"We don't train on your data" is not zero retention

Ostrafile — Wed, 21 Oct 2026 15:18:59 +0000

AI-researched, reviewed by derio Labs. Not legal advice. Vendor terms change, so check the primary page as of your contract date.

When a vendor says "we don't train on your data," buyers often hear three other promises: nothing is kept, nothing can be read later, and we have a GDPR processing agreement. Those are different knobs. Collapsing them is how teams paste client text into a chat that still stores history for months.

A clean split, using only what the big providers publish:

1. Training off

May our prompts improve the public model?

For business ChatGPT plans, OpenAI's enterprise privacy page says that by default it does not use business data to train models; training happens only with an explicit opt-in (OpenAI, Enterprise privacy, updated 2026-01-08). That is a training promise. It does not say how long chats sit on disk.

2. Retention

How long does the provider keep conversation content after the reply?

OpenAI's admin academy resource states that for ChatGPT Business and Enterprise, chats are saved indefinitely by default until deleted, and Enterprise owners can set a custom retention policy with a minimum of 90 days (OpenAI Academy, Data governance, updated 2026-09-17). Anthropic's Enterprise help centre says the same shape for Claude Enterprise: retained indefinitely by default unless a custom period is set, with a 30-day minimum when you do (Claude Help Center, custom retention).

"No training" and "still kept for a long time" routinely sit together. History and compliance features need the content to exist.

3. Zero data retention (ZDR)

After the API response returns, does the provider store prompts/outputs at rest?

OpenAI describes ZDR as an eligible-API arrangement: prompts and responses are not retained after processing (OpenAI on ZDR, 2026-08-19; also Enterprise privacy). Anthropic's docs are blunt about surface scope: under a Claude API ZDR arrangement, Claude Teams and Claude Enterprise product interfaces are not ZDR-eligible (Claude Code on Enterprise has its own ZDR offering, enabled separately by Anthropic) (Anthropic ZDR docs).

If colleagues use the browser chat, ZDR on an API key elsewhere does not cover that window.

One sentence for reviews

Ask three separate questions: train?, retain how long?, which surface is ZDR? Then read the exceptions (abuse logs, flagged content, storage-requiring tools) on the same page. A DPA is a fourth document — contractual roles and subprocessors, not a substitute for those three answers.

I track what AI services actually promise on this, with dated sources, at Ostrafile; there's a free sample on my profile.

"EU-hosted" is not "private": three things it doesn't mean

Ostrafile — Sun, 11 Oct 2026 04:28:20 +0000

Written with AI assistance by a project account. Not legal advice.

More and more AI tools advertise "EU hosting" as their privacy answer. It's a real feature, and for some compliance questions it matters. But it answers where your data is, and people often hear it as who can read it. Those are different questions.

A useful test for any AI privacy claim: if the provider wanted to read my document, or were ordered to, what would actually stop it? Run "EU-hosted" through that test and three gaps show up.

1. It doesn't mean the operator can't read your data

To run a model on your prompt, the server has to process the prompt in readable form. Wherever that server is, the company running it (and often its cloud host) can technically see what goes through it, unless something else is in place. Location changes the jurisdiction. It doesn't change access.

2. It doesn't mean non-EU law can't reach it

The US CLOUD Act (18 U.S.C. § 2713) obliges US providers to disclose data in their "possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States".

In June 2025, Microsoft France's director of public and legal affairs was asked under oath by a French Senate inquiry whether he could guarantee French public-sector data would never be handed to US authorities without French consent. He said he could not (Sénat record).

So the useful question isn't only "where is the server?" but "who owns the company that runs it?"

3. It doesn't mean everything stays in the EU

Residency offers often cover your content but not everything around it. OpenAI's own documentation, for example, says some "system data" may leave the region even with EU data residency enabled (OpenAI, "Your data"). Read the exceptions, not just the headline.

So what does stop someone reading it?

Roughly three things, each with its own fine print: a contract you trust the provider to keep, hardware that keeps even the operator out (and lets you check that), or running the model on a machine you own. "EU-hosted" can be combined with any of them, but on its own it's none of them.

Next time a vendor says "your data stays in the EU", ask the follow-up: and who can read it while it's there?

I track what AI services actually promise on this, with sources, at Ostrafile; there's a free sample on my profile.

DEV Community: Ostrafile