The Cobra Effect
What becomes worth producing once you pay people to remove it?
A reward can create more of the problem it was meant to solve when people are paid for the count instead of the outcome.
Paying for the count can change the system
Imagine paying people for every cobra they remove from a town. At first, the number brought in rises and the policy looks successful.
Then someone notices that a cobra now has a reliable cash value. Breeding cobras becomes a business.
That's the Cobra Effect: a reward meant to reduce a problem changes behaviour in a way that creates more of it.
The question is bigger than “Will this reward the behaviour we want?” Also ask what the cheapest way to earn the reward would be without producing the outcome you want.
A simple example
A digital wallet offers cash for every verified referral. Genuine customers invite friends, while others discover that fake accounts are cheaper to manufacture than real relationships are to build.
Signups rise alongside copied identities, batches of phones used to create accounts, questionable reward payments and a growing pile of accounts that staff must check by hand. The fraud team is exhausted while legitimate customers wait longer for their money.
The campaign increased signups, but many were fake and the number no longer represented a healthy customer base.
How the same incentive breaks elsewhere
Public policy. Bounties can turn an unwanted animal, object, or offence into an asset. When an authority verifies only removal and can't see how the item came to exist, production can grow behind the measurement. This weakens when origin is easy to verify and producing the problem costs more than the reward.
Product and growth. Paying for each new account gives genuine customers one reason to invite friends and gives people looking to exploit the offer a clear price for each fake account. The dashboard counts every new account as growth, but another team has to investigate the fakes while legitimate customers wait. Those later costs belong in the decision about whether the reward worked.
Organizations. Rewarding managers for team size can make hiring feel like evidence of importance. Headcount grows, coordination becomes heavier, and the organisation creates more management work to justify the capacity it added. Managers can respond to the reward exactly as designed and still make the organisation slower.
Environmental systems. Paying per unit of waste collected can improve cleanup. Without tracking where the waste originates, it can also make waste valuable enough to import, relabel, or produce. The desired outcome is less waste in the system, while the incentive pays for more waste passing through one checkpoint.
Check the cheapest way to earn the reward
Before launching a reward, ask:
- What is the real end state we want?
- What is the cheapest way to earn the reward without creating that state?
- Which costs move to another team, user, or part of the system?
- Can we pilot, cap, reverse, or withhold the reward until the outcome survives verification?
Where possible, reward durable outcomes such as retained genuine customers. Raw signups are easy to manufacture.
Keep the related models separate
Goodhart's Law says a measure becomes less informative when people optimize it as a target. The Cobra Effect goes further: the incentive can actively create or enlarge the problem it was meant to remove. Jevons Paradox needs no gaming at all; efficiency lowers the cost of use and demand expands.
Where incentives still work
Incentives aren't inherently corrupting. Direct rewards work well when the desired behaviour is hard to fake, verification is strong, actors share the underlying goal, and side effects remain visible.
The mistake is seeing a Cobra Effect behind every response to a reward and becoming afraid to incentivise anything. Before launch, test how the reward can be gamed. After launch, track the costs that appear later or move to another team, as well as the target number.