Animals must continually assess their uncertainty about their environment and future events in order to drive choice behavior and learning. Because uncertainty is a key decision variable for myriad behaviors (e.g. curiosity, attention), neuromodulatory systems, which engage and regulate diverse brain regions and brain states, are ideally suited to subserve uncertainty-guided modulation of local brain computations. Theoretical studies have hypothesized that the neuromodulator acetylcholine is responsible for encoding “expected uncertainty” in the brain (i.e., learned unreliability of predictive cues) (Yu & Dayan, 2005), but direct experimental evidence is scarce. To address this, we used a recently developed acetylcholine sensor (Jing et al., 2020) to monitor acetylcholine release in the basolateral amygdala during behavior. To study outcome uncertainty, we trained mice to associate 3 different odors with distinct probabilistic outcomes, each odor being assigned a 75, 50 or 25% reward probability. Hence, whereas the value associated with each odor increased with the reward probability, the uncertainty was maximum for the odor with a 50% chance of reward. Initially, reward delivery triggered acetylcholine release. After training, as previously reported, reward-related acetylcholine transients were greatly diminished (Sturgill et al., 2020). Further, acetylcholine release was rapidly triggered by the presentation of predictive cues. Remarkably, these cue responses increased with cue value but not uncertainty. The cholinergic sensor also revealed rapid responses, reminiscent of quantal release. To quantify these transients, we developed an event detection algorithm enabling us to assess the latency and amplitude of cholinergic release events across learning. We found that high-value cues elicited lower latency and larger magnitude acetylcholine transients compared to low-value cues. Neither of these metrics scaled with uncertainty during or after cue presentation. Our data supports the idea that acetylcholine release scales with reward prediction rather than uncertainty as previously hypothesized in theoretical models of reinforcement learning.