Illegal activity is widespread around the world, in part because of corruption and asymmetric information. Agents with observing responsibilities could be bribed to overlook illegal activities, while the enforcer does not have an independent source of information to detect it. We created a novel dataset using machine learning predictions on satellite imagery features to measure illegal mining. Then we disclosed our predictions in a $2\times 2$ randomized controlled trial to study the response of illegal activity to revealing its existence. Municipalities were randomly assigned to one of four groups: (1) the observer (local government) was informed of 5 potential mine locations in his jurisdiction; (2) the enforcer (National Government) was informed of five potential mine locations; (3) both observer and enforcer were informed, and (4) control group, where no agent was informed. In this paper we present results on the response of government agents to the information. We find that when the prediction model is wrong, according to independent verifications, local officials respond accurately that there is not a mine in the disclosed location. However, when the model is correct, local officials are less likely to confirm the existence, especially when the mine is illegal. The differential accuracy on legality of the mine is not present on the National Government verifications. We interpret these results as suggestive evidence of collusion between the local authorities and the miners.