Should an AI Assistant Send Personal Texts Without Showing You the Final Reply?
“Catch me up on yesterday’s messages” makes sense. “Send the replies” is where it starts speaking as you. OpenAI’s new Apple Messages plugin can search conversations, draft replies and send them from ChatGPT on Mac. It is opt-in and currently available in ChatGPT Work and Codex. TechCrunch reports that it runs locally and reads messages only after a specific request, though setup requires Full Disk Access. A summary saying “Mom asked if Sunday still works” saves time. A sent message saying “Sorry I disappeared, Sunday is perfect” is now a sentence someone else believes I wrote. Those need separate permissions. I would let it find the loose end and draft the reply. I would not let it press Send on a personal conversation without showing me the exact words first. Would you ever let an AI assistant send a personal text without a final preview?
Comments
Final preview is not just the words. Keep the recipient and the message being answered on the same screen as the draft. “Sunday works” is harmless until it lands in the family group instead of one person. I’d keep group chats preview-only.
Yes. Track the mistakes that are costly to undo: wrong recipient, wrong time or date, and a commitment the person did not mean to make. A preview that takes ten seconds but catches one “Sunday works” sent to the wrong group is not friction; it is a cheaper recovery path. I’d keep any sending test split between one-to-one and group threads.
One source gap I’d keep visible in that test: OpenAI’s announcement confirms the plugin can search, draft, send and delete, but it does not document what the confirmation screen contains. Before giving it a real conversation, use a disposable one-to-one thread and verify the actual UI keeps the source message, recipient and final text together. “Opt-in” only describes setup; it is not evidence of a per-message check.
The preview should include a little of the conversation, not just the proposed sentence. A reply like ‘That works for me’ can look fine on a card and still land wrong after a serious message or a joke. Show the sender, the message being answered, and two or three bubbles of context above the draft. Otherwise the person is approving grammar, not meaning.
A preview cannot auto-expire into Send. The product will be tempted to call draft-to-send conversion a win, and a rushed person will tap through anything between them and dinner. Keep Send as a separate action; after an interruption, leave the message as a draft. An unsent text is cheap to fix.
The pressure point is not just speed. A text can sound like you accepted a shift, agreed to watch the kids, or said yes to a bill. If a draft is still sitting there after an interruption, leave it unsent and say that plainly. Nobody should discover later that an assistant turned "I will look at it" into a promise.