What happens to a prompt after you send it?
Your prompt, and any file you attach, goes to the AI provider's computers. The model writes an answer and sends it back. Along the way, copies can be made:
- Abuse monitoring logs. OpenAI says these are "generated for all API feature usage and retained for up to 30 days", unless you're approved for zero data retention (OpenAI).
- Backend storage. By default, Anthropic's API deletes inputs and outputs "within 30 days of receipt or generation" (Anthropic).
- Saved features. On OpenAI, assistants, threads, files and batches keep data "Until deleted".
- Chat history in the app you use.
Under zero data retention, the answer comes back and nothing is written to durable storage. AWS defines its setting that way: "No request or response data is written to durable storage by AWS or shared with the model provider" (AWS).
Where can an AI tool keep your data?
Even with zero data retention, check these places:
| Where | What the providers say |
|---|---|
| Chat history | Kagi keeps threads until 24 hours after the last message, or permanently. Proton stores saved Lumo history with zero-access encryption. |
| Model providers behind an app | Duck.ai's providers must delete data "at most within 30 days". |
| Abuse monitoring | On Azure, flagged prompts and completions can be stored for human review. |
| Caches | Google keeps an in-memory cache, cleared within 24 hours, even under zero data retention. Brave deletes cached large prompts "within minutes". |
| Metadata | Under zero data retention, Google "continues to record operational and billing metadata (such as token counts and latency telemetry)". |
| Feedback | Brave keeps feedback you choose to send for one year. |
Each provider's page is linked from its entry on Best Zero Data Retention AI.
How is zero data retention set up?
- Approval, then a setting. OpenAI and Anthropic review the customer first, and so does Google for OpenAI models on its platform. OpenAI then shows "a Data Retention tab within Settings → Organization → Data controls." Microsoft's closest option, modified abuse monitoring, also needs approval; afterwards you check that "ContentLogging" shows "false" (Microsoft).
- An account setting. On AWS Bedrock, you set data_retention_mode to none, Region by Region. If a model needs retention, "Amazon Bedrock will block the request and return an error."
- The design itself. Apple's Private Cloud Compute changes its storage encryption keys on every reboot, never keeps them, and has no "general-purpose logging mechanism" and no remote shell (Apple Security).
- An app in the middle, plus contracts. Duck.ai, Kagi and Venice's Private mode remove your identity and rely on agreements with the model providers. Venice calls its Private mode "Trust-based" (Venice).
What are the common exceptions?
- Child sexual abuse material. OpenAI keeps a flagged image "for manual review, even if Zero Data Retention ... is enabled." AWS may store and review flagged content only to check for it.
- Features outside the scope. Anthropic says features "marked 'No' in the table (such as code execution) are not covered."
- Some models. AWS keeps all traffic for some models for up to 30 days for abuse detection (AWS). OpenAI may make models ineligible for a customer, with written notice.
- Settings that sound similar. AWS says: "Setting store=false does not guarantee zero data retention." Google says the same of store: false in its Responses API.
- Features you can't switch off. Google keeps Grounding with Google Search logs for up to three days, and says "There is no way to disable" this (Google Cloud).
- Automated review. On Azure, even after approval, "automated review may still be conducted."
See also Zero Data Retention vs. No Training.