What happens to the data you put into an AI chatbot?
Follow one paragraph.

It leaves your computer, gets processed on the provider's servers, and is stored for a period the provider chooses. Depending on your plan and settings, it may also be used to improve future models, and a sample may be read by human reviewers. None of that is hidden — the providers publish it. Almost nobody reads it.

So let's read it. Below, we follow a single pasted paragraph through each stage, quoting the providers' own documentation as it stood when we checked in September 2026. Terms change; the links go to the live pages.

Stage 1: It leaves the building

You paste a paragraph and press Enter. The text travels over an encrypted connection to a data center and is processed on hardware shared with the provider's other customers. This part is unavoidable: a cloud model can't answer a question it never receives. Everything after this point is a policy choice made by the provider, and it varies by product and plan.

Stage 2: It gets saved

Your conversation is normally stored with your account so you can scroll back to it. How long it stays is the provider's decision, with some settings left to you. Google's Gemini Apps Privacy Hub, for instance, describes an auto-delete setting that can be changed "from the default of 18 months to 3 months, 36 months, or indefinite" and says that "Temporary chats and chats you have when Keep Activity is off are retained with your account for 72 hours". (Google: Gemini Apps Privacy Hub(opens in new tab))

Anthropic ties retention to a choice the user makes about training: "We are also extending data retention to five years, if you allow us to use your data for model training. … If you don't choose this option, you will continue with our existing 30-day data retention period." That applies to its consumer plans — Free, Pro, and Max — and not to services under its commercial terms. (Anthropic: Updates to Consumer Terms and Privacy Policy(opens in new tab))

Stage 3: A person might read it

This is the stage that surprises people. Some providers use human reviewers as part of improving their services. Google says so plainly: "Human reviewers (including trained reviewers from our service providers) review some of the data we collect" for the purposes it lists. It also describes a safeguard — "Chats are disconnected from your account before being sent to service providers" — and gives a piece of advice worth taping to the monitor, which begins: "Please don't enter confidential information that you wouldn't want a reviewer to see".

Disconnecting a chat from your account is a real protection. It doesn't help much if the paragraph you pasted has your client's name and address in it.

Stage 4: It may help train the next model

Whether your text is used for training depends on the product, the plan, and a setting. On Anthropic's consumer plans it's an explicit choice: "We will train new models using data from Free, Pro, and Max accounts when this setting is on" and "You can change your selection at any time in your Privacy Settings." Business products often make the opposite commitment. Microsoft, for one, says that under enterprise data protection "the prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation models." (Microsoft: Enterprise data protection in Copilot(opens in new tab))

"Used for training" doesn't mean another customer can look up your document. It means your text can become one small part of the material a future model learns from. The practical problem is simpler than the technical one: once that has happened, you can't take it back, and you can't tell your client exactly where their information went.

Stage 5: You press delete

Deleting a conversation removes it from your history. Whether it removes it from everywhere is a separate question with a published answer. Anthropic: "If you delete a conversation with Claude it will not be used for future model training." Google, on conversations that were already selected for human review: they "are not deleted when you delete your activity. Instead, they are retained for up to three years."

Neither statement is alarming on its own. Both are the kind of detail you'd want to know before the paragraph you pasted was somebody's medical history.

What about ChatGPT specifically?

OpenAI answers several of these directly. On consumer products: "When you use our services for individuals such as ChatGPT and Codex, we may use your content to train our models." On business products: "By default, we do not train on any inputs or outputs from our products for business users, including ChatGPT Business, ChatGPT Enterprise, and the API." Consumer users can turn training off under Settings → Data Controls. (OpenAI: How your data is used to improve model performance(opens in new tab))

One detail is easy to miss. Feedback is an exception to opting out: "If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models." For business products OpenAI keeps a separate enterprise privacy page(opens in new tab). Read both with the five stages in mind.

Check your own settings in ten minutes

1. Find out which plan you're on. Personal email or company account? Free, paid individual, or a business agreement? This decides which document applies to you.
2. Open the privacy or data controls page. Look for a model-training or "improve the product" setting and note where it's set.
3. Find the history or activity setting and its auto-delete period.
4. Search the provider's privacy page for "review" and "retain." Those two words lead to the paragraphs that matter.
5. Ask who else in the office is using what. One careful person doesn't make a careful company.

It usually wasn't only your data

Here's the part the settings page can't fix. The paragraph you pasted probably wasn't about you. It was a tenant's application, a customer's complaint, an employee's review, a client's numbers. You can consent to a provider's terms on your own behalf. You can't consent on theirs — and if there's a confidentiality clause in your contract with them, a checkbox in your chatbot settings doesn't amend it.

The same paragraph, on a machine you own

Run the model in-house and the five stages collapse into one. The text travels from your keyboard to a GPU in your own building and the answer comes back. It's saved where you save it, for as long as you decide. Nobody reviews it unless you ask them to. It trains nothing. Delete means delete, because it's your disk.

We run this way ourselves: the assistant on this site answers from a single 24GB graphics card in our office, not from a cloud account. That setup isn't necessary for a blog draft or a recipe. It starts to matter the moment the paragraph belongs to someone who trusted you with it.