It’s a behavior produced by code which may be relied on by consuming code, though that behavior is not specified by the spec the code is written to. Hence, other implementations of the same spec may not exhibit the same behavior, and will break that consuming code. That’s why it’s bad to rely on them.
For instance, if you were to write some code against a list interface which specified an array sort but not the algorithm it used, and you needed the sort method to be stable, and a version of your code was used with a non-stable sort algorithm, then your code would break.