We report results from an experiment on a repeated sender/receiver game with twenty periods in which one of the periods has higher weight, the sender communicates about the realized state in each period, the receiver takes an action matching her belief about the state, and then learns whether the sender lied. Receivers are matched either with malevolent (human) senders who prefer the agents to take wrong decisions or with benevolent (machine) senders who always tell the truth. Our ndings do not support the predictions of the Sequential Equilibrium. The deceptive tactic in which malevolent senders tell the truth up to the key period and then lie at the key period is used much more often than it should and it brings higher expected payo. We suggest that our data are well organized by the analogy-based sequential equilibrium (ABSE) in which (some subjects) may reason coarsely when making inferences and forming expectations about others' types and behaviors.