Snapshot caveat: Provider data policies change often, so verify the current default and menu path in each app's own settings before you rely on it. Two recent shifts show how fast this moves. Anthropic flipped Claude's consumer default to training-on in August 2025, and OpenAI's retention practice changed twice during 2025 because of a court order. Reflects June 2026.
AI Privacy and Your Data
AI & You
6 min read
In Short
When you type into a cloud chat app like ChatGPT, Claude, or Gemini, your words travel to the company and are processed on its servers. On most consumer defaults your conversations can be used to train the model unless you turn that off, and each app hides the switch in a different place. Treat anything you paste as something you cannot fully un-share, so keep passwords and other people's private data out of the box. For prompts that must never leave your device, run a local model.
01. What It Is
A cloud AI app runs the model on the provider's servers, not on your device. Your prompt, any files you attach, and the reply all travel to the company, the same for ChatGPT, Claude, Gemini, and Copilot.
See chat-apps-and-subscriptions.
Three things can happen to that data. The company can use your text to answer you now, store the conversation, or feed it into future training. A fourth path, human review, means a person may read a sample. Privacy settings mostly control training, not storage or review.
02. Why It Matters
You cannot un-share a prompt. Once text reaches a provider, the company may keep a copy, a reviewer may read it, or it may be pulled into training, and each creates a lasting record outside your control. A password or a client's private detail pasted by mistake does not come back when you delete the chat, and each app controls this differently.
03. How It Works
What happens to what you type
A consumer chat app sends your message to the provider, the model replies on its hardware, and the answer comes back, so nothing about that trip stays on your device.
The only way to keep a prompt fully on your machine is to run a local model, covered below and in running-llms-locally.
Training is usually on by default, and where the opt-out lives
Paying for a consumer plan does not by itself stop training use. It is a separate, per-app setting, and on most free and default paid consumer tiers it starts switched on.
| App | Trains by default? | Where to turn it off | Default retention |
|---|---|---|---|
| ChatGPT (OpenAI) | Yes | Settings > Data Controls > "Improve the model for everyone" | Deletions cleared within 30 days |
| Claude (Anthropic) | Yes, since Aug 2025 | Privacy Settings (claude.ai/settings/data-privacy-controls) | 30 days off, 5 years if training on |
| Gemini (Google) | Yes, via "Keep Activity" | Gemini Apps Activity > "Keep Activity" | 18 months (72 hours if off) |
| Copilot (Microsoft) | Yes | Profile > Privacy > "Training on conversation activity" | 18 months |
OpenAI states ChatGPT trains on your conversations unless you opt out. Anthropic's August 2025 update reversed Claude's earlier default, so Claude now trains on Free, Pro, and Max chats and Claude Code sessions when the setting is on, and existing users had until 8 October 2025 to choose. Gemini's "Keep Activity" lets Google use your chats to improve its services, including training its AI models, and Google says plainly not to enter confidential information a reviewer should not see, nor login or payment details.
For other apps such as Grok, Meta AI, Perplexity, and DeepSeek, not verified here, assume the consumer default trains on your chats unless you see an off switch.
Retention and human review
Turning off training does not delete your history or stop every other use, and deleting a chat usually schedules removal within about 30 days rather than at once. The table above lists each app's default retention, and allowing Claude to train extends its retention to 5 years.
Human review is the part people miss. A sample of Gemini chats is read by reviewers, kept up to 3 years, and not removed when you delete your activity. OpenAI may review Temporary Chats for abuse though they never train the model, Anthropic may use safety-flagged chats to improve its safety classifiers with training off, and Microsoft's review for suspected Code-of-Conduct violations cannot be opted out of. Opting out still allows product improvement, advertising, and safety uses.
A court order can override deletion. As of October 2025, OpenAI says the New York Times copyright lawsuit no longer requires it to retain new consumer and API data indefinitely, with EEA, Switzerland, and UK conversations excluded and Enterprise, Edu, and the Zero-Data-Retention API never affected. As of January 2026, a federal judge affirmed an order to produce about 20 million de-identified ChatGPT logs, and the litigation is still live. In the EU and UK, GDPR gives enforceable rights to access and erase the personal data a provider holds, as Italy's 15-million-euro Garante fine on OpenAI in December 2024 showed.
See ai-regulation-governance.
Enterprise, business, and zero-retention tiers
The work-safe path is a business tier, where no-training is the default rather than a setting you must find. OpenAI does not train on ChatGPT Team, Enterprise, Edu, or API data unless the organisation opts in. Anthropic likewise excludes its commercial products, including Claude for Work, the API, Bedrock, Vertex, Gov, and Education. Microsoft does not train on work or school accounts or Microsoft 365 app content. OpenAI's Zero Data Retention on eligible API endpoints goes furthest, with inputs and outputs never logged, stored, or seen by reviewers, by request only.
Local models for full privacy
The only way to guarantee a prompt never leaves your device is to run the model locally, with no company in the loop, so nothing you type is sent out, stored, or used for training. The tradeoff is capability, since local models are smaller than the cloud frontier and need a capable machine.
See running-llms-locally and cloud-vs-local-which-to-choose.
04. What Not to Paste
Treat the chat box as a public counter, not a private vault. Keep these out of any cloud AI app unless you trust the tier for it.
- Passwords, API keys, and other secrets. Anything that grants access should never go into a prompt.
- Other people's personal data. Names tied to medical or financial details belong to them, not to a training set.
- Confidential or regulated work data. Unreleased figures, contracts, or code covered by an agreement. Use a business tier or a local model.
- Login or payment information. Google warns against this directly.
- Anything you could not un-share. If a kept copy, a reviewer reading it, or a training run would cause harm, do not paste it.
05. Key Terms
| Term | Plain meaning |
|---|---|
| Training opt-out | The per-app setting controlling whether your chats improve the model. Usually on by default on consumer tiers. Off excludes chats from training but does not delete them. |
| Data retention | How long the company keeps your conversation after you delete it, from 72 hours (Gemini, activity off) to 30 days (most deletions) to 5 years (Claude, training on). |
| Human review | A sample of chats read by trained people to rate quality or check for abuse. Often cannot be fully turned off, and reviewed copies can be kept up to 3 years at Gemini. |
| Zero Data Retention | A business API mode where inputs and outputs are never logged, stored, or seen by reviewers. The strongest cloud privacy option, by request. |
| Local model | A model that runs on your own device, so prompts never leave it. The only path to full input privacy. |
06. Common Misconceptions
"Paying for a plan makes my chats private."
It does not. Training use is a separate setting, independent of whether you pay, so a 20-dollar plan can still be training-on by default. Turn it off, or use a business tier that excludes training.
"A private or temporary chat is never kept or read."
Temporary and incognito modes keep chats out of training, but copies can still be retained for a window and reviewed for abuse. Gemini's reviewed chats can be held up to 3 years even after you delete your activity, and deletion itself is usually scheduled within about 30 days rather than instant.
"Turning off training turns off everything else."
It does not. The company can still store your history, run safety or abuse review, and operate the product. Microsoft states its training opt-out does not cover product improvement, advertising, or safety uses.