That's incorrect, unless you consider machine learning algorithms to suddenly be subjects/"knowers". They suffer from some of the same perceptive illusions that plague our optical processing.
So to answer your question directly, there is no "who", there need only be a "what", at which point any problem with calling qualia an illusion evaporates. There is no irreducible subject involved, there is only an information processing automoton that integrates facts acquired from perceptions. When a perception is taken to be accurate but it entails a conclusion that is false, that's an illusion.
I could argue that any machine learning algorithm is simply making probabilistic calculations based on input data in order to reach some kind of singular output. In this view (which I wouldn't discount), the machine learning algorithm is little more than a very advanced function. In my view, a function is not something I would consider a "subject" not would I consider it to have any ability to "perceive" an illusion. There's no sentience.
If we say the complexity of the algorithm is great enough that it can be said to "perceive", I would argue the algorithm does become a "who" instead of a "what". Still, there is a surprising amount of nuance to this concept that makes it very interesting.
Pure speculation, and almost certainly false. If you hadn't evolved the pleasurable feeling of eating an apple, or any other food, you would have died and your inability to feel would have gone with it. Feelings provide drive and motivation, they are functional.