ChatGPT Data Controls: Who Trains on Your Chats
ChatGPT’s training toggle is on by default for personal plans. Claude, Gemini, and local Llama are the other common answers. Check the matrix, not a rumor.
Earlier snapshot: AI privacy: which tools train on your data?.
If you paste payroll, source code, or a patient’s chart into a free chat box, assume it might be used to train a model unless you have proof otherwise. Our opt out of data training feature is the row to trust; this post is the short version for ChatGPT, Claude, and Gemini.
ChatGPT (personal)
On Free, Plus, and Pro personal workspaces, OpenAI’s help center says data sharing for training is on by default. You can keep chat history and still opt out:
Settings → Data Controls → Improve the model for everyone → off.
New chats then stay out of training. OpenAI’s FAQ is the source, not a screenshot from 2024. Turning that toggle off no longer wipes history.
Business, Enterprise, Edu, and the API do not use your inputs and outputs for training by default.
Matrix: ChatGPT has a training opt-out.
Claude and Gemini
Claude and Gemini both have yes on data-training-optout in our matrix. Anthropic’s commercial terms are the usual reason people prefer Claude for client work; still open the in-product privacy settings rather than trusting a blog. Google will store Gemini chats and has historically used human review — do not put secrets in the free consumer box.
Claude Code documents training opt-out on individual plans; Team and Enterprise default to no training.
Local models
The only “the vendor never sees it” answer is weights on your GPU: Llama via Ollama, LM Studio, or a self-hosted API. Llama’s enterprise-privacy row is self-hosted for that reason.
Grok and Perplexity are no on training opt-out in our current matrix — do not assume an X or search chat is private.
The live table: /features/data-training-optout.