What is the learning phase, and what does leaving it actually tell you?
The learning phase is the period after a new or significantly edited ad set during which delivery is still unstable and results are not yet representative. Meta describes it as ending once an ad set has had roughly 50 optimisation events within a week. Leaving it is a delivery milestone, not a verdict: an ad set can exit the learning phase and still be losing money.
Automated delivery systems need a minimum flow of outcome signals before their bidding settles down. Meta names this period explicitly and shows it in the interface; Google's Smart Bidding strategies have the same behaviour after a strategy change or a target change, without the same label. During it, cost per result swings widely, and reading a two-day slice of that swing as performance is how good ad sets get killed and bad ones get scaled.
The practical trap is that significant edits restart the clock. An account that is adjusted every couple of days can keep several ad sets permanently in learning, which produces exactly the unstable numbers that prompt the next adjustment. The related status, learning limited, means the opposite problem: the ad set is not receiving enough optimisation events to ever stabilise, which is usually an audience-size or budget-per-ad-set issue and almost never a creative issue, though it is very often diagnosed as one.
The error worth naming is the second one, because it looks like patience. People correctly wait for learning to finish before judging, and then treat the finish itself as the judgement. Exiting learning means the numbers have become readable. Whether they are good is a completely separate question, answered against your cost ceiling, and the answer is sometimes no.