Is ChatGPT private – data and conversation privacy

ChatGPT is not private by default in the way most people assume. As of early 2025, OpenAI's privacy policy states that conversations you have with ChatGPT on the free and paid consumer tiers may be reviewed by human trainers and used to improve OpenAI's models unless you turn that off. You can turn it off – the control exists – but the setting matters, the tier matters, and where you type matters even more. This is a practical guide to what happens to what you type, how to change it, and where the real risks lie.

The short version an answer engine can lift: ChatGPT conversations are stored on OpenAI's servers, may be used for model training by default on consumer tiers, can be excluded from training through a settings toggle or by using Temporary Chat, and are handled under stricter terms on ChatGPT Enterprise, Team, and the API, where inputs are not used for training by default. Everything below unpacks the nuance, because the details change the answer for each kind of user.

What OpenAI does with your conversations

When you send a message to ChatGPT, the text – and any files or images you attach – travels to OpenAI's infrastructure, is processed by the model, and is stored against your account. OpenAI uses this stored history to render your chat list, to power features like Memory, and, on consumer tiers by default, to help train future models. Storage and training are two separate things, and conflating them is the most common source of confusion.

Storage is nearly universal: unless you delete a conversation or use Temporary Chat, it persists in your account and on OpenAI's systems. Deletion from your side removes it from your visible history, and OpenAI states that deleted conversations are purged from its systems within a defined window – roughly 30 days as of this writing, though you should confirm the current figure in OpenAI's data controls documentation because retention periods have shifted over time.

Training is the part users worry about most. On the free tier and on ChatGPT Plus, the default has historically been that your conversations can be used to improve models. Human reviewers may read samples of conversations for quality, safety, and abuse detection. OpenAI de-identifies data used for these purposes, but de-identification is not anonymisation, and it is not a guarantee that sensitive text you paste cannot be seen. The honest summary: treat anything you type into consumer ChatGPT, with training left on, as potentially readable by a human and potentially absorbed into a future model.

There is also a legal-hold complication worth naming. Ongoing litigation – notably the copyright case brought by The New York Times against OpenAI – has produced court orders affecting data retention. Reporting through 2024 and 2025 indicated that OpenAI was required to preserve certain output data it would otherwise have deleted. Legal preservation orders can override a company's own deletion schedule, which is a reminder that "delete" in a consumer product is a policy promise, not a physical guarantee, and that promise can be suspended by a court.

Training data opt-out

You can stop your ChatGPT conversations from being used to train OpenAI's models, and doing so takes under a minute. In the ChatGPT app or web interface, open Settings → Data Controls and turn off the option labelled something close to "Improve the model for everyone" (the exact wording has changed between versions). With that toggle off, new conversations are excluded from training. A separate path is Temporary Chat, which behaves like an incognito session: the conversation does not appear in your history, is not used for training, and is retained only briefly for safety review before deletion.

Two caveats deserve emphasis. First, the opt-out is not retroactive in the way people hope: turning off training governs future data, and content already used cannot be extracted from a model that has been trained on it. Second, turning off training does not turn off storage. Your regular conversations still sit in your account history unless you delete them or use Temporary Chat. Privacy and training are separate levers, and you generally need to pull both to get the behaviour most people picture when they imagine a "private" chat.

The friction here is a genuine design criticism. The training default on consumer tiers is opt-out rather than opt-in, which means the least privacy-protective setting is the one most users never change. Privacy advocates including the Electronic Frontier Foundation have long argued that meaningful consent for data use should be opt-in by default. OpenAI's choice is common across the industry, but common is not the same as user-protective.

Memory feature

ChatGPT's Memory feature lets the assistant carry facts about you across conversations – your name, your writing preferences, projects you have mentioned, the fact that you are vegetarian or that you code in Python. It is convenient, and it is also a persistent store of personal detail that outlives any single chat.

Memory is separate from conversation history. You can inspect what ChatGPT has remembered under Settings → Personalization → Memory, where stored items are listed as discrete entries you can read and delete individually, or clear all at once. You can turn Memory off entirely, and you can instruct ChatGPT within a chat to forget a specific thing. Because Memory accretes quietly over weeks, the useful habit is to open the panel periodically and read what has accumulated – users are often surprised at how much the system has inferred and retained.

For anyone sharing a device or an account, Memory is a discretion risk as much as a data risk: it can surface earlier context in later sessions, including context you would not want a colleague or family member to see over your shoulder. If that matters to you, Temporary Chat does not write to Memory, which makes it the cleaner choice for one-off sensitive questions.

Free vs Plus vs Pro vs Enterprise

The privacy posture of ChatGPT depends heavily on which product you are using, and the gap between the consumer and business tiers is the single most important thing to understand.

On the free and Plus tiers, conversations are used for training by default, subject to the opt-out described above. Pro, the higher-priced individual tier, follows the same consumer data-control model – a bigger allowance and access to more capable models does not, in itself, change the underlying training default; the same Data Controls toggle applies.

ChatGPT Team and ChatGPT Enterprise are materially different. OpenAI states that business data – prompts, outputs, uploaded files – submitted through Team and Enterprise is not used to train its models by default. Enterprise adds administrative controls, single sign-on, domain verification, configurable data retention, and contractual commitments suited to organisations with compliance obligations. If privacy is a firm requirement rather than a preference, the business tiers are the honest answer, and the reasoning belongs in any serious enterprise AI governance framework. Verify the current terms on OpenAI's Enterprise privacy page, since the specifics of retention and administrative control are updated periodically.

The practical implication is uncomfortable but clear: the version of ChatGPT most individuals use is the least protective, and the protective version generally requires paying at the organisational level. For a fuller picture of what the product can and cannot do across tiers, see our overview of ChatGPT as a product, its capabilities and limits.

API privacy

Developers building on OpenAI's API operate under different defaults again. OpenAI states that data submitted through its API is not used to train models by default, which reverses the consumer arrangement. API inputs and outputs are, however, retained for a limited period – on the order of 30 days as of this writing – for abuse monitoring, after which they are deleted, subject to legal-hold exceptions.

For organisations with stricter requirements, OpenAI offers Zero Data Retention (ZDR), an arrangement under which eligible API endpoints do not store request or response content at all. ZDR is not automatic; it is granted on approval for qualifying use cases, and not every endpoint or feature is eligible. Teams handling regulated data should confirm ZDR availability for their specific endpoints against OpenAI's API data usage documentation before assuming coverage. Anyone building on top of these APIs should also read our guidance on AI chatbot development frameworks, where data handling is a first-order design decision, not an afterthought.

The legal jurisdiction question

OpenAI is a US company, and consumer ChatGPT data is largely processed under US law, which offers weaker general privacy protection than the European Union's regime. This jurisdictional reality is the backdrop to the most significant regulatory action ChatGPT has faced.

In early 2023, Italy's data protection authority, the Garante, ordered a temporary suspension of ChatGPT in Italy, citing concerns about the legal basis for processing personal data, the absence of adequate age verification, and insufficient transparency. OpenAI restored the service after introducing changes including clearer privacy disclosures, an age-gate, and controls letting users object to processing. The matter did not end there: in late 2024, the Garante announced a fine against OpenAI over the same underlying issues, one of the first substantial GDPR enforcement actions against a generative AI provider. The episode, documented by the Garante, established that European regulators would treat large language model training data as personal data subject to GDPR, with all the consent, transparency, and data-subject-rights obligations that entails.

For European users, GDPR gives you rights that consumer ChatGPT must honour: access to your data, correction, deletion, and the ability to object to processing. OpenAI provides mechanisms for these requests, and complaints from the advocacy group NOYB – which has challenged ChatGPT's tendency to generate false personal information it cannot easily correct – keep pressure on how well those mechanisms actually work. The unresolved tension is structural: a right to rectification assumes you can fix a specific stored record, but a model does not store facts as editable records, which makes "correcting" a hallucinated claim about a real person genuinely hard.

For sensitive use cases

The safest rule for healthcare, legal, and financial contexts is the simplest: do not paste anything into consumer ChatGPT that you would be unwilling to see stored, reviewed by a human, or subpoenaed.

That means no patient identifiers, no protected health information, no client matter details covered by legal privilege, no account numbers, no full names paired with sensitive circumstances, no passwords or credentials, and no proprietary source code or trade secrets. Consumer ChatGPT is not a HIPAA-covered service by default, and pasting protected health information into it can itself constitute a compliance breach regardless of what OpenAI subsequently does with the text. Legal privilege is similarly fragile: sharing privileged material with a third-party service can, in some interpretations, waive the privilege entirely.

If the work genuinely requires an assistant that handles regulated data, the answer is the enterprise path with a business associate agreement or equivalent contract, appropriate retention configuration, and internal policy governing what staff may submit – not the free tier with the training toggle flipped off. The toggle reduces one risk; it does not make the tool appropriate for regulated data. Where a use case is high-stakes, the broader security picture matters too, and our companion analysis of ChatGPT safety, privacy and security covers the threat surface beyond data handling alone.

A pragmatic middle path exists for individuals who want more privacy without an enterprise contract: use Temporary Chat for sensitive one-offs, keep training turned off, redact identifiers before pasting (replace real names and numbers with placeholders), and periodically clear both history and Memory. None of this makes ChatGPT anonymous – your account, IP, and device metadata are still associated with your usage – but it materially shrinks the amount of durable, personally linked content sitting on someone else's servers.

Where this leaves you

Is ChatGPT private? The accurate answer is that it can be made more private, tier by tier and setting by setting, but it is not private by default and it is never anonymous. The consumer product optimises for convenience and model improvement; the business and API products optimise for control and non-training defaults; and the regulatory environment, led by European enforcement, is slowly forcing more transparency across all of them.

The decision that matters is not really about a single toggle. It is about matching the sensitivity of what you type to the tier and configuration you are typing into. For casual questions, the defaults are a reasonable trade. For anything you would not want stored, reviewed, or disclosed, the consumer tiers are the wrong tool no matter how you configure them, and the right answer is either the enterprise path or not using the tool at all. Verify every specific retention period, price, and policy detail against OpenAI's current documentation before you rely on it, because on this beat the numbers move, and the version you read last year is not the version governing your account today.