Six companies run the chatbots millions of people type into every day, each with its own rule for your words. Rules differ. Almost everyone skips the fine print.
Checked this week, five of those six give you a real setting. Flip it, and new chats skip training. Meta AI skips that switch for most readers. It works only under European Union, United Kingdom or Brazilian privacy law135679.
What "Training" Actually Means Here
Training means a company feeds your past chats into its next model. Pick "yes," and a line you typed last week can shape tomorrow's answer to a total stranger.
Every service below separates that choice from basic processing. Your message travels to a server either way, and a reply comes back either way, switch on or switch off. The setting only decides whether that message also joins the training pile.
Six Services, Six Different Switches
OpenAI turns training on by default for ChatGPT Free, Plus and Pro. Open Settings, then Data Controls, then switch off "Improve the model for everyone"1. Business, Enterprise, Edu and Healthcare workspaces skip training from the start.
Anthropic made the same choice for Claude Free, Pro and Max in August 2025. A pop-up asked every user to pick a side, with the toggle already sitting on "on"10. Settings, then Privacy, then "Help improve Claude" turns it off. Claude for Work, Enterprise, Education and Gov skip the question entirely3.
5 of 6 Major AI Chatbots Let U.S. Users Turn Off Training
Whether a U.S. consumer account gets a working setting to stop training on new chats, checked Oct. 29, 2026
Each cube is one of the six services in this guide: magenta cubes are services where a U.S. user can switch training off, and the gray cube, Meta AI, limits that switch to the European Union, the United Kingdom and Brazil.
Sources: OpenAI Help Center [1]; Anthropic [3]; Google Gemini Help [5]; Meta Privacy Center [6]; xAI [7]; Microsoft Support [9]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| U.S. users can turn it off | 5 |
| EU, UK, Brazil only (Meta AI) | 1 |
| Whole | 6 |
Google's Gemini calls its version "Keep Activity." On by default, it saves your chats. It trains on them too. Turn it off at myactivity.google.com. Future chats then skip both your history and the model5.
xAI built two doors out of Grok's training. Settings, then Data Controls, flips "Improve the Model" off for chats you keep. Or open a Private Chat: that conversation skips training entirely78.
Microsoft's Copilot works chat by chat too. A personal account's privacy settings carry a toggle for "conversation activity." An organization's Entra ID account skips training by default instead9.
Meta AI is the odd one out. Its privacy center lets European Union, British and Brazilian users object and stop new training. It reserves that form for accounts under those three laws612.
Melissa Heikkilä covered the gap for MIT Technology Review. American users sit outside that process, she found, and so do people in most other countries lacking a comparable privacy law. Every other route she checked fell short of a guarantee.
The Switch Only Covers What Comes Next
Anthropic defends the trade. "By participating, you'll help us improve model safety, making our systems for detecting harmful content more accurate and less likely to flag harmless conversations," the company wrote when it rolled the choice out3. It also promised sharper coding and reasoning in return.
Jennifer King disagrees with the framing. A privacy and data policy fellow at Stanford's Institute for Human-Centered Artificial Intelligence, she told Big Technology the design itself misleads people. The language, she said, is "obscuring what they're really doing by appealing to your social good, essentially"11.
Claude@claudeaiWe're updating our consumer terms and privacy policy to help us deliver even more capable, useful AI models.
Our users will now be given the choice to allow their data to be used to improve Claude.
King's sharper point sits a layer down. A flipped switch protects your next message only. Your last one stays untouched, exactly where your earlier choice put it.
Alex Kantrowitz counted the same six companies. He reached a blunt verdict. "Amazon, Anthropic, Google, OpenAI, Meta, and Microsoft have all built default settings that allow them to train on anything you input into the chatbot window," he wrote11.
Retention Runs on Its Own Clock
Turning training off rarely erases you instantly. Google's own help page says it plainly: "Your future chats won't appear in your Activity, and won't be used to train our AI models, unless you choose to send Google feedback"5. Past chats still take up to 18 months to clear on their own schedule.
Google Cuts an Opted-Out Chat to 3 Days; Rivals Keep It 30
Retention window after a user turns off training, in days, for the services that shorten it
Each column is the number of days a service keeps a chat after training is switched off: Google's drop to 3 days is far shorter than the 30-day window Anthropic and xAI use.
OpenAI's ChatGPT and Microsoft's Copilot keep a saved chat on the same clock whether training is on or off, so each skips a separate opt-out retention number here.
Sources: Anthropic [3]; Claude Privacy Center [4]; Google Gemini Help [5]; xAI [7]; xAI [8]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Anthropic's Claude | 30 days |
| xAI's Grok | 30 days |
| Google's Gemini | 3 days |
Anthropic cuts retention from five years to 30 days the moment a user declines training34. Google's drop is sharper. An opted-out Gemini chat sits for 72 hours, just enough to answer you and guard against a system failure.
That compares against 18 months when Keep Activity stays on5. OpenAI and Microsoft run the opposite way. Each keeps a saved chat on its usual clock regardless of the training setting19.
What to Do
Open the setting before you send anything sensitive. The choice protects only what comes after it. Anthropic's and xAI's own terms say a changed choice covers new or resumed chats alone. Last month's medical question already sits wherever your earlier setting put it37.
Meta AI readers outside the EU, UK or Brazil face a simpler rule. Treat every message as training data from the start612.
Readers of the other five should flip the switch once per account, then check it again after any major policy update. Anthropic's own pop-up proved a company can quietly reset a choice you already made10.
By the Numbers
- Five of six: Major AI chat services that give a U.S. user a working opt-out switch for training on new chats, Meta AI the exception135796.
- Half a decade: How long Anthropic may keep a Claude chat once a user agrees to training. Decline, and the window drops to 30 days3.
- 72 hours: How long Google keeps a Gemini chat once "Keep Activity" is off, down from 18 months when it stays on5.
- 18 months: Default retention for a Microsoft Copilot conversation, regardless of its training toggle9.
- 30 days: How long xAI keeps a Grok "Private Chat" before deleting it, with training skipped entirely78.
- Three: Countries, the European Union, the United Kingdom and Brazil, where Meta's privacy center accepts an objection to AI training612.
- August 2025: Month Anthropic gave Claude's consumer plans a default-on training choice at signup, reversing its earlier policy310.
What to Watch
Watch whether Meta extends its opt-out form to U.S. users the way it already has for the EU, UK and Brazil. A change in U.S. state privacy law could force that move faster than a policy memo would. Anthropic already proved a company can reopen the training question for millions of users in one pop-up. The next quiet reset is only a matter of timing.
Sources
- OpenAI, "Data controls in ChatGPT," OpenAI Help Center, checked Oct. 29, 2026, https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt
- OpenAI, "Data usage for consumer services FAQ," OpenAI Help Center, checked Oct. 29, 2026, https://help.openai.com/en/articles/7039943-data-usage-for-consumer-services-faq
- Anthropic, "Updates to Consumer Terms and Privacy Policy," Anthropic, Aug. 28, 2025, https://www.anthropic.com/news/updates-to-our-consumer-terms
- Anthropic, "How long do you store my organization's data?" Claude Privacy Center, checked Oct. 29, 2026, https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data
- Google, "Gemini Apps Activity," Google Gemini Help, checked Oct. 29, 2026, https://support.google.com/gemini/answer/13594961
- Meta, "How Meta uses information for generative AI models and features," Meta Privacy Center, checked Oct. 29, 2026, https://www.facebook.com/privacy/genai
- xAI, "Privacy Policy," xAI, checked Oct. 29, 2026, https://x.ai/legal/privacy-policy
- xAI, "Consumer FAQs," xAI, checked Oct. 29, 2026, https://x.ai/legal/faq
- Microsoft, "Privacy FAQ for Microsoft Copilot," Microsoft Support, checked Oct. 29, 2026, https://support.microsoft.com/en-us/microsoft-copilot/privacy-faq-for-microsoft-copilot
- Rebecca Bellan, "Anthropic users face a new choice: opt out or share your data for AI training," TechCrunch, Aug. 28, 2025, https://techcrunch.com/2025/08/28/anthropic-users-face-a-new-choice-opt-out-or-share-your-data-for-ai-training/
- Alex Kantrowitz, "Hey, You Should Probably Check Your Chatbot's Privacy Settings," Big Technology, March 6, 2026, https://www.bigtechnology.com/p/hey-you-should-probably-check-your
- Melissa Heikkilä, "How to opt out of Meta's AI training," MIT Technology Review, June 14, 2024, https://www.technologyreview.com/2024/06/14/1093789/how-to-opt-out-of-meta-ai-training/