Can I opt out of my input or output data being used for training?
Mistral AI’s change to train on user prompts by default for non-enterprise tiers has raised concerns among paying customers who expected stronger privacy guarantees, especially from a European provider. Commenters debate confusing “opt-in vs opt-out” wording, the removal of an organization-wide privacy toggle on the Team plan, and how much users can realistically trust any AI vendor’s promises about not using data for training. The thread situates Mistral’s approach alongside competitors like Anthropic, Google, and Grok, and highlights a broader shift toward local or proxy-based LLM use for those who want tighter control over sensitive data.
Policy change and training defaults
- Discussion centers on Mistral now using user input/output for training by default on non‑enterprise tiers, with an opt‑out mechanism.
- For Vibe (consumer/Team): users are “in” by default and must disable training per account.
- For Enterprise: training is off by default and can be managed centrally by admins.
- Some participants see this as a reversal from earlier documentation that claimed Team was opted out by default.
Org controls and confusion over toggles
- Several users report previously having an organization‑wide admin toggle to disable training; newer Team accounts reportedly lack this.
- One commenter says support confirmed the org‑wide toggle was moved to Enterprise only.
- Others still see global toggles, suggesting behavior differs between old and new accounts, which is a source of confusion.
- Docs and UI reportedly lagged behind the policy change, leading to unintended training on some test prompts.
Opt‑in vs opt‑out terminology
- Long sub‑thread debates the meaning of “opt in by default.”
- Consensus among many:
- “Opt‑in” = off by default; user must choose to turn it on.
- “Opt‑out” = on by default; user must choose to turn it off.
- Some argue corporate marketing has muddied these terms in user‑hostile ways.
Comparisons to other AI providers
- Claude Team/Enterprise: cited as disabling training on prompts by default; individuals report opposite defaults on personal plans.
- Grok: praised for training off by default, but its ToS are criticized as extremely broad.
- Google/Microsoft/Anthropic/OpenAI: many commenters say similar “on by default, opt‑out available” patterns are common.
Trust, privacy, and legal angles
- Strong skepticism that any large provider truly refrains from training on opted‑out data; others argue contractual and GDPR risk makes secret training unlikely.
- GDPR is mentioned as making unauthorized secondary use of data explicitly illegal, though some doubt enforcement strength.
- Concerns raised about PII in prompts, retention durations, breaches, and irreversibility once data is in a model.
Why use or avoid Mistral
- Pro‑Mistral arguments: EU jurisdiction, data‑sovereignty, local models, good OCR/small models, ability to serve third‑party open models like GLM.
- Anti‑Mistral sentiment: perceived “enshittification,” privacy‑hostile defaults, Saudi partnerships, patents, and a sense that they mirror US Big Tech behavior.
- Several conclude that local/self‑hosted models are ultimately the only reliable way to avoid training on one’s data.