Quick warning before you type another word into ChatGPT: hitting “delete” on a chat doesn’t make it disappear. According to a post I just read from this savvy professional on LinkedIn, your deleted conversations may already be sitting in a training pile, and no amount of clicking will pull them back out.
I’ll be honest, I assumed delete meant delete. Most people do. The original poster put together a list of the ways your prompts leak out of your control in 2026, and it’s sobering. I’m sharing the key points here, plus the practical steps you can take today to close the gaps.
✦ The story that made me sit up
The post opens with an example that stuck with me. Two mathematicians used Codex to work on a Millennium Prize problem, one of the hardest open questions in math. Three weeks later, OpenAI solved it too. When someone asked whether the model had learned from their sessions, the answer was blunt: “Yes. And so does every LLM company.”
Think about what that means for your own work. If a model can pick up a math proof from a user’s session, it can pick up your client proposal, your unreleased product roadmap, or the salary spreadsheet you pasted in for “a quick summary.” The expert’s point is simple: whatever you type is potential fuel for the next model version, unless you switch that off.
✦ Training is on by default
This one surprised me the most. Both ChatGPT and Claude ship with data training switched ON. You don’t opt in. You’re already in, and you have to opt out. The author’s fix is short: go to Settings, then Privacy, and flip the training toggle off. It takes under a minute. Do it on every account you use, including the one on your phone.
✦ Your thumbs up is a training signal
Most of us never think twice about this habit. When you tap thumbs up or thumbs down on a reply, that feedback goes straight into the pile used to train the models, and the conversation goes with it. The creator’s rule: never rate a chat you wouldn’t want a stranger to read. If the conversation contains anything personal or confidential, skip the rating entirely.
✦ Share means publish
The share button feels harmless. It isn’t. In 2025, Google indexed roughly 100,000 shared ChatGPT chats, which means anyone with the right search query could read them. Those “private” links were public pages. The post’s author recommends going back through your account and deleting every share link you’ve ever created. If you can’t remember which chats you shared, assume the worst and clean up all of them.
✦ Delete doesn’t undo training
This is the core warning, and it deserves its own section. Deleting a chat removes it from your view. It doesn’t reach back in time. If your conversation was copied into a training set before you deleted it (usually stripped of your name, but still containing the content), that copy stays exactly where it is. In other words, deletion protects you going forward, not backward. That’s why the order of operations matters: turn training off first, then start chatting.
✦ Incognito isn’t invisible
Temporary or incognito chats hide the conversation from your sidebar. That’s all they do. On a Team or Enterprise plan, the workspace owns the data, and your admin can still see it. So incognito hides your chat from you, not from your boss. If you’re using a company account for anything you’d rather keep between you and the model, that’s the wrong tool for the job.
Your defensive checklist
Pulling this LinkedIn creator’s advice together, this is what I’d do today, in order:
- Open Settings, then Privacy, in ChatGPT and Claude and switch data training off on every account.
- Stop rating replies in any conversation that contains personal, financial, or client information.
- Go through your shared links and delete every one of them.
- Treat incognito or temporary chats as a sidebar cleanup feature, not a privacy feature.
- Assume anything you typed before today may already be in a training set, and act accordingly for future chats.
- On a work account, ask your admin what the workspace retention and training policies are before pasting anything sensitive.
One more habit worth adding, from my own experience: before pasting a document into any chatbot, strip out names, account numbers, and anything that could identify a person or a company. The model will still summarize it just fine, and you lose nothing.
Why it matters: AI assistants are becoming the default place we think out loud. Contracts, health questions, code, strategy. The tools are built to learn from that input unless you tell them not to, and by the time you remember to delete something, the useful part may already be out of your control. A minute in the settings menu today prevents a problem you can’t fix later.
The original post has more detail on each of these risks, and the person who shared it clearly knows this space inside out. Go read the full LinkedIn post, then spend that one minute in your privacy settings. Future you will be glad you did!