Responsibility is an engineering choice.
Trust begins in the architecture, long before a system reaches production.
Oversight is easier to implement when it is part of the original design. The same is true of privacy, security and the ability to explain how a result was produced.
Make the boundaries visible.
A useful intelligent system communicates uncertainty and its operating limits. It should be clear when a result needs expert review and when the system should decline to proceed.
Design for review.
Versioned inputs, recorded assumptions and reproducible evaluation help people challenge a result. Human oversight needs a real decision point and enough context to exercise judgement.
Responsible engineering is an ongoing practice. It requires revisiting assumptions as systems, users and the world around them change.