Research Snapshot · March 2026

AI Model Training & Data Retention

Does the model train on your conversations? Is data retained? Can you turn it off? Covers consumer GUI, paid plans, and direct API access. Policies shifted significantly in late 2025 — most providers moved to opt-out defaults for consumer users.

Yes / Safe / Off by default
Opt-out available (on by default)
No opt-out / Trains by default, can't disable
Requires paid / enterprise plan to disable
Not applicable / open weights
Model / Lab GUI — Trains on chats? GUI — On by default? GUI — User can disable? API — Trains on data? API — On by default? API — User can disable? Notes
ChatGPT OpenAI Yes (opt-out) On by default Yes — free & Plus No Off by default N/A — not trained Consumer GUI defaults to training. Disable via Settings → Data Controls → "Improve the model for everyone." Disabling also disables chat history. Team/Enterprise plans: never trained. API never trained by default; 30-day abuse-monitoring retention only.
Claude Anthropic Yes (opt-out) On by default Yes — Free, Pro, Max No Off by default N/A — not trained Policy shifted Aug 2025. Free/Pro/Max GUI: trains by default; opt out via Settings → Privacy → "Help improve Claude." Opt-in = 5-year retention; opt-out = 30-day retention. Deleted chats excluded always. API, Claude for Work/Enterprise, Bedrock, Vertex AI: never trained.
Gemini Google Yes (opt-out) On by default Yes (free & Adv.) No Off by default N/A — not trained As of Sep 2, 2025, GUI defaults to using a sample of uploads/chats for training ("Keep Activity"). Disable via Gemini Apps Activity / Keep Activity toggle. Paid Gemini Advanced still defaults on — not enterprise-safe. Workspace/enterprise: not trained. Vertex AI API: not trained. Chats kept 72 hrs even with activity off.
Copilot Microsoft Yes (opt-out) On by default No (Azure) Off by default N/A — not trained Consumer Copilot is among the least transparent on opt-out options. Enterprise/M365: governed by Microsoft Purview; data never used for training; full admin control via Customer Lockbox. Azure OpenAI API: not trained; 30-day abuse monitoring only.
Grok xAI Yes (limited opt-out) On by default Cannot fully disable No Off by default N/A — not trained Deeply tied to X (Twitter). xAI trains on public X posts and Grok conversations. Users can delete threads but cannot fully disable training use of chat data. Unauthenticated users have almost no privacy controls. API data not trained by default.
Meta AI Meta Yes (no opt-out) On by default No opt-out available Open weights N/A N/A Ranked worst for privacy in multiple 2025 studies. No dedicated opt-out mechanism for training on Meta AI chats. Governed by Meta's broad platform Data Policy — also trains on public Facebook/Instagram posts. Llama model weights are open-source; self-hosting means you control your own data entirely.
Le Chat Mistral AI Yes (opt-out) Opt-in framing Yes — all tiers No Off by default Zero retention option Ranked most privacy-friendly consumer AI in 2025 Incogni study. GUI chat stored until account deletion or manual deletion; API inputs retained 30 rolling days for abuse monitoring unless zero data retention (ZDR) is enabled. GDPR-aligned; EU-based.
DeepSeek DeepSeek (China) Yes (no opt-out) On by default No opt-out Opaque Opaque No controls published Significant privacy concerns. Collects account info, prompts, chat history. No published opt-out mechanism. Data subject to PRC law (including government access). Ranked near bottom in all 2025 privacy analyses. Self-hosting open-weight models avoids cloud data risks.
Llama (self-hosted) Meta / open weights N/A N/A Full control N/A N/A Full control Open weights — no data leaves your infrastructure when self-hosted. Training, retention, and logging are entirely your responsibility. No usage data sent to Meta. Gold standard for sensitive/regulated workloads.
Sources: Anthropic, OpenAI, Google, Mistral, xAI privacy policies and help documentation; Incogni LLM Privacy Ranking 2025; Drainpipe.io AI Privacy 2026; Bonfireci Shadow AI report Sep 2025; DigitalInformationWorld, WinBuzzer Aug 2025. Policies are subject to change — always verify current terms before sharing sensitive data. Last verified: March 2026.

⚠ Note: OpenAI, Google, and Anthropic all made significant privacy policy reversals in Aug–Sep 2025, moving consumer products to opt-out training defaults. Even paid consumer tiers (ChatGPT Plus, Claude Pro, Gemini Advanced) are not enterprise-safe without explicit opt-out.