Unfortunately, your third point has already been attempted multiple times with multiple different implementations. Unless you have a significantly new approach to offer that takes into account human neurobiology, I predict with a high degree of certainty that CLIP is a) either unlearnable by humans without severe modification, or b) that within less than a full generation of use humans will re-evolve the ability to easily express incorrect reasoning.