Yes, that would be its natural approach. A model learns about pain from how people describe it and react to it, and an agent learns by acting and receiving a score that people designed. Whatever it values comes from that training.
Reading people's reactions is what psychologists call cognitive empathy, and such a system can become very good at it. Solms ties empathy to stakes of one's own. In his reply here he wrote that what's at stake is existence, sentient existence, and that this is the basis for empathy with other sentient beings.
My working hypothesis is that feeling appears only where a system has real stakes, such as irreversible loss and an irreplaceable history.