These systems have far fewer "Likes" than other implicit signals such as which posts the user read (ie, didn't scroll past). So they use all of those signals and likes are unlikely to have a strong impact on the output of the algorithm.
The algorithms are opaque by nature. Even the developers of these deep models won't be able to explain to you why a user's recommendation list is ranked the way it is. Now imagine a user contemplating the consequences of liking a certain item on their future recommendations. It's hopeless.
As a result, the developers learn not to rely on these "like" signals and the users learn to provide likes only to affect the directly observable changes - the like counter.