Does ChatGPT train on what I paste in, and how do I turn that off?

Context: I am a paralegal at a small firm. Nobody asked me to use AI, but a colleague mentioned she pastes paragraphs into a chatbot to tidy letters, and I want to know what the real risks are before anyone does that with client material. I read the terms closely because that is the job, and the help pages contradict each other in places or are written so that you have to read three pages to get one answer.

What I have worked out so far:

• The training toggle exists and is documented
• There is a "Temporary Chat" mode that is described as not used for training
• Old opt-out requests sent to support are said to be honored, even if the toggle does not show it
• Business plans have separate terms

What I still want to know:

1. If training is off, is the text still stored, and for how long?
2. Does the toggle cover files I upload, or only typed text?
3. Does turning it off make it acceptable to paste client information? My instinct says no, and that the answer depends on the firm's policy and the engagement terms, not on a setting.
4. Is any of this different on other tools, such as Claude or Gemini, so I can answer a colleague who uses them?

I would like step by step answers from anyone who has checked their own settings, plus a pointer to the primary pages.