What Data Do You Give an AI When You Type?

Typing into a chat assistant feels private, roughly like writing in a notebook. It is closer to sending email to a company you cannot see, where the terms decide what happens next. Your text leaves your device, crosses the network, is processed on somebody else's computer, and may be logged, reviewed or used to train another model. Here is what is at stake and how to reduce it.

Where your text goes

Every prompt is transmitted to a remote server unless you run the model locally on your own hardware. The provider has to receive it to generate a reply, and most providers keep some record to operate the service, debug faults, prevent abuse and bill accounts. Retention periods vary widely, from a stated short window to an indefinite archive tied to your account.

Deleting a chat from your own interface usually removes it from view. Whether it is gone from backups, safety logs and abuse records is a different question, and provider documentation is often vague on exactly that point. Exporting your data and then formally requesting deletion is the only route some providers offer, and it is worth doing before anything important depends on that account.

What the settings actually change

Consumer accounts and business accounts often behave differently even on the same provider. Business and enterprise tiers commonly state that submitted data is not used for training and is kept for a defined period. Free and consumer tiers may use conversations to improve models unless you switch the setting off. Temporary or incognito chats are still sent to the server to be answered; they are simply not kept in your history.

Before you paste anything sensitive, find three facts in the provider's documentation: whether your inputs train future models, how long they are retained, and whether human reviewers can read them. If a page hides all three, plan for the worst case.

What should never go in a chat box

Full names paired with health details. Other people's personal data, especially without their permission, because their consent is not yours to give. Financial account numbers and tax identifiers. Source code under a confidentiality agreement or an unreleased product plan. Privileged legal matters. Passwords, keys and access tokens. Anything your employer classifies as confidential. Student records and anything about a child.

For workplace use the test is simple. If it would not go in an email to an outside supplier, it should not go into a general chat assistant. Where the task genuinely needs that material, the answer is an approved internal tool with a data agreement behind it.

The data that travels with your message

Text is not the only thing transmitted. Photos often carry hidden metadata such as the location and the device that took them. Prompts carry account identifiers, timestamps, your network address and sometimes approximate location and device type. Uploaded files are stored at least long enough to be processed, and in many products they are indexed so that later retrieval works, which means the file is kept rather than merely read. Links you ask a model to open may be fetched from the provider's servers, which reveals the address to whoever runs the destination.

None of these details are usually secret on their own. Combined over months they build a detailed profile, and they persist in logs even after the visible conversation has been deleted.

Two different data questions

People mix up two separate issues. The first is the training data, the large corpus collected before the model existed, which is the source of the consent arguments and the copyright lawsuits. The second is your data, the messages and files you supply now, which may become part of a future model's training set.

Both matter, but the second is the one you control today, and it is the one most people ignore. The first is a policy argument you can support or oppose. The second is a habit you can change in the next thirty seconds, and habits beat terms of service, because policies get revised quietly and nobody sends you an email when they change.

Before you press send

• Check the retention setting - Know whether your inputs train future models and how long they are kept.
• Strip the identifiers - Replace names, account numbers and addresses with placeholders before pasting.
• Use the right account - Keep work on the approved tier rather than a personal login.
• Avoid sensitive categories - Health, legal, financial and other people's data do not belong in a general assistant.
• Clean your uploads - Remove metadata and redact documents before you attach them.
• Assume permanence - Write as if the text will be stored, read by a person, and quoted back to you later.