Without human verification, an LLM can generate correct or incorrect proofs but it can't tell the difference. A human is necessary to be able to tell one from the other.

Saying that's a solution "done autonomously by an automated AI pipeline" is like saying that a self driving car that can only take you to the nearest train station after which you have to ride the rain to where you're going is "autonomously" driving you to your destination. Which is exaggerating the autonomy of the system, rather.