Skip to content
Call wiki

Does OpenAI train on what Codex sees on your computer?

Last updated 2026-08-17 · Confidence: documented — OpenAI’s data-controls docs and the Codex CLI’s own source; the separate Codex-settings environment toggle is reported from help-center excerpts.

By default on a personal ChatGPT plan, yes — one data-controls toggle turns it off. Business plans and API-key sign-in default to no training. Opting out stops training, not the upload: whatever Codex reads still goes to OpenAI’s servers for processing.

Personal plans: training on unless you opt out

Section titled “Personal plans: training on unless you opt out”

Codex signed in with a ChatGPT account inherits that account’s data controls — the CLI’s sign-in screen says so: “Uses your plan’s rate limits and training data preferences”, linking straight to ChatGPT settings. On Free, Plus, and Pro, content may train OpenAI’s models until you turn off Settings → Data Controls → “Improve the model for everyone”, or file “do not train on my content” at the privacy portal.

One catch: OpenAI’s Codex help page describes a separate Codex-settings control for training on full environments, which the ChatGPT toggle does not change.

Business, Enterprise, API: off by default

Section titled “Business, Enterprise, API: off by default”

ChatGPT Enterprise has no training on business data by default; Team and Business terms say the workspace owns and controls its business data — so the trade is training privacy for company ownership, as with a Claude Team account. With an API key, “data sent to the OpenAI API is not used to train or improve OpenAI models” unless you opt in.

What opting out doesn’t cover

Section titled “What opting out doesn’t cover”

Everything Codex pulls into context — code, notes, whatever its file access reaches — is sent to OpenAI for inference regardless; opting out governs training, not transmission or retention. The pattern shows in Computer History: interaction events are processed on OpenAI’s servers, deleted after processing, and feed training only “if allowed by your ChatGPT data controls”. Local telemetry is quieter: the CLI’s OTel export is off by default.

Sources

Section titled “Sources”