AI Privacy Scorecard

Which AI apps train on your chats?

Each scorecard answers seven questions, from training to deletion, using quotes from the company's own privacy policy and help pages. Every quote is re-checked word for word each week, and changes are logged.

8 apps · 152 quoted passages · scored on a published rubric. Last weekly check: .

Scorecards

ChatGPT

OpenAI · Chat assistant

12 of 15

Trains on your chats?
On by default, setting to turn it off · 2 of 3
Opt-out
Turn off "Improve the model for everyone" in ChatGPT's settings, or use OpenAI's Privacy Portal.

As of

Claude

Anthropic · Chat assistant

14 of 15

Trains on your chats?
On by default, setting to turn it off · 2 of 3
Opt-out
Turn off "Help improve our AI models" in Claude's privacy settings.

As of

DeepSeek

Hangzhou DeepSeek Artificial Intelligence · Chat assistant

7 of 15

Trains on your chats?
On by default, opt-out by request · 1 of 3
Opt-out
DeepSeek's documents describe opting out of training as a right exercised by request, and name an email address for it.

As of

Gemini

Google · Chat assistant

12 of 15

Trains on your chats?
On by default, setting to turn it off · 2 of 3
Opt-out
In Gemini's History settings, uncheck "Chat content including text & uploads". Turning History off entirely also stops training on future chats unless you send feedback.

As of

Grok

SpaceXSI · Chat assistant

11 of 15

Trains on your chats?
On by default, setting to turn it off · 2 of 3
Opt-out
Turn off "Improve the model" in Grok's data controls, on the app and on Grok.com.

As of

Microsoft Copilot

Microsoft · Chat assistant

9 of 15

Trains on your chats?
Off by default · 3 of 3
Opt-out
Microsoft says chats in the updated app are not used to train foundation models, so there is no training setting to change. You can turn off ad personalization.

As of

Mistral Vibe

Mistral AI · Chat assistant (formerly Le Chat)

9 of 15

Trains on your chats?
On by default, setting to turn it off · 2 of 3
Opt-out
Turn off "Allow your interactions to be used to train our models" in Vibe's privacy settings, or untick "Enable data sharing" in the mobile app.

As of

Perplexity

Perplexity AI · AI search assistant

8 of 15

Trains on your chats?
On by default, setting to turn it off · 2 of 3
Opt-out
Turn off the AI data retention toggle in your account preferences.

As of

All answers side by side

AppScoreTraining on your chatsHow long chats are keptDeleting your chatsPeople reading your chatsSale and advertisingNo-history modeDated policy and past versions
ChatGPT12/15On by default, setting to turn it off · 2 of 3Specific period or rule · 2 of 2Delete it yourself, removal time stated · 2 of 2May be reviewed to improve the service · 1 of 2Not sold, may be used for ads · 1 of 2Yes, with stated retention · 2 of 2Dated, past versions published · 2 of 2
Claude14/15On by default, setting to turn it off · 2 of 3Specific period or rule · 2 of 2Delete it yourself, removal time stated · 2 of 2Limited to stated cases · 2 of 2Not sold, not used for ads · 2 of 2Yes, with stated retention · 2 of 2Dated, past versions published · 2 of 2
DeepSeek7/15On by default, opt-out by request · 1 of 3Specific period or rule · 2 of 2Deletion available, no removal time · 1 of 2Not stated · 0 of 2Not sold, not used for ads · 2 of 2None described · 0 of 2Dated, no past versions found · 1 of 2
Gemini12/15On by default, setting to turn it off · 2 of 3Specific period or rule · 2 of 2Deletion available, no removal time · 1 of 2May be reviewed to improve the service · 1 of 2Not sold, not used for ads · 2 of 2Yes, with stated retention · 2 of 2Dated, past versions published · 2 of 2
Grok11/15On by default, setting to turn it off · 2 of 3Specific period or rule · 2 of 2Delete it yourself, removal time stated · 2 of 2May be reviewed to improve the service · 1 of 2Sale not addressed · 0 of 2Yes, with stated retention · 2 of 2Dated, past versions published · 2 of 2
Microsoft Copilot9/15Off by default · 3 of 3General wording only · 1 of 2Deletion available, no removal time · 1 of 2May be reviewed to improve the service · 1 of 2Not sold, may be used for ads · 1 of 2None described · 0 of 2Dated, past versions published · 2 of 2
Mistral Vibe9/15On by default, setting to turn it off · 2 of 3Specific period or rule · 2 of 2Deletion available, no removal time · 1 of 2Not stated · 0 of 2Not sold, not used for ads · 2 of 2None described · 0 of 2Dated, past versions published · 2 of 2
Perplexity8/15On by default, setting to turn it off · 2 of 3Specific period or rule · 2 of 2Deletion available, no removal time · 1 of 2Not stated · 0 of 2Not sold, may be used for ads · 1 of 2Yes, details limited · 1 of 2Dated, no past versions found · 1 of 2

Scores add up the points in the rubric. They describe what the documents say, not how the apps behave in testing.

Not covered yet

These apps are popular but their documents could not be read in a way this project allows. Read them yourself at the links.

How it works

  1. Read the company's own words. Privacy policies and help pages, fetched politely and quoted exactly.
  2. Score them on a fixed rubric. Seven questions, each answer defined in advance. Read the rubric.
  3. Watch for changes. A weekly job re-reads every document and flags changed text and missing quotes. Change log · RSS.