If you have ever pasted a client contract, a snippet of proprietary code, or your own medical history into ChatGPT, here is an uncomfortable truth: by default, some of that text can be reviewed by human trainers, stored on OpenAI's servers, and in specific circumstances shared with third-party service providers. OpenAI's own privacy policy is explicit about this. The company works with vendors for hosting, analytics, and safety review, and consumer chats may be used to improve the models unless you turn that off.
Most people never touch these settings. In fact, a 2024 Cisco Data Privacy Benchmark study found that 48% of employees admitted entering non-public company information into generative AI tools, and roughly a quarter had no idea whether their employer even allowed it. That gap between what we type and what we understand about where it goes is exactly the problem.
This guide walks through how to actually stop ChatGPT sharing data with third parties and reduce what OpenAI itself retains. We will cover the exact toggles to flip, the difference between consumer and business plans, a real before-and-after scenario, a comparison of privacy-focused approaches, and the limits of what any setting can promise. No fear-mongering, just the practical steps I use myself.
Key Takeaways
- Turn off "Improve the model for everyone" in ChatGPT settings to stop your chats from training future models.
- Use Temporary Chat for anything sensitive so conversations are not saved to your history or used for training.
- API, Team, and Enterprise data is not used for training by default; consumer free/Plus data can be unless you opt out.
- Even with opt-out, OpenAI retains data for up to 30 days for abuse monitoring, and third-party sub-processors still handle infrastructure.
- The safest option for truly confidential material is a local AI model that never leaves your machine.
- Audit the browser extensions and integrations connected to ChatGPT; they are a common and overlooked leak point.
What "sharing data with third parties" actually means for ChatGPT
Before you change any settings, it helps to be precise about the three separate things people usually lump together as "data sharing." They are not the same, and they carry different risks.
1. Model training
By default on the free and Plus consumer plans, your conversations can be used to train and improve OpenAI's models. That means a human reviewer might read a sample, and fragments of your text could indirectly influence future model behavior. This is the piece most people want to stop first.
2. Third-party sub-processors
OpenAI does not run every piece of infrastructure itself. It uses cloud hosting (Microsoft Azure), analytics providers, payment processors, and content-moderation vendors. These are contractually bound "sub-processors," but your data does technically pass through systems OpenAI does not solely own.
3. Integrations and plugins you connect
When you connect a third-party GPT, a browser extension, or a "connector" to Google Drive or GitHub, you are authorizing data to flow to those companies too. This is the leak nobody thinks about. If you run a lot of extensions, it is worth reading our walkthrough on how to audit browser extension permissions for AI data leaks before you trust any of them near your prompts.
You can meaningfully control items 1 and 3. Item 2 you cannot fully eliminate on a hosted service, which is why the last section of this guide covers running AI locally.
How to stop ChatGPT from using your chats for training
This is the single highest-impact change, and it takes under a minute. Here is the exact walkthrough on the web app as of early 2026.
- Log in at
chatgpt.comand click your profile icon in the bottom-left corner. - Choose Settings.
- Open the Data controls tab.
- Find the toggle labeled "Improve the model for everyone" (older versions call it "Chat history & training").
- Switch it off. On some accounts this is a single toggle; on others you will confirm in a pop-up.
Once this is off, new conversations are not used to train OpenAI's models. Note an important nuance: on the classic version of this setting, turning off training also disabled saved chat history. OpenAI has since decoupled them for many users, so you can now keep your history while opting out of training. Check which version your account shows.
Submitting a formal opt-out and data deletion request
If you want to go further, OpenAI provides a privacy portal at privacy.openai.com where you can:
- Request that your personal data not be used to train models (a stronger, account-level request).
- Submit a data access request to see what is stored.
- File a deletion request under GDPR, CCPA, or similar laws.
Deletion requests are honored, but remember that anything already absorbed into a trained model cannot be surgically removed. That is why prevention beats cleanup.
Use Temporary Chat for anything sensitive
The training toggle is a global setting. But sometimes you want a one-off private session without changing your defaults. That is what Temporary Chat is for.
To start one, click the model name at the top of a new conversation and select Temporary Chat (or the small "ghost" icon on the new-chat screen). In a Temporary Chat:
- The conversation does not appear in your history.
- It is not used to train models.
- It does not use your saved memories or personalization.
- OpenAI still keeps a copy for up to 30 days for safety and abuse monitoring, then deletes it.
Think of Temporary Chat as the incognito mode of ChatGPT. It is my default for anything I would not want quoted back to me.
A worked before-and-after example
Say you are a freelance developer reviewing a client's authentication code. You have 3 files, roughly 480 lines total, including an API key in a config example and a comment describing an internal endpoint.
Before (default settings): You paste all 3 files into a normal chat with training left on. That API key, the endpoint, and the client's logic are now stored, potentially sampled for review, and eligible to influence future models. If the client's contract has a confidentiality clause, you may have just breached it.
After (hardened setup): You first redact the API key and replace the endpoint with https://api.example.com. You open a Temporary Chat with training already disabled account-wide. You paste only the 2 files that actually contain the bug. Result: no history entry, no training use, minimal 30-day retention, and no real secrets exposed even during that window.
The "after" version took about 90 extra seconds and eliminated nearly all the meaningful risk. That is the ratio to aim for.
Consumer vs Team vs API vs local: which keeps data most private?
Not all ChatGPT access is equal on privacy. The plan you use changes the default rules dramatically. Here is how the main options compare on the criteria that matter.
| Option | Trains on your data by default? | Data retention | Third-party exposure | Best for |
|---|---|---|---|---|
| Free / Plus (consumer) | Yes, unless you opt out | Until you delete | Sub-processors + optional integrations | Casual, non-sensitive use |
| Team | No | Configurable, retained short-term | Sub-processors only | Small teams with mild confidentiality needs |
| Enterprise | No | Admin-controlled | Sub-processors, with DPA | Regulated businesses |
| API | No | Up to 30 days (0 with ZDR) | Sub-processors only | Developers building products |
| Local model (Ollama etc.) | Never | Only what you store | None | Maximum-confidentiality work |
The pattern is clear: paid business tiers and the API do not train on your inputs by default, and local models remove third parties entirely. If you handle client or regulated data, the consumer plan is the weakest choice even after you flip every toggle.
Lock down connectors, custom GPTs, and browser extensions
You can perfectly harden your ChatGPT settings and still leak data through the things bolted onto it. This is where a lot of quiet exposure happens.
Review connected apps
- In Settings > Connected apps (or "Data controls > Connectors"), list every service linked to your account.
- Revoke anything you no longer use, especially cloud-storage and calendar connectors that read files automatically.
- For each remaining connector, confirm what scope it has. Read-only access to one folder is very different from full-drive access.
Be skeptical of custom GPTs
Third-party GPTs in the store can send your conversation to the builder's own servers through "Actions." A GPT that summarizes PDFs may route your document to an external API you have never heard of. Only use GPTs from sources you trust, and never feed them confidential material.
Extensions are a top leak vector
Browser extensions that "enhance" ChatGPT often request permission to read page content across all sites. That is a wide-open door. The same discipline applies to AI overviews and other injected features; if you want to reduce your AI footprint at the browser level, our guide on how to block AI overviews without losing search pairs well with an extension audit.
For teams managing web properties, the same "audit everything you install" principle scales to your CMS. If you run WordPress, the discipline in our piece on vetting WordPress plugins for security maps directly onto vetting AI integrations.
When settings are not enough: run AI locally
Here is the honest limit of everything above: as long as your prompts travel to a hosted service, you are trusting that company's contracts, security, and sub-processor list. For the most s
Cover image: Innovate Maryland Emerging Technology Center by MDGovpics, licensed under BY 2.0 via Openverse.
