LLM Training Policies
ZeitMind routes customer data to AI model providers via API endpoints only. This page documents the data protection posture of each supported foundational model provider with respect to model training, and assesses compliance against ZeitMind's contractual no-training commitment.
All customer data flows exclusively through provider API endpoints under commercial or enterprise terms. No consumer-facing chat products are used in any customer data path. No in-product feedback mechanisms (e.g. thumbs up/down) are enabled on customer-facing deployments.
Summary
| Provider | Access Path | No-Training Default | Contractual Basis | No-Training through use in ZeitMind |
|---|---|---|---|---|
| Anthropic Claude | Direct API + Vertex AI | ✅ Flat policy — no opt-in/out required | Commercial Terms + Privacy Center | ✅ Yes |
| OpenAI GPT | Direct API | ✅ Off by default; sharing disabled in org settings | DPA — processor role | ✅ Yes |
| Google Cloud Vertex AI | Hosting platform for EU-served Claude | ✅ Off by default | Service Specific Terms §18 — contractual clause | ✅ Yes |
| Mistral | La Plateforme API (Scale Plan) | ✅ Scale Plan default; toggles disabled | DPA opt-out structure — see caveat below | ✅ Yes (DPA caveat noted) |
All four access paths, as currently configured, satisfy ZeitMind's no-training commitment for API-delivered customer data. Mistral carries a structural DPA caveat that should be considered when onboarding customers with heightened contractual requirements.
Provider Assessments
Anthropic Claude — Direct API + Google Cloud Vertex AI
Anthropic explicitly excludes all API usage from model training. By default, inputs and outputs from commercial products — including the Anthropic API and third-party platforms such as Google Cloud's Vertex AI — are not used to train models. API logs are retained for 7 days for abuse monitoring only, then automatically deleted. This is a flat policy for all API and commercial customers, with no opt-in or opt-out mechanism required.
Anthropic's 2025 consumer policy update (which introduced opt-in training for Claude Free/Pro/Max users) explicitly does not apply to API usage or Vertex AI.
Key sources
- Privacy Center — commercial products (is my data used for model training?)
- Data input controls and audit statement
- Consumer terms update — confirms API/Vertex exclusion
OpenAI GPT — Direct API
As of March 1, 2023, data sent to the OpenAI API is not used to train or improve models unless a customer explicitly opts in. OpenAI's Data Processing Addendum formalises this: OpenAI acts as a data processor and will only process Customer Data for the purposes of delivering the Services.
Model feedback sharing and input/output sharing are both disabled in ZeitMind's OpenAI organisation settings. Abuse monitoring logs are retained for up to 30 days for safety purposes only and are not used for model training. Zero Data Retention (ZDR) is available for eligible enterprise customers who require full log elimination.
Key sources
Google Cloud Vertex AI — EU-Hosted Claude
Google's Service Specific Terms for Google Cloud Platform include a dedicated Training Restriction clause (Section 18) which states:
"Google will not use Customer Data to train or fine-tune any AI/ML models without Customer's prior permission or instruction."
This is a contractual term in the standard GCP agreement — not merely a policy statement — making it directly enforceable. It applies to all models served via Vertex AI, including the Anthropic Claude models ZeitMind serves from the EU multi-region endpoint.
Key sources
- Google Cloud Service Specific Terms — Section 18 (Training Restriction)
- Vertex AI and Zero Data Retention
- Google Cloud GenAI & Privacy whitepaper
Mistral — La Plateforme API (Scale Plan)
Mistral's La Plateforme API does not use customer data for model training by default. ZeitMind is on the Scale Plan, which explicitly excludes API data from training and improvement. This is confirmed in the Mistral admin console under API Keys → Privacy: "You are on the Scale Plan, therefore no data is used to improve our service." The training toggle is disabled.
DPA structural caveat
Unlike the other three providers — where the no-training commitment is either a flat policy (Anthropic) or an explicit contractual clause (Google, OpenAI DPA) — Mistral's DPA is structured as an opt-out rather than an absolute prohibition. Specifically, Mistral's DPA (Section 2.3) states that Mistral AI is authorised to process personal data as Controller for model training purposes "unless (a) Customer opted-out of training or (b) uses a Mistral AI Product that is opted-out by default and has not opted-in."
This means:
- The protection is conditional on the opt-out remaining in place, rather than being an unconditional contractual prohibition.
- Additionally, if users submit thumbs-up/thumbs-down feedback within a Mistral-powered interface, Mistral reserves the right to use that feedback and associated inputs/outputs for training — even for customers who are otherwise opted out. This is not applicable to ZeitMind's current deployment (no feedback UI is exposed), but is noted for completeness.
In ZeitMind's current configuration — Scale Plan, training disabled, no feedback UI — the operational risk is mitigated. However, the contractual structure is weaker than the other providers.
Key sources
Last reviewed: August 2026. Provider policies are subject to change.