled how to avoid the wrong turns of economic policy. What he needed was an opening parable that policymakers would never forget, not a piece of historical evidence. That is exactly the selection pressure cautionary tales face: the cleaner the story and the prettier the reversal, the more likely it gets picked for a book title, a speech or a slide, and the transmission chain does not grade whether an archive exists. The combination of real mechanism and fictional case travels freely on that basis, until someone goes back to the archive to reconcile the books.

This class of failure now has its own name in the formal taxonomy. Manheim and Garrabrant in 2018 named the situation where "an agent applies pressure to a measure in a non-causal way" the non-causal cobra effect: the rat farmer's response to the measure is genuine, neither the corpses nor the tails are faked. This is different from adversarial manipulation that tampers with the measurement directly, where the problem is not that the measure itself was touched. It lies in the causal structure: the causal response to the measure and the causal response to the goal part ways. A folk parable is thereby promoted into a taxonomic term. The story is a parable, the mechanism is a theorem, and the term is the bridge. The full map of the four failure mechanisms is in B08 (the four Goodhart types).

The boundary conditions are equally well documented: bounties and piece rates are not inherently perverse. When the task is one dimensional and the output can be verified directly, piece-rate pay (payment per unit produced) works. Lazear's 2000 study of Safelite windscreen fitters is the standard evidence: after the switch to piece rates, output rose substantially and quality did not collapse, because one fitted windscreen is one fitted windscreen and there is no seam to prise open between evidence and outcome. The condition for perverse incentives is proxy distance: the further the bounty hangs from the outcome you actually want and the more easily an intermediate step can be produced on its own, the faster the backfire arrives. The question is never "are bounties right," it is which link of the causal chain the bounty is hung on (the full checklist for benign indicators is in B14). Reading this boundary backwards and concluding "all incentives are harmful" is the same class of error as reading Campbell's law (B02: indicators corrupt the processes they measure) as opposition to measurement.

Finally, set Hanoi next to its neighbours in the family and separate three diseases. Campbell's clearance rate is manipulation at the measurement end, where false confessions contaminate the statistic itself (B02). Kerr's folly is the reward designer hanging the target in the wrong place, rewarding something that was never what was wanted (B06). The cobra effect is a third thing: the target is not hung outrageously (a rat tail really is reasonable evidence of a killed rat) and the measure is not faked, but the failure lies in the bounty itself opening a profitable business model for manufacturing the problem. The three often get told as one, but the diagnosis and the prescription differ: the first needs auditing, the second needs the target rehung, the third needs an inspection of the seams in the causal chain.

One open question: nobody has checked the German original of Siebert's book to see whether he gave a source for the Delhi story and if so what it was; who first attached the place name "Delhi" is also still an unowned mystery. A field whose business is verification has not finished verifying the transmission history of its own signature story. Of the cases you have cited, how many have you traced to a primary archive? And are the ones you did trace also the ones that hold up best when an opponent checks them?

The one-line takeaway: a bounty does not buy the outcome, it buys evidence of the outcome; the moment the evidence can be manufactured on its own, the bounty starts paying for the problem to be manufactured.