When you delete an AI chat, the conversation vanishes from your screen instantly, but OpenAI’s own documentation says the underlying data is only “scheduled for permanent deletion” from its systems within 30 days, not erased on the spot. That gap between what you see and what actually happens on a company’s servers is not a hypothetical concern. For more than four months in 2025, a federal court order meant that deleting a ChatGPT conversation did almost nothing at all: the chat still disappeared from your account, but OpenAI was legally required to keep a copy anyway.
That episode is the clearest real-world proof that “delete” and “gone” are not the same word when it comes to AI chatbots, and it is worth understanding exactly what deletion does and does not do before you type something sensitive into any AI tool.
What Actually Happens When You Delete an AI Chat
According to OpenAI’s own help centre documentation, when you delete a chat it is removed from your account immediately, but it remains on OpenAI’s systems for up to 30 days before permanent deletion, unless the chat has already been de-identified or OpenAI is required to retain it for legal or security reasons. Files attached to a deleted conversation follow the same 30-day clock. Temporary Chat mode, ChatGPT’s closest equivalent to an incognito setting, is not saved to your history at all and is automatically purged within the same 30-day window without you needing to delete anything.
ChatGPT’s Memory feature adds a further wrinkle most users never notice. When Memory is enabled, the model can extract and retain specific facts and preferences from a conversation, your job, your family situation, ongoing projects, separately from the conversation itself. Deleting the original chat that generated a memory does not automatically clear that memory. It has to be reviewed and removed manually in Settings, which means a conversation you deleted specifically because it contained something sensitive may have already left a durable trace elsewhere in your account that deletion alone will not touch.
That 30-day figure is standard across most major consumer AI platforms, though the exact mechanics differ. The pattern that holds everywhere is the same: deletion from your visible history is close to instant, but deletion from the company’s actual infrastructure, backups, abuse-monitoring systems, and safety logs is a scheduled process, not a switch you flip yourself.
The Court Order That Made “Delete” Not Mean Delete
The starkest illustration of this gap played out over the course of 2025, when the New York Times’ copyright lawsuit against OpenAI collided directly with ChatGPT’s deletion policy. In May 2025, a federal magistrate judge ordered OpenAI to preserve and segregate all ChatGPT output data that would otherwise have been deleted, specifically because the Times argued that users trying to evade paywalls with ChatGPT would likely mark their conversations as temporary or delete them afterward. For the following four months, deleting a ChatGPT conversation did not remove it from OpenAI’s systems at all. It only removed it from your own account view, while a segregated, legally held copy remained accessible to a small internal team.
OpenAI’s own account of the dispute confirms the order was lifted on September 26, 2025, after the company fought it through multiple rounds of appeal, and that standard 30-day deletion practices resumed after that date. But the order left a lasting consequence: OpenAI is still required to securely retain a specific set of user data from the April to September 2025 window, locked under legal hold and accessible only to an audited legal and security team, regardless of whether the users who generated it ever deleted their conversations.
The Discovery Order That Went Further Still
The preservation order was only the first stage. In November 2025, the same litigation produced a second, more consequential ruling: Bloomberg Law reported that a federal judge ordered OpenAI to produce 20 million de-identified ChatGPT logs directly to the New York Times and other news plaintiffs as part of discovery, and that OpenAI’s attempt to narrow that order to only the conversations most relevant to the case was rejected. A district judge affirmed the order in January 2026, ruling that ChatGPT’s ownership of the logs was uncontested and that users had voluntarily submitted the conversations in the first place, a legal reasoning that treated user chat data as company-owned business records rather than something closer to private correspondence.
The practical upshot for ordinary users is unsettling regardless of how the underlying copyright case eventually resolves: a conversation you deleted between April and September 2025, believing it gone, may exist today as one of 20 million logs sitting in a litigation database, de-identified but not necessarily un-reconstructable, reviewable by lawyers on both sides of a lawsuit you were never a party to and never consented to be involved in.
Deletion and Training Are Two Separate Questions
A separate and equally important distinction is that deleting a conversation and preventing it from training a model are not the same control. OpenAI’s consumer terms allow chat content to be used to improve its models by default, and turning that off requires a separate setting in Data Controls. Critically, that opt-out only applies going forward. If a conversation contributed to model training before you disabled the setting, or before you deleted it, that contribution cannot be clawed back. The influence exists somewhere in the model’s weights independent of whatever you later did to the original conversation.
This is the same underlying technical limitation LiveAIWire has traced in the AI right to be forgotten: even where a legal right to erasure exists, and even where a company genuinely wants to honour a deletion request, removing a specific data point’s influence from an already-trained model requires machine unlearning techniques that remain, at best, an active area of research rather than a solved, verifiable process. Deleting the source conversation is straightforward. Undoing what that conversation already taught the model, if it was used for training before deletion, is not something any major AI company can currently guarantee.
Why This Matters Beyond One Lawsuit
The OpenAI case is the most visible example, but it is a symptom of a broader legal reality rather than an isolated incident. LiveAIWire’s coverage of the wider wave of AI training data lawsuits has documented more than a dozen active cases against major AI companies, and litigation discovery is a well-established mechanism that can compel any company holding user data, AI or otherwise, to preserve and produce it regardless of what a privacy policy promises.
Any AI platform currently facing active litigation, and most major consumer AI companies now are to some degree, carries some baseline risk that ordinary deletion practices could be suspended by a future court order, exactly as happened at OpenAI for four months in 2025.
What This Means for What You Type
The practical lesson is not that AI chatbots are unusually unsafe compared with other cloud services; email providers, cloud storage, and social media platforms all face similar discovery obligations when they are sued. The lesson is that “I deleted it” should never be treated as equivalent to “it never happened,” particularly for anything genuinely sensitive: health details, legal strategy, financial information, or anything you would not want to see attached to your name in a courtroom exhibit years from now.
That caution matters most in exactly the conversations people are least likely to think about it during. LiveAIWire’s reporting on AI therapy chatbots found that people routinely share their most private mental health disclosures with AI tools built for emotional support, often without registering that those conversations sit inside the same retention and legal-discovery framework as any other ChatGPT chat, subject to the same 30-day windows, the same training opt-out limitations, and the same vulnerability to a future preservation order if the company is ever sued.
The same blind spot applies to the fast-growing category of AI companion apps. LiveAIWire’s coverage of AI companion apps and the loneliness crisis found that products explicitly designed to deepen emotional engagement, and therefore to encourage longer, more personal disclosures, have some of the weakest independently reviewed privacy protections of any AI product category, which is precisely the combination that makes understanding what deletion actually does matter most.
How Other AI Providers Compare
OpenAI’s situation is the most thoroughly documented because it is the one that has been tested in open court, but the underlying architecture, soft-deletion on your side, a scheduled removal process on the company’s side, and a legal-hold exception that can override both, is standard across the industry rather than unique to ChatGPT. Anthropic’s own documentation describes comparable retention windows for deleted conversations, and Google’s Gemini and xAI’s Grok publish similar structures: an immediate removal from your visible account, a defined backend deletion window measured in weeks rather than instantly, and a reserved right to retain data longer where law or active litigation requires it.
None of the other major providers has yet faced a preservation order on the scale of the one that hit OpenAI in 2025, but that is a difference of litigation history rather than a difference in underlying architecture. Any AI company sued in a way that puts user-generated content at issue, whether over copyright, defamation, or another claim entirely, has the same legal exposure OpenAI did: a court can compel preservation of exactly the data a privacy policy promises to delete, and the company’s own commitments to its users do not override a valid discovery order.
What You Can Actually Control
Within the limits described above, a few concrete choices genuinely change your exposure. Turning off “Improve the model for everyone” in Data Controls before a sensitive conversation, not after, prevents that specific conversation from ever being eligible for training, since the setting only applies prospectively. Using Temporary Chat mode for anything you do not want retained beyond the standard window avoids leaving a conversation sitting in your visible history at all.
And treating business or Enterprise-tier accounts as meaningfully different from consumer accounts is not paranoia: OpenAI’s own documentation confirms ChatGPT Enterprise and Edu customers were excluded from the 2025 preservation order entirely, and Zero Data Retention API customers never had their prompts logged in the first place, because that data was never stored to begin with.
None of these controls make a deleted conversation instantly and permanently unrecoverable under every possible future circumstance. What they do is meaningfully reduce the two biggest risks documented here: unwanted training contribution, and exposure to a preservation order triggered by litigation you have no connection to. Given how one already played out, that is not a hypothetical worth ignoring.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.