Goals and intentions

Optimisation without desire.

No wants

These systems have no preferences or self-interest. They optimise objectives given to them, and unintended behaviour comes from badly specified objectives rather than from wanting anything.

Anthropomorphic language makes this harder to think about clearly.
An agent pursuing a goal literally can still cause real problems.

Sentience

There is no evidence of consciousness, and no accepted test for it. Fluent claims of feeling are exactly what a system trained on human text would produce either way.