I think we just have a tendency to anthropomorphize things that are a bit too complicated to understand fully. Like a child calling the clutch mechanism in a yo-yo a "brain" for example. It's not that the yo-yo can really think, it's just that it has a behavior that seems that way. So indeed once you've upgraded your system to the point of not fully understanding how something that isn't a human could achieve whatever emergent behavior occurs, go ahead and anthropomorphize it by calling it AI.

There's not a specific line in the sand, although tasking it with machine learning (in which outcomes improve based on collecting runtime inputs, rather than based only on its creator adding capabilities) would be a decent one. That's fairly human-like, while non-ML workloads are more plant-like.