CMIIW currently AI models operate in two distinct modes:

1. Open mode during learning, where they take everything that comes from the data as 100% truth. The model freely adapts and generalizes with no constraints on consistency.

2. Closed mode during inference, where they take everything that comes from the model as 100% truth. The model doesn't adapt and behaves consistently even if in contradiction with the new information.

I suspect we need to run the model in the mix of the two modes, and possibly some kind of "meta attention" (epistemological) on which parts of the input the model should be "open" (learn from it) and which parts of the input should be "closed" (stick to it).