Rating a Claude or ChatGPT answer feels like grading the tool. On Anthropic's commercial products, that click is permission to keep the whole conversation and train on it. On OpenAI's, it depends on which product you are in and a setting most lawyers have never seen. The same click can be a disclosure decision - and the client was never asked.
Your AI Vendor Keeps a Copy. That Is Not the Problem.
An AI vendor's copy of a client trade secret, held under confidentiality terms, does not lose the secret. What loses it is an obligation that expires while the copy lives on, or a copy you cannot require the vendor to delete.
Every enterprise AI tier promises not to train on your data. None of them promises not to keep it. Lawyers are treating the first as if it were the second.
AI Did Not Break Privilege. There Was Never Privilege to Break.
Everyone is reading United States v. Heppner as the case where a chatbot destroyed attorney-client protection. Read it beside a quieter ruling from the same weeks that came out the other way, and the real lesson is one lawyers already knew.
There is no single answer to what the tool does with your inputs. The terms depend on how you reach the model and which one you pick - and they changed three times while you were not looking.
