Encryption, training and deletion are different controls

A conversation can be encrypted as it travels over the internet and still remain stored when it reaches the provider. Encryption protects content against certain kinds of access during transmission; it does not, by itself, determine how long data is kept, who can process it or whether it is used to improve models.

It is also useful to distinguish three settings that are often presented together: using conversations for training, the history visible in your account, and internal retention needed to provide the service or address security matters. Changing one does not automatically change the others. Turning off training may prevent future conversations from being used for that purpose without deleting them immediately; deleting your history does not prove that every internal copy has disappeared at once.

The Spanish Data Protection Agency (AEPD) advises people not to enter personal or sensitive information into AI tools, such as medical, financial or contractual details, and to follow their employer’s security policies when handling work information. Its practical guidance starts with a simple rule: share only what is necessary, even when a service offers privacy options.

The distinction matters because interfaces tend to present understandable controls—such as “improve the model” or “save activity”—while retention may depend on terms, account type and security or legal exceptions. Users can limit certain uses and keep less history, but a toggle should not be read as a universal promise of deletion.

What to check before changing settings

Start by identifying the specific service and account: personal, work, education or developer. Rules can differ among plans and products from the same company; moreover, a chatbot’s settings do not necessarily govern data in other applications connected to that account.

Look separately for controls for training, history or activity, temporary chats, deletion and export. Read the explanation beside each option and note whether it applies to future conversations, keeps data for a period or includes exceptions. Scope matters more than a setting’s name: “turn off activity” may affect different uses from “delete conversations.”

Also check permissions for connected apps, files, the microphone and the camera. Spain’s National Cybersecurity Institute (INCIBE) advises limiting permissions to what is necessary, protecting accounts with a strong password and two-step authentication, and reviewing options regularly because services can change.

To reduce risk without depending on a particular setting, replace names, addresses, case numbers and other identifiers with fictional details. For work-related questions, confirm that your organization authorizes the service before uploading customer information or internal documents. Anonymization reduces exposure, but it does not make every text harmless: a combination of distinctive details can still identify someone.

ChatGPT: turn off use for model improvement

In ChatGPT, account data controls let users turn off the option called “Improve the model for everyone.” The route may vary slightly between app and website versions, but the starting point is the profile menu and the data settings section. Check the setting in the account you actually use; do not assume that changing it in one profile affects another.

Wired’s coverage of OpenAI’s controls describes opting out of training for new conversations through this option. So if your goal is to limit future training, change the control; if you also want to remove chats from your account view, manage the history separately.

OpenAI also offers temporary chats, but check the product’s current help information for availability and retention terms before entering sensitive information. A temporary chat should not be treated as a confidential channel or as a substitute for your company’s security rules, professional advice or a service designed for protected communications.

After changing the preference, return to the screen and confirm that it remains off. Avoid sending a sensitive conversation as a test: you can check the control’s status without exposing real information. If you switch between web and mobile, check the account and setting in the session you use; do not assume one toggle deletes data that has already been processed.

Gemini: activity and conversation retention

In Gemini, activity settings let users manage the saving of interactions associated with their account. The interface may call this Gemini Apps Activity or use an equivalent name. Wired’s guide explains how to turn off that activity and notes that data already reviewed by people may be handled differently; turning off the setting does not necessarily delete all earlier history.

Before changing the setting, check which account is open, what activity will be retained and what deletion options are available. Retention periods and handling can depend on the configuration and specific product; do not apply a condition described for one version or period to every account.

Then review your history and use the account tools to delete conversations you no longer need. That step may remove chats from view or start a deletion process, but it does not guarantee that every technical copy is erased at the same moment: providers may apply timeframes or exceptions set out in their policies.

If Gemini is connected to other Google services, also review permissions and enabled apps. The chatbot’s history control is not necessarily the same as disabling access to email, files or other Workspace data. For work or customer information, use only the settings and environment authorized by your organization.

Gemini: actividad y conservación de conversaciones En Gemini, la configuración de actividad permite gestionar el guardado de interacciones asociado a la cuenta. En la interfaz pued...
Editorial illustration by BytePreview

Claude and Copilot: don’t generalize across plans

Anthropic has described controls that allow users to manage whether their conversations may help improve Claude. In Wired’s coverage, the company said it does not use user-submitted data for model training by default and may receive voluntary feedback on responses. Check the current privacy option in your account: controls can vary by product and plan.

Claude also distinguishes model-improvement use from retention needed to operate the service. In September 2026, Bolsamania reported a change announced by Anthropic for enterprise customers: a solution offering corporate controls over data review and management. That announcement concerns an enterprise setting and does not establish that the same terms apply to individual users or every model.

For Microsoft Copilot, Wired reported that personal accounts did not then have an option to exclude queries from service improvement, while work accounts may be subject to different terms. Interfaces and policies can change, so check the documentation that applies to your plan before sharing data. Do not transfer instructions for a personal account to a business one, or the other way around.

For any assistant, identify the exact product—general chat, an integrated app or an enterprise service—and the policy date. If you cannot find a specific control, do not assume an equivalent opt-out exists: minimize the data and, where relevant, ask your organization’s administrator.

Alternatives and the limits of privacy promises

Some tools describe architectures or policies that differ from those of general-purpose assistants. Venice, for example, describes options that include anonymized processing, open-source models hosted by the service and protected execution environments, depending on the selected model. Protection is not identical across every route: check which provider processes the request and what level applies to the specific feature.

Duck.ai, from DuckDuckGo, has also been presented as an interface intended to limit the information sent to model providers. An academic description hosted by the University of Barcelona highlights its privacy approach and access to third-party models, but it does not replace the service’s current terms. An intermediary tool adds a layer between the user and the model; that alone does not mean content cannot be identified or retained.

Running a model locally can reduce the need to send prompts to a remote provider, provided the application and model really run on the device and do not transmit data through connected features. Capabilities, hardware requirements and ease of use vary in return. “Local” describes where the model runs, not a complete guarantee about every component of an application.

When comparing alternatives, look for concrete answers to these questions: what data leaves the device, who operates the model, how long data is retained, whether it is used for training, what technical logs are kept and what happens when the account is deleted. Privacy claims describe commitments or designs; their scope depends on configuration and implementation.

Checklist for reducing exposure

A brief, regular review is more useful than assuming settings will stay the same. Menu names and terms can be updated, so repeat the steps after significant changes to an account, product or policy.

  1. Identify the account and product you plan to use, including the plan and connected apps.
  2. Check training, history, activity, temporary chats and deletion separately.
  3. Confirm the setting’s status and read its explanation of scope and retention.
  4. Remove identifiers and sensitive data before sending a prompt; do not upload work information without authorization.
  5. Check file and app permissions, secure the account and review settings regularly.

The practical choice should match the content: for general questions, minimizing personal details and reviewing controls may be enough; for personal or sensitive information, including medical, financial or contractual details, follow the AEPD’s advice not to enter it into AI tools. For work information, follow your organization’s security policies. If a task requires sensitive details, use only a service approved for that purpose and confirm its terms with the responsible organization.