# What happens to your prompts at the provider
> Training, retention, abuse logs and EU hosting at OpenAI, Anthropic, Google and Microsoft as of November 2026, and why consumer apps and business tiers follow different rules.
Source: https://rfriedmann.de/blog/what-happens-to-your-prompts/
Published: 2026-11-20 · Track: learn · Level: Ops
The question I hear most often about AI tools at work is "do they train on our data?".
It is only half the question. A prompt that is never used for training can still sit in
a log for weeks, and a person at the provider can still read it. So I went through the
official policy pages of OpenAI, Anthropic, Google and Microsoft. Everything below is as
of November 2026; these pages change, so the links are the source, not my summary.
## Two tiers, two rulebooks
Every one of these providers runs two different sets of rules for the same models.
- **Consumer apps**: ChatGPT Free/Plus/Pro, Claude Free/Pro/Max, the Gemini app on a
personal Google account, consumer Copilot. You accept consumer terms as an individual.
Training on your chats is either on by default or a setting you choose.
- **Business tiers and APIs**: ChatGPT Business/Enterprise, the OpenAI API, Claude
Team/Enterprise and the Claude API, Gemini in Google Workspace, Vertex AI, Azure
OpenAI, Microsoft 365 Copilot. These run under commercial terms with a data
processing agreement, and training on customer content is off by default at all four.
The split matters more than the vendor. An employee pasting a contract into the free
ChatGPT app and the same contract sent through your company's API key are two different
legal situations, even though the model may be identical.
## Consumer apps
| Provider | Training on chats | Retention | Opt-out |
|---|---|---|---|
| OpenAI: ChatGPT Free, Plus, Pro | On by default ([help](https://help.openai.com/en/articles/8983130-what-if-i-want-to-keep-my-history-on-but-disable-model-training)) | Deleted chats are removed within 30 days; Temporary Chats are kept up to 30 days for safety ([help](https://help.openai.com/en/articles/8809935-deleting-and-archiving-chats-in-chatgpt)) | Settings, Data controls, "Improve the model for everyone". Thumbs-up/down feedback can still be used for training |
| Anthropic: Claude Free, Pro, Max | Your choice in the privacy settings ([announcement](https://www.anthropic.com/news/updates-to-our-consumer-terms)) | 5 years if you allow training, otherwise 30 days; chats flagged for policy violations up to 2 years ([privacy](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data)) | Privacy settings, model improvement toggle |
| Google: Gemini app, personal account | Depending on your settings, chats train generative models; a subset is read by human reviewers ([privacy hub](https://support.google.com/gemini/answer/13594961?hl=en)) | History auto-deletes after 18 months by default; reviewed chats are kept up to 3 years, even after you delete them; with History off, chats are still kept about 2 months | Turn off Gemini History |
| Microsoft: Copilot, personal account | Used for training unless you opt out or fall into an excluded group ([FAQ](https://support.microsoft.com/en-us/microsoft-copilot/privacy-faq-for-microsoft-copilot)) | Conversation history stored 18 months by default; human review cannot be opted out | Privacy settings, "Training on conversation activity" |
Two details stand out. Google keeps chats a human reviewer has seen for up to three
years, and deleting your history does not remove them. Microsoft's FAQ says "an opt-out
of human review is not available"; its training section covers the Copilot app before
18 August 2026, so check the setting in the app you actually use.
## Business tiers and APIs
| Provider | Training | Retention and abuse logs | Zero data retention | EU option |
|---|---|---|---|---|
| OpenAI: API, ChatGPT Business, Enterprise, Edu | Not used by default ([enterprise privacy](https://openai.com/enterprise-privacy/)) | API abuse monitoring logs kept up to 30 days, longer if the law requires it ([your data](https://developers.openai.com/api/docs/guides/your-data)) | ZDR or Modified Abuse Monitoring, after approval by OpenAI | API region Europe (EEA and Switzerland) with regional storage and processing, chosen per new project ([data residency](https://developers.openai.com/api/docs/guides/data-residency)) |
| Anthropic: API, Claude Team, Enterprise | Not used on customer content under the Commercial Terms ([commercial terms](https://www.anthropic.com/legal/commercial-terms)) | API inputs and outputs deleted within 30 days, flagged content kept up to 2 years ([privacy](https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data)); the API docs say conversation content is not retained by default, except for "Covered Models" (Fable and Mythos 5 family), which require 30 days ([docs](https://platform.claude.com/docs/en/manage-claude/api-and-data-retention)) | By arrangement via sales; Covered Models only if Anthropic authorizes it | Own API: inference only "us" or "global" ([data residency](https://platform.claude.com/docs/en/manage-claude/data-residency)); EU regions through Bedrock, Google Cloud or Microsoft Foundry ([regional compliance](https://claude.com/regional-compliance)) |
| Google: Gemini API (paid), Vertex AI, Workspace | Paid API: not used to improve products ([terms](https://aistudio.google.com/docs/terms)); Vertex AI and Workspace: not without your permission ([Vertex](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/vertex-ai-zero-data-retention), [Workspace](https://knowledge.workspace.google.com/admin/gemini/generative-ai-in-google-workspace-privacy-hub?hl=en)) | Gemini API: 55 days for abuse detection ([policy](https://aistudio.google.com/docs/usage-policies)); Vertex AI: flagged prompts up to 90 days, only for customers on the online terms ([abuse monitoring](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/abuse-monitoring)) | Vertex AI: customers with a Google Cloud Master Agreement are exempt from prompt logging; some features still store data | Vertex AI "eu" multi-region endpoint: EU member states only, not the UK or Switzerland ([data residency](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/data-residency)) |
| Microsoft: Azure OpenAI in Foundry, Microsoft 365 Copilot | Not used to train foundation models ([Azure](https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy), [M365](https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-privacy)) | Azure: content selected for human review is stored in your resource's geography, no fixed period stated; M365 Copilot has opted out of abuse monitoring, retention via Purview | Azure: Modified Abuse Monitoring for customers who meet Limited Access criteria ([abuse monitoring](https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/abuse-monitoring)) | Azure DataZone in the EU: processing in EU member states; M365 Copilot is an EU Data Boundary service, but Anthropic models in Copilot are excluded |
One useful quirk: Google's Gemini API terms apply the paid-tier data rules to free usage
for anyone in the EEA, Switzerland or the UK. Outside those, the free tier lets human
reviewers read your input and output.
## What "not used for training" does not mean
The business tiers all promise no training. That promise covers one arrow in the
diagram below. The others remain.
Where a prompt goes after you press Enter
[diagram omitted — see the page for the chart]"Not used for training" closes the right-hand path. The abuse monitoring log and human review stay unless you have zero data retention or a modified monitoring agreement, and your own logs are always your job.
Concretely, "not used for training" still allows:
- **Storage.** OpenAI keeps API abuse logs up to 30 days, Anthropic deletes API data
within 30 days at most, Google keeps Gemini API prompts 55 days. Your data exists on their
disks during that window.
- **Human reading.** All four describe staff who can look at flagged content. Microsoft
documents that, for Azure deployments in the EEA, those reviewers sit in the EEA. Ask
the others the same question.
- **Legal holds.** OpenAI's API retention applies "unless longer retention is required by
law". A court order can keep data that would otherwise have been deleted.
- **Stateful features.** Files, threads, batch jobs and stored responses follow their
own clock. Anthropic keeps batch data up to 29 days; Vertex AI lists grounding and research
features whose storage cannot be turned off.
None of this makes the business tiers unsafe. It means "no training" is the start of the
conversation with the provider, not the end.
## Rules I use for a company
1. **Business tier with a signed DPA, nothing else.** Company data goes only through
accounts under commercial terms with a processor agreement. That is also what the
GDPR asks for; [the legal side has its own lesson](/blog/ai-and-the-law-gdpr-ai-act/).
2. **No secrets in consumer apps.** Not on a personal account, not "just this once". An
employee cannot sign a DPA for the company.
3. **Pick the region on purpose.** EU processing exists at every provider, but often only
on a new project, a specific endpoint or deployment type, or a cloud platform.
4. **Ask for ZDR where the data is sensitive**, and read the exclusion list that comes
with it.
5. **Count your own logs.** Proxies, gateways and chat histories in your own tools keep
prompts too, under your retention rules.
6. **Settle output rights separately.** Who owns what comes back is a different clause,
[covered here](/blog/who-owns-ai-output/).
If the data is too sensitive for any of this, the remaining option is a model on your own
hardware. Then the abuse log and the training path simply do not exist, which is the
main reason [I run models locally](/blog/why-run-ai-locally/) for some work.
## Bottom line
All four providers keep business and API data out of training by default and treat
consumer apps differently. "No training" still leaves abuse monitoring storage, possible
human review and your own logs. Use the business tier with a DPA, choose the EU region
explicitly, ask for zero data retention where the data warrants it, and keep company
secrets out of personal accounts.