Software that can has been (or can be) trained to make decisions based on rules of its own invention, as opposed to rules explicitly programmed by the software author. Instead of starting with rules, an ML system starts with sample data. The ML system learns from that data, and constructs its own complex rules to make more accurate decisions.
The software design is called a neural net, in part because it is designed to mimic the neurons in a human brain. It learns, much like a human brain, although in a much more limited way. Each ML system is generally designed for just one limited task, but a device may include multiple ML systems.
ML systems are excellent for recognizing objects or patterns based on one or more inputs. In a phone, those inputs might be visual data from the camera, audio, other sensor data, or phone usage.
For example, an ML system can be taught to recognize apples in photographs. Instead of a human designing specific rules to recognize apples, the system is trained (taught) by inputting many (thousands) of different photos, and told which ones contain apples. The system learns from that sample data, automatically creating its own "rules" for recognizing apples.
ML systems are often referred to as AI (artificial intelligence). This is not accurate in the strict sense, but the terms have become somewhat interchangeable. (A true AI would not be limited to just one task.)
In many cases, ML systems can exceed the recognition ability of humans.