Let's take the AI example in a slightly different direction: Consider an AI built as a neural net with many input lines and output effectors, and a few well-chosen reward signals. One of the input lines goes to a Red Detector; the other input lines go to many other types of sensors but none of them distinguish red things from non-red things. This AI then gets named Mary and put into a black and white room to learn about optics, color theory, and machine learning. (Also assume this AI has no ability to alter its own design.)

Speculation: At the moment when this AI Mary steps out of the room into the colorful world, it cannot have any immediate perception of red (or any other color), because its neural net has not yet been trained to make any use of the sensory data corresponding to redness (or any other color). Analogously to how a young child is taught to distinguish a culturally-specific set of colors, or to how an adult can't recognize lapiz versus cerulean without practice, our AI cannot so much as distinguish red from blue until adequate training of the neural net has occurred.

If that line of reasoning is correct, then here's the conclusion: Mary does not learn anything new (perceptually) until she learns something new (behaviorally). Paradox dismissed.