Inês deleted the ChatGPT conversation the moment she was done. She had just walked the model through a tricky client case, with a name, an account number and a few sensitive details, to get a clean summary faster. Done, deleted, gone. Or so she thought.
But the chat log outlives the chat. On screen, the conversation has vanished from her history. In the systems behind it, the story is different. And that is exactly the misunderstanding: we treat deleting as the same thing as erasing. With AI tools it rarely is.
Deleting is not erasing
Take the most common example. With ChatGPT a conversation disappears from your history the moment you delete it, but permanent deletion from the systems normally follows within about 30 days. Temporary Chats, which never appear in your history, are also kept for up to 30 days before they are deleted. Source: OpenAI Help Center, chat and file retention policies, accessed June 2026.
Usually that one-month window is not a problem. Until it is. During that period the conversation still exists somewhere, with everything in it. And the timeline is not a law of nature: a legal or regulatory obligation can stretch it.
A deleted conversation that was not allowed to go
This is not theory. In May 2025 a US magistrate judge, Ona Wang, ordered OpenAI to preserve and segregate all ChatGPT output log data, in the lawsuit The New York Times brought against OpenAI. The order explicitly covered conversations users had already deleted, and Temporary Chats too. That paused the usual deletion within 30 days: data that would normally be gone now had to be kept.
The order applied to the ordinary accounts: Free, Plus, Pro and Team, plus the standard API. It did not apply to ChatGPT Enterprise, ChatGPT Edu or API usage with Zero Data Retention, where the data is not retained in the first place. OpenAI objected, calling it at odds with the privacy commitments it had made to users. The preservation requirement ended on 26 September 2025. Source: OpenAI, "How we're responding to The New York Times' data demands", and legal reporting on the case, May to October 2025.
The uncomfortable part of this story is not the outcome. It is what it reveals. A user who hit delete in June 2025 did everything right as far as the settings were concerned, and the chat stayed anyway, beyond their control, because a judge in a case they had no part in decided that deleted data had to be kept. Your delete button is not the final word.
Memory remembers what you forget
There is another route by which a detail lingers, and it sits outside your chat history. ChatGPT can store information from a conversation as a memory, and can also reference past chats to make later answers more personal. What Inês explained in one chat can therefore surface in a later one, even after she deletes the original.
This deserves its own piece, and it gets one: see the separate explainer on how AI memory works. For now, the point is simple: deleting a conversation does not automatically erase what the model retained from it.
On a work account, the admin can read your chats
Then there is the workspace. On the consumer version your employer cannot simply look at your conversations. On a work account, things are different. On ChatGPT Business, workspace admins can view, export and delete the conversations in the workspace. On ChatGPT Enterprise, ChatGPT Edu and ChatGPT for Healthcare, admins can also retrieve conversation logs and metadata through the Compliance API, built for audits and eDiscovery. Business does not have that API. Sources: OpenAI, enterprise privacy, and OpenAI Help Center, Compliance API for Enterprise and Edu, both accessed 20 October 2026.
That is reasonable in itself, and often required in regulated sectors. But it has consequences for what an employee types. Who knows what their employer's plan lets an admin see? Almost nobody. What your employer can see of your AI chats differs by tool and by plan.
What the settings do not solve
You can do a lot in the settings. You can stop training on your data, you can limit memory, you can use Temporary Chat. Good steps, and they do what they promise. But they all share the same gap: they do not pull a pasted detail back out of the log.
The preservation order overrode the delete button. Memory works independently of the chat history. The admin can read your chats regardless of your settings. Three different routes, one thing in common: once the detail is in the log, you have lost control of it. The settings decide what happens to the log. They do not decide what goes into it.
So the moment that really counts comes earlier: when the detail goes into the text box. That moment belongs to the employee, not to the settings. What never goes in does not need protecting afterwards, because it is not in the log. The same root cause lies behind the broader privacy risks of ChatGPT: not the tool itself, but what ends up in it.
What you can do before anything reaches the log
You can act at the moment it counts: before the conversation begins. BeeSensible highlights sensitive data while you type, across AI tools, email and other web apps. The text is sent to BeeSensible's EU detection service, where it is analysed in memory and discarded after detection. None of it is stored. Sensitive details (a name paired with an account number, a card number, a national ID) are highlighted before the text is sent. Inês sees the account number light up as she types out her case, and decides for herself whether to mask or remove it before she hits send.
Here is what that looks like the moment someone pastes client data into an AI tool:
Outlook
