Every major AI assistant trains on your conversations by default. The companies behind ChatGPT, Gemini, Copilot, and others have built their default settings to harvest your prompts, questions, and uploads—unless you explicitly tell them not to.
I reviewed the current privacy policies and data practices of eight major AI platforms to find out exactly what they collect, how long they keep it, and whether you can stop them.
The Quick Summary
| Platform | Trains by Default | Can Opt Out | Data Retention | Human Review |
|---|---|---|---|---|
| ChatGPT | Yes | Yes | 30 days (minimum) | Yes |
| Claude | Hybrid | Yes | 30 days or 5 years | Yes |
| Gemini | Yes | Yes | 3-36 months | Yes |
| Copilot | No (M365) | N/A | Per compliance | Limited |
| Perplexity | Yes | Yes | Varies | Yes |
| DeepSeek | Yes | No | Indefinite | Yes |
| Grok (X) | Yes | Yes | Unknown | Unknown |
| Meta AI | Yes | Limited | Indefinite | Yes |
ChatGPT (OpenAI)
What they collect: All conversations, uploaded files, images you generate, and custom GPT configurations.
Training default: Your data trains their models unless you opt out. Even if you opt out, conversations are retained for at least 30 days for “safety and abuse monitoring.”
Human access: OpenAI staff can review conversations for safety purposes.
How to Opt Out of ChatGPT Training
On desktop:
- Click your profile icon
- Select “Settings”
- Go to “Data Controls”
- Toggle off “Improve the Model for Everyone”
On mobile:
- Tap the two horizontal lines (top-left)
- Scroll down and tap your account name
- Select “Data Controls”
- Turn off the training toggle
Alternative: Use “Temporary Chat” mode (toggle at the top of new chats). These conversations aren’t used for training but are still kept for 30 days.
Claude (Anthropic)
What they collect: Conversations, uploaded documents, and any files you share.
Training default: Anthropic shifted to an opt-in/opt-out hybrid in September 2025. If you don’t respond to policy update prompts, the default becomes consent. Opting in extends retention from 30 days to 5 years—a 60x increase.
Human access: Anthropic staff can access conversations for safety and quality purposes.
How to Opt Out of Claude Training
- Open Claude (web or mobile)
- Go to Settings
- Find “Privacy” or “Model improvement” section
- Set the training toggle to Off
Note: Claude’s paid tiers (Pro, Team) don’t use conversations for training by default.
Gemini (Google)
What they collect: Conversations, uploaded images, voice queries, and—critically—integrates this with your broader Google data including search history, Gmail, and YouTube activity.
Training default: Conversations are saved for 18 months by default (adjustable to 3 or 36 months) and are used to improve models. Human reviewers may examine your conversations.
How to Opt Out of Gemini Training
- Go to Gemini
- Click “Activity” (now called “Keep Activity” as of August 2025)
- Choose “Turn Off”
- Select “Turn Off and Delete Activity”
- Click Next, then Delete
Alternative: Gemini now has “Temporary Chat” mode similar to ChatGPT—these sessions aren’t saved or used for training.
Microsoft Copilot
What they collect: Prompts and responses within M365 Copilot.
Training default: Microsoft 365 Copilot works within your existing Microsoft compliance framework. Prompts and responses are not used to train public AI models.
Copilot Training Settings
For consumer Copilot (free):
- Click your profile icon
- Select your account name
- Go to “Privacy”
- Turn off “Model Training on Text” and “Model Training on Voice”
Enterprise M365 Copilot doesn’t require opt-out—it’s excluded from training by design.
Perplexity
What they collect: Search queries, conversation history, and any follow-up questions.
Training default: Free and Pro users have AI Data Retention enabled by default. Your data can be used for model training unless you opt out.
How to Opt Out of Perplexity Training
- Log into your account
- Go to Settings
- Find the AI training opt-out toggle
- Disable it
Important: Previous data may still be retained even after opting out. Enterprise and API users have stronger protections—Perplexity maintains a “Zero Data Retention Policy” for the Sonar API.
DeepSeek
What they collect: Everything. Account details, all inputs (text, audio, prompts, files), device data, IP addresses, and—uniquely concerning—keystroke patterns and typing rhythms.
Training default: Yes, and there’s no way to opt out.
Storage location: All user data is stored on servers in the People’s Republic of China.
Human access: Conversations can be reviewed by staff.
DeepSeek Privacy Status
You cannot opt out of DeepSeek using your data for training. Any prompt you submit may be used to refine future models. The collection of behavioral biometrics (keystroke patterns) is particularly invasive—typing rhythm is unique to each person, like a fingerprint.
If privacy matters to you, don’t use DeepSeek’s cloud service. If you need their models, run them locally through Ollama or another local inference setup.
Grok (X/Twitter)
What they collect: Your X posts, direct messages to Grok, and interaction patterns.
Training default: Yes. X’s terms allow Grok to train on your posts unless you opt out. The January 2026 terms update treats all prompts and outputs as “Content” under the platform’s licensing framework.
How to Opt Out of Grok Training
- Go to X Settings
- Select “Privacy and Safety”
- Click “Grok & Third-party Collaborators”
- Uncheck the training box
Note: Opting out only affects future data. Past posts may have already been ingested.
Meta AI (Facebook/Instagram/WhatsApp)
What they collect: Public posts, comments, likes, AI chat interactions, and profile information across Facebook, Instagram, Messenger, and WhatsApp.
Training default: Yes. As of December 2025, Meta’s privacy policy explicitly allows using AI chat interactions to personalize recommendations and deliver targeted ads, plus train their Llama models.
How to Object to Meta AI Training
Meta doesn’t offer a simple toggle. You must submit an “objection request”:
On Facebook:
- Go to your profile
- Click Settings and Privacy
- Scroll to “More info and support”
- Click “About,” then “Privacy policy”
- Find “Learn more about your right to object”
- Fill out the form explaining how the processing impacts you
On Instagram:
- Tap your profile icon
- Open the menu (top right)
- Go to “More info and support” → Privacy Center
- Select “AI at Meta”
- Tap “Submit an objection request”
- Complete the form
Critical limitations:
- Meta must accept your objection before it’s honored
- Your objection doesn’t cover photos others post of you
- US users have no real opt-out option
- European users had a deadline that passed in May 2025
What This Means
The trend is clear: default settings favor data collection. Every platform has made it harder, not easier, to protect your privacy over the past year. Claude’s shift to an opt-in hybrid model, Meta’s expanding scope, and DeepSeek’s complete lack of controls show where things are heading.
Even when you opt out, minimum retention periods mean your data sits on company servers for weeks to months. Human reviewers at every major company can read your conversations. And opting out today doesn’t erase yesterday’s data.
What You Can Do
Immediate actions:
- Go through each platform you use and check/change settings now
- Use “Temporary Chat” modes when available for sensitive queries
- Review your settings quarterly—companies rename and reset these controls without notice
Never paste into any cloud AI:
- Passwords, API keys, or 2FA codes
- Credit card or bank account numbers
- Unpublished legal, medical, or financial records
- Trade secrets or unreleased product information
For maximum privacy:
- Run models locally using Ollama, LM Studio, or similar tools
- Use Claude Pro or Team tiers (training opt-out by default)
- Consider Microsoft 365 Copilot for work (enterprise compliance)
- Avoid DeepSeek’s cloud service entirely
The most private choice among cloud services remains Claude with training disabled, followed by ChatGPT with training disabled and Temporary Chat enabled. But “most private” is relative—if your data truly can’t be shared, don’t type it into any cloud AI service.