benchr
Here's the part nobody flags when you sign up: for most consumer chatbots, the privacy choice was already made for you, and it was made in the provider's favor. You didn't agree to share your chats with the training pipeline so much as you didn't go find the setting that says don't. That setting exists on every major service. Almost nobody flips it.
This isn't a "be careful out there" piece. It's a list. Below is exactly who learns from your conversations the moment you start typing, where each opt-out lives, and how long your words stick around. The pattern is the same everywhere once you see it, so let's go provider by provider.
Who trains on you by default
The table is the whole article, so start here. Each row is one provider's consumer tier: whether it trains on your chats out of the box, the exact menu where you turn that off, and how long your data lives once you do. Read across, find the service you actually use, and change that one setting before you read the rest.
| Provider | Trains by default? | How to opt out | Retention |
|---|---|---|---|
| ChatGPT (Free, Plus, Pro) | Yes | Settings → Data Controls → turn off "Improve the model for everyone," or use Temporary Chat | Opt-out stops future training; not retroactive |
| Claude (Free, Pro, Max) | Yes, on new or resumed chats unless you opt out | Privacy Settings → decline the training option (you can change it anytime) | 30 days if you decline; 5 years if you allow it |
| Gemini | Yes, plus a subset goes to human reviewers | Turn off Gemini Apps Activity | 72 hours after opt-out; reviewed chats kept up to 3 years |
| Copilot (signed-in consumer) | Yes, unless you opt out | Account settings while signed in | Opt-out stops future training |
| API / Enterprise / Business (all major providers) | No | Nothing to do; not used for training by default | Set by your contract or DPA |
ChatGPT: on by default, opt-out tucked away
Consumer ChatGPT, meaning Free, Plus, and Pro, may use your conversations to train OpenAI's models unless you say no. The switch is in Settings under Data Controls, labeled "Improve the model for everyone." Turn it off and new chats stop going into training. Worth saying twice: the opt-out isn't retroactive, so anything already absorbed stays absorbed. You're protecting what you type next, not what you typed last week.
The cleaner escape hatch for one-off private questions is Temporary Chat. It skips your history, doesn't touch memory, and isn't used for training. Think of it as the incognito tab of chatbots. And if you run anything through OpenAI's API or a ChatGPT Team, Business, or Enterprise account, you're already covered: those aren't used for training by default. One caveat on the business side, giving a thumbs-up or thumbs-down can opt that single conversation into review, so feedback isn't free.
Claude: the default that flipped in 2025
This is the one that changed, and changed the wrong way for privacy. Anthropic used to delete consumer Claude chats within about 30 days and never trained on them. Under the updated Consumer Terms, that flipped. Free, Pro, and Max users, including Claude Code on those plans, now have their new or resumed chats used for model training unless they opt out. Say yes and your retention jumps from 30 days to five years. Say no and the old 30-day window holds.
TechCrunch flagged the framing as a dark pattern: a big "Accept" button with the training toggle pre-set to On sitting right beneath it. Existing users had until October 8, 2025 to make the call, and you can still change it anytime in Privacy Settings. None of this touches the commercial side, Claude for Work, Government, Education, and API use, including through Amazon Bedrock and Google Cloud's Vertex AI, all sit outside the change. If you've been letting an assistant handle purchases on your behalf through a consumer plan, the five-year retention is the line to reread.
Anthropic didn't ask you to share your chats. It asked you to notice the toggle that was already on.
Gemini: the human-review wrinkle
Gemini carries a risk the others mostly don't: people read some of it. Google's Gemini Apps Privacy Hub says a subset of chats are reviewed by human reviewers, including trained service providers, to improve Google's services and the models behind Gemini. Those reviewed chats get kept for up to three years, disconnected from your Google Account, and here's the kicker, they aren't deleted when you delete your activity. Deleting your history doesn't reach them.
Google says the quiet part out loud in its own help text: don't enter confidential information you wouldn't want a reviewer to see or Google to use to improve its services. Take that at face value. Turning off Gemini Apps Activity stops future chats from being used for training, but chats still sit in short-term storage for up to 72 hours so Gemini can answer you. So "off" means off for training, not gone in an instant.
Copilot and Meta AI: the rest of the field
Signed-in consumer Copilot trains by default too. Microsoft's privacy FAQ says it uses Copilot conversation activity, voice, and uploaded files for AI training unless you opt out, and signed-in users can change that in account settings. Plenty of people are excluded automatically: work and school (Entra ID) accounts, Microsoft 365 personal and family subscribers, signed-out users, anyone under 18, and users in a handful of countries including Brazil, China (excluding Hong Kong), Israel, Nigeria, South Korea, and Vietnam. If you're on a work Copilot, you're already in the safe tier.
Meta AI is a different shape. From December 16, 2025, Meta's updated Privacy Policy uses your interactions with Meta AI across Facebook, Instagram, Messenger, and WhatsApp to personalize content and ads. That covers what you say to Meta AI, not the private messages you send friends and family, unless you pull those into a Meta AI chat yourself. The line to watch is the boundary between talking to the assistant and talking through the app.
What to actually do
Pick your service and change the one setting. Go with ChatGPT's Data Controls toggle or Temporary Chat if that's your daily driver. In Claude, decline the training option in Privacy Settings so you keep the 30-day window instead of the five-year one. In Gemini, turn off Apps Activity and keep anything confidential out of it entirely, given the human-review storage. In consumer Copilot, flip the setting in your signed-in account.
And the broad rule that survives any policy update: if the data is sensitive, use the business or API tier, where not-training is the default. The whole reason these settings stay hidden is that defaults are sticky, the same dynamic that powers zero-click search, where most people never click past what's handed to them. Don't be the default. Flip the switch.
Frequently asked
Does ChatGPT train on my conversations by default?
On consumer plans (Free, Plus, Pro), yes. Your conversations may be used to improve OpenAI's models unless you opt out. Turn it off in Settings then Data Controls by switching off "Improve the model for everyone," or use Temporary Chat, which isn't used for training. The opt-out applies going forward and isn't retroactive. OpenAI's business products, the API and ChatGPT Team, Business, and Enterprise, are not used for training by default.
Did Claude change its privacy policy?
Yes. Anthropic reversed its consumer default. Before, consumer Claude chats were deleted within about 30 days and weren't used for training. Under the updated Consumer Terms, Free, Pro, and Max users, including Claude Code on those plans, have new or resumed chats used for model training unless they opt out, and allowing it extends retention from 30 days to 5 years. Existing users had until October 8, 2025 to choose, and you can change it anytime in Privacy Settings. Commercial plans, Work, Government, Education, and API, are not affected.
How do I opt out of training across providers?
Each provider hides the switch in a different menu. In ChatGPT, go to Settings then Data Controls and turn off "Improve the model for everyone." In Claude, open Privacy Settings and decline the training option. In Gemini, turn off Gemini Apps Activity. In consumer Copilot, change the training setting in your account settings while signed in. None of these are retroactive, so they stop future use, not data already absorbed.
Does the API train on my data?
No, not by default. OpenAI does not train on API, Team, Business, or Enterprise data by default; Anthropic excludes Work, Government, Education, and API use (including via Amazon Bedrock and Google Cloud's Vertex AI) from its consumer training change; and Microsoft excludes work Entra ID Copilot accounts. The business and enterprise tiers are the privacy-safe default. The free consumer tiers are where you most need to check the settings.
Is anything I type ever truly private?
Close, but read the fine print. ChatGPT's Temporary Chat isn't used for training and doesn't create memories. Turning off Gemini Apps Activity stops future training, though chats are still stored for up to 72 hours so Gemini can respond. The cleanest default privacy comes from the business and API tiers, which aren't used for training. Treat any consumer chat as something a person could conceivably see until you've changed the setting yourself.
Can a human actually read my Gemini chats?
Yes. Google's Gemini Apps Privacy Hub says a subset of chats are reviewed by human reviewers to help improve its services and models. Those reviewed chats are kept for up to three years, disconnected from your Google Account, and aren't deleted even if you delete your activity. That's why Google explicitly warns against entering confidential information you wouldn't want a reviewer to see. Turning off Gemini Apps Activity stops future chats from training, though chats are still stored for up to 72 hours so Gemini can respond.
Changelog
- May 30, 2026 — Originally published. Provider defaults, opt-out paths, and retention figures verified against Anthropic's consumer-terms post, Google's Gemini Apps Privacy Hub, Microsoft's Copilot privacy FAQ, and reporting on OpenAI's data controls.
References
- Anthropic, "Updates to Consumer Terms and Privacy Policy," anthropic.com, accessed May 2026.
- TechCrunch, "Anthropic users face a new choice — opt out or share your data for AI training," techcrunch.com, accessed May 2026.
- Google, "Gemini Apps Privacy Hub," support.google.com, accessed May 2026.
- Microsoft Support, "Privacy FAQ for Microsoft Copilot," support.microsoft.com, accessed May 2026.
- Yahoo Tech, "How to keep your ChatGPT data private by opting out," tech.yahoo.com, accessed May 2026.
- Secure Privacy, "GPT-5 training data opt-out," secureprivacy.ai, accessed May 2026.
- OpenAI Help Center, "How do I turn off model training," help.openai.com, accessed May 2026.
- Social Media Today, "Meta's not going to scan your private DMs for AI training," socialmediatoday.com, accessed May 2026.