What Happens to Your Prompts After You Send Them?

Short version: yes, it usually gets saved, at least for a while, and whether anyone sees it depends entirely on which product you typed it into. A prompt sent through a paid API call and a prompt typed into a free consumer chat app go down two different pipes with two different retention clocks, and the difference is bigger than most people assume when they paste a client contract or a database schema into a chat window. This post traces what happens to your prompts after you send them to an ai, using the actual current data-handling policies of two major providers, not the vague reassurances in their marketing copy.

That's the whole answer for casual use. The rest of this post is the plumbing behind it, because the plumbing is where the actual risk lives.

The path a prompt actually takes

Every prompt goes through roughly the same five stops, whichever provider you use. What differs is how long the data sits at each stop and who is allowed to look at it.

• Input. Your text (and any images, files, or system instructions attached to it) leaves your device over TLS and arrives at the provider's inference infrastructure.

• Inference. The model reads the prompt, generates a response, and the request-response pair exists in memory on the serving machine for the length of that call.

• Logging. Separately from inference, most providers write a copy of the input and output to a log store, usually for abuse monitoring and debugging, not for the model itself to learn from in that moment.

• Training-data selection. Whether that logged copy is later used to train or fine-tune future models is a policy decision, not a technical inevitability, and it is the single biggest variable between providers and between plans.

• Retention and deletion. Logs sit for a defined window, then get deleted, unless something flags them for longer holds, such as a suspected policy violation or a legal request.

Every step is configurable at the provider level, and on business and API tiers, often by you. The defaults are what most people never check.

Does AI save my prompts? It depends which door you walked through

There are two doors into the same model family, and providers treat them differently:

• Consumer chat apps (ChatGPT free/Plus, Claude Free/Pro/Max, Gemini's free tier) typically save your conversation history by default and, in several cases, may use it to improve future models unless you turn that off.

• Developer APIs (calls made through an app you or someone else built) run under separate, generally stricter terms: providers commonly commit that API inputs aren't used for training by default, with a shorter, narrower logging window kept mainly for abuse detection.

This is why the same underlying model can feel more or less trustworthy depending on the product wrapped around it. An app built on a provider's API is usually operating under the API's data terms, not the consumer chat app's terms, even though it's calling the same model, and that includes tools like AI coding assistants and customer-facing chatbots built on top of these APIs.

What OpenAI's API actually does with a prompt

OpenAI's developer documentation states that data sent to the API has not been used to train or improve OpenAI's models since March 1, 2023, unless a customer explicitly opts in. On the logging side, the default is to retain API inputs and outputs for up to 30 days, solely for abuse and misuse monitoring, after which they're deleted (longer only if law requires it or in an active abuse investigation).

Eligible business customers can additionally apply for Zero Data Retention, which excludes prompts and completions from those abuse-monitoring logs entirely, and in August 2026 OpenAI expanded that Zero Data Retention option to cover more of its frontier models for qualifying API customers. See OpenAI's own data-controls documentation: Data controls in the OpenAI platform.

Consumer ChatGPT (Free and Plus) runs the opposite default: content from those tiers can be used to improve the product unless you opt out in settings.

What happens to a prompt sent to Claude's API

Anthropic's platform documentation lays out a similar but distinct structure. For the Claude API, conversation content, meaning your prompts and Claude's replies, is not retained by default beyond what a specific feature technically requires, and any data that is retained is never used for model training without your explicit permission.

Anthropic's commercial data-retention policy specifies that where storage is needed, inputs and outputs are automatically deleted from the backend within 30 days of receipt, with narrower exceptions: content flagged for a Usage Policy violation can be held up to 2 years, and the associated trust-and-safety classifier scores up to 7 years. Eligible organizations can also negotiate a Zero Data Retention arrangement, under which Anthropic doesn't store prompts or responses at rest once the API response is returned. See Anthropic's own retention documentation: API and data retention.

Consumer Claude.ai again runs a different default: as of Anthropic's 2025 policy update, conversations can be used for training unless a user opts out, and opted-in data can be retained in de-identified form for up to five years.

A third data point: Google's Gemini API

Google's Gemini API documentation shows the same pattern again: on billing-enabled (paid) projects, prompts and responses are not used to improve or train Google's products by default, though logs are kept up to 55 days for abuse detection. On the free tier, the terms flip: submitted content can be used to improve Google's products, including by human reviewers. See Data logging and sharing on Google AI for Developers.

Three companies, three retention windows, one identical structural pattern: paid and business tiers default to no training use and a short, bounded log window; free consumer tiers default to broader use of your content unless you go into settings and turn it off.

Is it safe to type sensitive info into AI?

Given the above, a practical rule of thumb:

These terms change (Anthropic's training-opt-in policy shifted in 2025, OpenAI expanded Zero Data Retention again in August 2026), so treat a live product's current settings page as the source of truth, not a blog post from six months ago, including this one. Policy changes are also worth watching for their own sake: see what to check when an AI vendor changes its terms of service. A simple working rule in the meantime: if a piece of information would be sensitive in an email to the wrong person, apply the same caution before pasting it into a prompt. The model itself doesn't gossip, but the logging and review layer behind it is operated by people, and people occasionally see flagged content as part of their job.

Frequently asked questions

Does AI save my prompts?

Usually yes, at least temporarily. Consumer chat apps typically save conversation history to your account by default. Developer APIs commonly log inputs and outputs too, but for a shorter, narrower window (commonly around 30 days) aimed at abuse monitoring rather than product features, and some providers offer a zero-retention option for eligible business customers.

What is a typical AI data retention policy in 2026?

There's no single industry standard, but the common pattern is: business and API tiers skip training use by default and delete logs within roughly 7 to 55 days depending on the provider, while free consumer tiers may use content to improve products unless you opt out, and retain opted-in data much longer.

Is it safe to type sensitive information into an AI chatbot?

It depends on the tier and the provider's current policy. It's generally lower-risk on a paid or business tier with training off and a short retention window, and highest-risk on a free consumer tier with training use turned on. When in doubt, check the specific product's data controls before pasting anything you wouldn't put in a normal work email.

Can an AI company's employees read my prompts?

It's possible, though not routine. Prompts flagged by automated abuse-detection systems, or involved in a legal request, can be reviewed by trust-and-safety staff regardless of the provider's normal retention policy. On some free consumer tiers, human review is also part of the standard product-improvement process unless you've opted out. The same logging layer is also why a chatbot can end up leaking a customer's data if it's misconfigured to retain or expose more than it should.

Do AI providers delete my data if I delete a conversation?

Deleting a conversation in the app usually removes it from your visible history, but backend systems commonly keep a copy for a defined grace period, often up to 30 days, before it's permanently purged, unless it's been flagged for a longer hold. That routine deletion window is a different question from what happens to your data if a vendor shuts down entirely, which is its own risk worth planning for separately.

For the broader picture of what can go wrong once you've handed sensitive work to an AI system, see the AI safety and risk overview, and for how AI agents with their own accounts change this calculus further, whether an AI agent should have its own user account.