The average outcome if you repeated something many times, weighting each possibility by how likely it is. Useful for deciding, and misleading on its own, because the average of a distribution that is mostly small losses and occasionally catastrophic looks identical to one that is mildly negative every time.
Technical view
The probability-weighted mean of a random variable. Training minimises expected loss over the data distribution, approximated by the empirical mean over samples. Its weakness is insensitivity to tail risk, which is why practical evaluation reports percentiles and worst-case behaviour alongside means, and why expected-value reasoning alone is a poor guide to decisions with irreversible downside.
Related
Part of AI Terminology Index, section: Maths and Statistics.