As it applies to visual perception: The fact that the in-context state influences how AIs (at least Claudes) seem to even perceive images at a low level (e.g. seeing cats or hamburgers in generic forms) is quite interesting. It's imaginable that their image perception faculties would operate mostly independently of / fixed with respect to the rest of their state, but that is not the case. It's seems similar to how humans' visual perception can become more hallucinatory under the influence of some psychedelics or when they have some kinds of vision loss such as optic nerve damage (and AIs have poor vision by human standards as well).