Does OpenAI train on your ChatGPT data?

Training default

Your content is used for training unless you opt out.

Opt-out

Settings → Data Controls → turn off 'Improve the model for everyone' (account-level, syncs web+mobile); or privacy.openai.com 'do not train on my content'; legacy support/form opt-outs still honored.

Retention

Chats kept until deleted; on deletion removed from account immediately and permanently deleted from OpenAI systems within 30 days (unless already de-identified for training, or legal/security retention applies). — Temporary Chats auto-delete within 30 days and aren't trained on. Saved Memories stored separately and persist until deleted (deleted-memory logs kept up to 30 days). Even with training off, content sits ~30 days for abuse monitoring. NYT indefinite-hold ended 26 Sep 2025; residual April–Sept 2025 preserved logs and flagged-account logs may remain. US Free/Go tiers show ads (adults, logged-in); advertisers get only aggregate metrics, not chats.

Human review

OpenAI staff and specialized contractors may review content to monitor/investigate abuse and safety; if training is on, sampled conversations may be reviewed by human trainers. Temporary Chats may be reviewed only for abuse.

Feedback carve-out

If you submit feedback (e.g., thumbs up/down), 'the entire conversation associated with that feedback may be used to train our models,' even if otherwise opted out.

Go ($8/mo) shares these consumer defaults. Codex has separate full-environment training controls in Codex Settings.

Training default

Your content is used for training unless you opt out.

Opt-out

Same as Free: Settings → Data Controls → turn off 'Improve the model for everyone' (or privacy portal).

Retention

Same as Free: chats kept until deleted; deleted chats purged from systems within 30 days unless legal/security retention or already de-identified. — Identical consumer data policy to Free — Plus buys capability, not privacy. Temporary Chats auto-delete within 30 days. 30-day abuse-monitoring floor applies even with training off. Was in scope of the (now-ended) NYT preservation order.

Human review

Same as Free: abuse/safety review by staff/contractors; sampled human-trainer review when training is on.

Feedback carve-out

Same universal carve-out: submitting feedback licenses training on the entire associated conversation even after opt-out.

US Privacy Policy governs; 'Reference chat history' memory available. $20/mo, monthly only.

Training default

Your content is used for training unless you opt out.

Opt-out

Same as Free/Plus: Settings → Data Controls → turn off 'Improve the model for everyone.'

Retention

Same as Free/Plus: deleted chats purged within 30 days unless legal/security retention or already de-identified. — Consumer defaults identical to Plus. Pro ($200/mo; a $100/mo Pro tier added April 9, 2026) buys usage/model access, not different data handling. Was in scope of the (now-ended) NYT preservation order.

Human review

Same as other consumer plans: abuse/safety review; sampled review when training on.

Feedback carve-out

Same universal feedback carve-out.

OpenAI's NYT page lists 'ChatGPT Free, Plus, Pro' together for retention practices, confirming Pro follows consumer defaults.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

No opt-out needed — not trained on by default. It is opt-IN: training only occurs if the org/user explicitly opts in (e.g., feedback mechanisms). Admins control workspace data/privacy settings.

Retention

Workspace admins control retention; deleted or unsaved conversations removed within 30 days, unless longer retention is required by law or reasonably necessary to protect the services or a third party. — Renamed from 'ChatGPT Team' on Aug 29, 2025 (name change only; contracts/features/pricing/limits unchanged). $25/user/mo monthly or $20 annual (monthly cut $5 effective Apr 2, 2026), 2-seat minimum. Memories tied to individual accounts, not trained on by default, admin-controllable. Was listed in the (now-ended) NYT order scope for 'Team.'

Human review

Access limited to (1) authorized employees for engineering support, abuse investigation, and legal compliance, and (2) specialized third-party contractors bound by confidentiality, 'solely to review for abuse and misuse.' Workspace admins can view/export/delete end-user conversations.

Feedback carve-out

Opting in — e.g., submitting feedback — allows OpenAI to use that shared data to train models; otherwise no training.

ZDR is API-only; the ChatGPT chat product isn't ZDR because persistent chat history is intrinsic.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

Not trained on by default; opt-IN only via explicit agreement (e.g., feedback mechanisms). Admin-controlled workspace settings.

Retention

Workspace admins control retention duration; deleted conversations removed from systems within 30 days unless legally required to retain. — Contractual guarantee in Services Agreement §4.2. SOC 2 Type 2 audited; SAML SSO, Enterprise Key Management, Compliance API. Never impacted by NYT preservation order.

Human review

Authorized OpenAI employees 'will only ever access your conversations for the purposes of resolving incidents, recovering end user conversations with your explicit permission, or where required by applicable law.' Business data may run through automated classifiers (metadata only).

Feedback carve-out

Only if the customer explicitly opts in (e.g., feedback mechanisms) does OpenAI use shared data to train; otherwise never.

The Enterprise chat product is not a ZDR product (persistent chat history is part of how it works); ZDR is API-only.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

Not trained on by default; opt-IN only. Admins have the same controls as ChatGPT Enterprise.

Retention

Workspace admins control retention; deleted conversations removed within 30 days unless legally required to retain. — Processed under OpenAI's Student Data Privacy Agreement (SDPA) rather than the standard DPA. Never impacted by the NYT preservation order (explicitly exempt). Data export off by default; admin-enabled.

Human review

Same Enterprise-grade conditions: employee access only for incidents, explicit-permission recovery, or where legally required; automated classifiers produce metadata only.

Feedback carve-out

Only via explicit opt-in (e.g., feedback mechanisms); otherwise never.

Chat product not ZDR (API-only). ChatGPT for Teachers/Healthcare follow the same no-training-by-default model but are out of scope.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

Not trained on by default; opt-IN only (e.g., sharing feedback/evals in the Playground). Unless explicitly opted in, organizations are opted out of data-sharing.

Retention

Abuse-monitoring logs retained up to 30 days for all API usage, then removed, unless longer retention is required by law or reasonably necessary to protect the services/third parties. — Stateful endpoints persist application state 'until deleted' (Conversations, Threads, Assistants, Vector Stores, Files, fine-tuning files). /v1/moderations and audio transcription/translation have no abuse-monitoring retention. Fine-tuned models are private to the customer. Was in scope of (now-ended) NYT order for non-ZDR API.

Human review

Authorized employees (engineering support, abuse investigation, legal compliance) plus specialized third-party contractors 'solely to review for abuse and misuse.'

Feedback carve-out

API customers can opt in to share data (e.g., Playground feedback/evals) which OpenAI then uses to improve models; no training without explicit opt-in.

Governed by the Services Agreement + DPA. Data residency available (per-project) for eligible regions; non-US regions require MAM/ZDR amendment.

Training default

Your content is not used for training unless you explicitly opt in.

Opt-out

ZDR is itself the strongest opt-out of retention; content is never trained on (same default as standard API) and is excluded from abuse-monitoring logs. Configured in Settings → Organization → Data controls → Data Retention after approval.

Retention

Inputs and outputs are never logged and not retained for application state on ZDR-eligible endpoints; store parameter forced to false for /v1/responses and /v1/chat/completions. — Eligibility is approval-gated: customers must apply via sales and provide qualifying use-case justification; not available on free/standard pay-as-you-go by default. ZDR-eligible (stateless): /v1/chat/completions, /v1/responses, /v1/embeddings, /v1/moderations, /v1/completions, /v1/audio/*, /v1/realtime, gpt-image generation. NOT eligible (stateful): /v1/conversations, /v1/threads, /v1/assistants, /v1/vector_stores, /v1/files, /v1/fine_tuning/jobs, /v1/evals, /v1/batches, /v1/videos. Image/file inputs flagged as potential CSAM are retained for manual review even under ZDR. Never affected by the NYT order.

Human review

None under ZDR ('not accessed by human reviewers'), except potential-CSAM image/file inputs retained for manual review. Related controls: Eyes Off (no human review even if retained) and Safety Retention (retain + review only classifier-flagged severe-risk content).

Feedback carve-out

No training regardless; ZDR content is excluded from logs and human review. Opting into feedback would be a separate, explicit choice.

Eligible only for API/business customers, granted per organization/project; the ChatGPT consumer/chat products are not ZDR. Extended prompt caching may store encrypted KV tensors in GPU-local storage (not retained after 24h).