does not actually understands the meaning
First, I want to say the next paragraphs are not to argue against what you wrote. The next paragraphs are to show a conclusion I've come up some time ago, before this topic and unrelated to AI. Bringing it because it seems to apply well to AI, too.
I was trying to clarify what we mean when we say we "understand" (how something works, or what that something is), and came up to the conclusion that
we think we understand only after we are able to assign a causality chain.
The funny thing is we "build" that causality in terms of what we know, or experienced, those are the building blocks. For example, it is very easy for a kid to say "I understand Kermit the frog, when in fact that's a sock over somebody's hand". It is irrelevant if those building blocks are correct or not. Maybe a little too far fetched example, my point is the building blocks doesn't have to be correct.
Back to our AI, it was trained in a very different context than we, humans, are trained. AI's world building blocks are very different from ours (for now), because humans and AIs were trained in totally different worlds (a world of text for AI vs. the physical world for us). But the training environment can be changed in the future, so no problem with that.
We must note that the capability of learning is there, and if we apply the above definition of "understanding", then the AI also understands its world (a world of text

).
After all, even for us, humans,
mind is nothing but a driven illusion. Physics tells us it should be all fields and particles and thermal wiggling, and other very-very-very strange and different things than what we experience.
It looks to me that the AI's abilities to learn and understand are real, and similar with ours. By learning I mean the ability of making sens and finding structure in a given data stream (so learning not as in learning a poem, but learning as in noticing patterns). AI can notice and detect patterns, just like us. That's learning.
For now, it's a matter of size. We are much bigger neural networks, though the computers can compute faster, so an AI might not need as many neurons as a human. AI could be a close match to human intelligence.