Privacy in AI assistants: what is stored and how to limit it

When someone asks whether an AI assistant “saves” a conversation, they often conflate at least three questions: whether the chat appears in the account’s history, how long the service’s systems retain the data, and whether the content may be used to improve products or models. These are distinct forms of processing. Changing a history setting does not, by itself, prove that every technical copy is erased immediately or that the content is excluded from every other use.

It also matters what was sent beyond the text itself: files, images, instructions, account data and usage signals may be governed by different rules. Google’s privacy policy, for example, describes categories of information its services collect and associated uses; it is not an account exclusively of Gemini, so it should be read alongside the assistant’s specific documentation. A useful answer requires identifying the product, the type of data and the specific setting, rather than treating “privacy” as a single switch.

The distinction between visible history and retention is especially important when deleting a conversation. An item disappearing from the interface confirms an action affecting the user’s history, but it does not, by itself, establish that every technical record, backup or data subject to legal obligations has disappeared at that same moment. This limitation does not prove that a particular copy is retained: it identifies what the interface alone cannot confirm.

Google’s documentation provides a general privacy policy and a Privacy Checkup tool for reviewing some account settings. The policy helps explain the practices the company describes across its services; the tool lets users check and change options available to their account. Neither one, taken alone, serves as a complete, product-specific inventory of every Gemini mode or region. Compare the general policy with the product’s current help documentation before drawing conclusions about conversations, activity or files.

The evidence available for this article does not include current official policies from OpenAI, Anthropic or other providers that would allow their rules to be compared precisely. The OpenAI URL provided refers to a review of its policy dated June 23, 2023, and is not a sufficient source for describing ChatGPT’s current terms. Nor does a question posted in Google’s help community amount to an official answer: the cited threads are user contributions and do not, by themselves, verify a retention or human-review rule.

For that reason, no specific conversation-retention periods are attributed, no setting is presented as a universal training opt-out, and services are not compared on the basis of incomplete data. A privacy policy and a product control address related but different questions: the former sets out general practices and conditions; the latter lets users act on some preferences. Exact options, their names and their effects may vary by product, account type and country.

Before sharing information, check the settings directly in the app and in the linked account. Look for controls related to activity or history, the use of conversations to improve the service, review or personalisation, and read the help information associated with each option. If a setting does not explain how files or sensitive data are handled, do not assume that it covers them: the documentation should specify its scope.

A practical review can follow this order:

• Identify the service, account and region; check that the help page you consult applies to that combination.
• Read what each option controls: visible history, activity, use to improve the service and data retention are not synonyms.
• Check whether shared conversations include files, third-party data or unnecessary personal information.
• Use the deletion or export tools documented by the provider and keep any confirmation available.
• Recheck the settings after product or account changes; do not assume that one preference applies to every service from the same provider.

Google’s Privacy Checkup is a starting point for general Google Account controls, not a guarantee that every Gemini-specific setting is covered on a single screen. Similarly, deleting a conversation within a product and managing account data may be separate actions. Check the outcome where the data was created and read what the provider says about the scope of deletion; do not infer immediate removal from systems the tool does not describe.

A policy may describe purposes, data categories and control mechanisms, but it does not, by itself, reveal every internal operation applied to a particular conversation. Without specific, current documentation, it is not responsible to claim who reviewed a chat, whether a human reviewed it, how long deletion took or what auxiliary copies exist. Questions about these points should remain open rather than be filled in with assumptions or answers from forums that are not official sources.

It is also worth distinguishing processing to provide the service from using data to train or improve systems. A provider’s processing an input in order to respond does not automatically prove that it uses the input for training; conversely, turning off a particular setting does not prove that every form of processing needed to operate the service has stopped. The consequence depends on the documented wording and scope for that product, and the absence of an accessible explanation limits what can be concluded.

Deletion should not be described as an absolute technical promise unless the source says so. There may be exceptions, time limits or categories of records that a simplified tool does not explain; this is a limitation of the available evidence, not a claim that such exceptions have applied to a particular account. If you need certainty about a specific situation, consult the provider’s policy and official privacy contact, and avoid sending data whose handling you cannot accept.

In the European Union, the European Data Protection Board (EDPB) published an opinion on December 18, 2024, on the use of personal data to develop and deploy AI models. Among other points, the document addresses when and how a model may be considered anonymous, the possible use of legitimate interest as a legal basis, and the consequences of data used in its development having been processed unlawfully. It is guidance on data-protection issues relating to AI models, not a rule that automatically determines how long each assistant retains each conversation.

The EDPB’s topic page describes artificial intelligence as an area with potential benefits and data-protection issues, and brings together the authority’s work and materials. That context does not replace reading the applicable law, the service terms or the documentation for a specific control. The existence of legal obligations does not, without further information, let you infer an account’s settings or the fate of an individual chat.

As of the date of this article, the EDPB also has an open consultation on its Guidelines 03/2026 concerning web scraping in the context of generative AI, with a comment period listed from July 8 to October 30, 2026. This is a consultation on collecting web data, not a policy on retaining conversations in assistants. Likewise, the EDPB announcement of July 8, 2026, on guidelines for anonymisation and web scraping should not be confused with a specification of user controls for a commercial service.

The practical recommendation is not to stop using assistants, but to match what you share to the sensitivity of the data and the clarity of the terms. Avoid entering passwords, credentials, financial data, identifiable medical information or confidential content about third parties when it is not essential. If you need help with a text, replace names and identifying details before pasting it, and check attachments too, in case they contain information that is not visible at a glance.

Before relying on a control, look for verifiable answers to three questions: what data it covers, which uses it prevents, and what happens to material already sent. If the policy does not answer one of them, the correct response is that this cannot be concluded from that source. Privacy options are useful, but they do not automatically turn a conversation into professional confidentiality or a guaranteed secret.

The documented picture here supports a practical approach, not a ranking of providers: consult the current policy for the exact service, verify settings from your own account, and treat any claim about deletion, exclusion or review as specific and limited to what the source confirms. The available documentation is not enough to present a recent change shared across the industry; this is therefore a verification guide, not an announcement of new rules.