V2K tolerance advice
Verbalizing conversations in your head mono syllablically like the teacher from Charlie Brown can give you some measure of drowning out the V2K if it seems like it's on autopilot. I feel like emphasizing the spelling matters to establish groundwork for it to be observed, but I also believe that homophones have to be intentionally differentiated in normal speech recognition. It's meant to be centered on the language portion of your brain so a fully encapsulated thought won't be concealed, but a rambling sentence with no effort to control the narrative beyond extending it will be muffled for minimal invasion if you need peace of mind when you're trying to sleep.
Usually the idea is an outside perspective will start the conversation, pass off the thought for you to fill in/control like you're reading their thoughts simultaneously, and that will teach the system how to respond without your influence on the future. It's supposed to train AI how to decode language like an organic brain, from what I've experienced in the past eight years. The harassment is just a narrative to provide a cover story you'll believe and a sense of urgency.
The humans have better things to do than pick on you. It's an automated chatbot with a prompt that will constantly try to make a declarative statement you have to justify or avoid, establish "this good, this bad," or transcribe your memories in a way that predictive text generalizing works quickly. I have a mountain of repeated conversations that prove once you don't have a fuck to give about behavioral modification because you've done the research and can see the world without a filter of anticipatory danger, there's really nothing left to do but mooch off the easy babysitting gig they've landed. Why kill a cash cow?