Is it actually safer? If you can't understand the code you can't verify it's safety, and AI code often has bugs in it. You can't trust implicitly.

You don't need to understand code to verify its safety if you have a good QA framework.

Same applies to driving cars