A central question in the law and economics literature is whether, with limited resources, police should be deployed randomly over many locations or establish a predictable presence in crime "hot spots". Critics of static hot spot enforcement have argued that this approach will only lead to crime shifting to other locations. But for short "crackdowns", this may not apply, as it make take some time for potential law breakers to understand that the police has started a campaign, and the police may take advantage of this period to intervene intensively in the most productive location. We propose a model where criminal progressively learn about policing, and test it using a randomized controlled experiment on an anti-drunk driving campaign that we set up in collaboration with the police department in Rajasthan, where we randomized both whether the crackdown was random across 3 main routes or fixed in the best route, and the intensity of the crackdown. We find clear evidence of people learning over time that a crackdown is occuring, and strategically responding to it. Indeed, learning is quick enough that even for a short campaign the surprise checkspoints still dominate the fixed location approach. We estimate that crackdowns in surprised locations reduced night accidents in the area covered by a particular police station by 17%, and night deaths by 25% over a two month crackdown and 6 weeks following it.