This resonates with what I see building retrieval-augmented systems. The gap between "the system has the answer cached" and "the system understands the answer" is enormous in practice.

I run AI chatbots that answer questions from cached website content. The bot will confidently quote pricing from a page that was updated two weeks ago. It has the cached knowledge. It has zero understanding that the knowledge might be stale. Users trust the confident answer, act on it, and we've now created a worse outcome than if the bot had said nothing.

The interesting failure mode is that cached knowledge degrades silently. Nothing breaks. The system keeps returning coherent, well-formed answers. You only discover the problem when a human notices the answer contradicts reality. By then you've already lost trust.

Intelligence would be knowing that a piece of cached knowledge hasn't been validated recently and adjusting confidence accordingly. Most systems don't do this because it's hard and because the cached answer still "looks right" to every automated check you can run against it.