🧪 This platform is in early beta. Features may change and you might encounter bugs. We appreciate your patience!
cobra_effect
The cobra effect describes situations where an attempt to solve a problem through incentivizing a measurable proxy not only fails but actively worsens the underlying phenomenon it was meant to address. Named after the apocryphal British colonial policy of offering bounties for dead cobras, which incentivized snake farming and increased the cobra population. Distinct from Goodhart's Law in that the cobra effect involves genuine backfire.
A hospital reduces readmission rates (a quality metric) by discharging patients less ready for discharge. The metric improves, but patients deteriorate at home and present to emergency rooms not counted in the readmission statistic. The intervention improved the metric while worsening the health outcome it was proxying.
A school district ties teacher bonuses to standardized test score improvements. Teachers respond by spending most of the year drilling test-specific strategies and excluding untested subjects like art, physical education, and critical thinking. Test scores rise modestly, but broader educational outcomes and student engagement deteriorate.
A city government offers a bounty for every rat tail turned in to reduce the rodent population. Entrepreneurs begin breeding rats specifically to harvest their tails and collect the bounty. The rat population increases rather than decreases, and the city ends up paying for a problem it created.
Binary (yes/no) questions an LLM must answer to identify this aspect:
Does the intervention optimize a measurable metric rather than the underlying goal?
Type: binaryCould the metric be gamed or optimized independently of the true outcome?
Type: binaryDoes improving the metric create incentives that worsen the underlying phenomenon?
Type: binaryIs this distinct from Goodhart's Law (metric ceases to be good measure) by involving active backfire?
Type: binaryThe cobra effect describes situations where an attempt to solve a problem through incentivizing a measurable proxy not only fails but actively worsens the underlying phenomenon it was meant to address. Named after the apocryphal British colonial policy of offering bounties for dead cobras, which incentivized snake farming and increased the cobra population. Distinct from Goodhart's Law in that the cobra effect involves genuine backfire.
Metrics are always imperfect proxies for underlying goals. Incentives attached to metrics create pressure to optimize the metric by any means available, including means that improve the measurement while worsening the underlying reality.
Use multiple metrics that capture different aspects of the goal. Assess the plausibility of gaming. Test whether metric improvement is correlated with improvement in harder-to-game indicators. Apply logic of anticipated behavioral responses to metric incentives.
Soviet production quotas measured by weight led factories to make nails as heavy as possible. Nails counted by number led to tiny, useless nails. Both are cobra effects: the metric was optimized at the expense of the goal.
Use these tools to detect, analyze, or train this aspect.