Learn

What each service does with your data

Updated

In short

  • Five AI services, three types of account: what each provider writes about training, retention, hosting and settings.
  • The type of account matters more than the service: business plans do not train models by default, personal accounts often do.
  • Each row links to its source and shows the date it was checked. Policies change: check the provider’s page before you decide.

ChatGPT (OpenAI)

ChatGPT (OpenAI): what the official pages say, by type of account
Account TrainingRetentionHostingSettings Sources

Free
Yes, unless you opt out. Even after opting out, rating an answer (thumbs up or down) can make the whole conversation available for training. Until you delete the conversation, then erased within 30 days, unless a legal or security obligation applies. Temporary chat: a copy kept for up to 30 days. Outside the European Economic Area: in the United States, or in the countries where OpenAI’s affiliates, partners and service providers are. “Improve the model for everyone” (data controls); temporary chat; deleting conversations.

Checked on


Plus, Pro
Yes, unless you opt out. Even after opting out, rating an answer (thumbs up or down) can make the whole conversation available for training. Until you delete the conversation, then erased within 30 days, unless a legal or security obligation applies. Temporary chat: a copy kept for up to 30 days. Outside the European Economic Area: in the United States, or in the countries where OpenAI’s affiliates, partners and service providers are. “Improve the model for everyone” (data controls); temporary chat; deleting conversations.

Checked on


Business, Enterprise, Edu
Not by default. Rating an answer does not allow training on the conversation. Enterprise and Edu: period set by administrators; a deleted conversation is erased within 30 days, unless a legal obligation applies. Enterprise and Edu: a new workspace can keep its content stored in Europe (or in another region offered). Business: not stated in the pages read. Administrators: retention period (Enterprise, Edu), turning feedback on answers on or off.

Checked on

Claude (Anthropic)

Claude (Anthropic): what the official pages say, by type of account
Account TrainingRetentionHostingSettings Sources

Free
Your choice, with the “Model Improvement” setting. Deleted conversation: erased within 30 days. If you accept training: data kept for up to 5 years, in a form separated from your identity. Conversation flagged as breaking the usage policy: up to 2 years. No storage country stated: Anthropic may process data in the countries where it or its partners operate. “Model Improvement” in the privacy settings; deleting conversations.

Checked on


Pro, Max
Your choice, with the “Model Improvement” setting. Deleted conversation: erased within 30 days. If you accept training: data kept for up to 5 years, in a form separated from your identity. Conversation flagged as breaking the usage policy: up to 2 years. No storage country stated: Anthropic may process data in the countries where it or its partners operate. “Model Improvement” in the privacy settings; deleting conversations.

Checked on


Team, Enterprise, API
Not by default, except feedback you send or explicit agreement. Conversations kept in the product; deleted ones: erased within 30 days. API: erased within 30 days. Flagged conversation: up to 2 years. Data stored in the United States; processing possible in the United States, Europe, Asia and Australia. Enterprise: owners set a retention period (30 days at least); without it, data is kept with no limit.

Checked on

Gemini (Google)

Gemini (Google): what the official pages say, by type of account
Account TrainingRetentionHostingSettings Sources

Personal Google account
Yes: chats may be used to improve Google’s services, including its machine learning technologies. Deleted automatically after 18 months by default (3 months, 36 months or never, your choice). Chats reviewed by a person: up to 3 years, even after you delete your activity. Servers around the world: your data may be processed outside your country. The setting that keeps your activity (« Conserver l’activité » on the French page; turning it off stops your future chats from being reviewed to improve the services); auto-delete; temporary chats.

Human reviewers, including at service providers, look at some of the chats.

Checked on


Google AI subscription on a personal account
As for the free account: the Gemini guide sets different terms only for a work or school account; the subscription appears there only among the information collected. As for the free account. As for the free account. As for the free account.

Checked on


Google Workspace (work or school account)
Not without your permission: no human review and no training outside your organisation. Gemini app: up to 36 months, set by the administrator (18 months by default). From Business Standard up: the administrator can have prompts and responses stored in the United States or in Europe (“data regions”); otherwise, no preference. Gemini Notebook is not covered. Administrators: whether conversations are saved, and how long before automatic deletion (3, 18 or 36 months).

Checked on

Vibe, ex-Le Chat (Mistral AI)

Vibe, ex-Le Chat (Mistral AI): what the official pages say, by type of account
Account TrainingRetentionHostingSettings Sources

Individual plans
Yes by default, unless you opt out. Rating an answer allows Mistral to use that rating, whatever the plan. Conversation erased when you delete it, after a short grace period (length not stated), unless the law requires keeping it. By default in the European Union; temporary transfers possible to the subprocessors Mistral lists. Mobile app: “Enable data sharing” (Data & Account Controls); deleting conversations.

Le Chat has been called Vibe since August 2026. Members of the Mistral team can see a conversation you report.

Checked on


Teams, Enterprise
Enterprise: not by default, the administrator can turn it on. Teams: the administrator can turn it off for the whole organisation. Conversation erased when you delete it, after a short grace period (length not stated), unless the law requires keeping it. By default in the European Union; temporary transfers possible to the subprocessors Mistral lists. Admin console: “Allow your interactions to be used to train our models”.

Checked on

Copilot (Microsoft)

Copilot (Microsoft): what the official pages say, by type of account
Account TrainingRetentionHostingSettings Sources

Personal Microsoft account (Copilot app since 18 August 2026)
No: prompts, responses and files are not used to train the foundation models. Microsoft still uses them to provide and improve the service, advertising included. History kept so you can come back to it; a conversation or the whole history can be deleted at any time. Period not stated in the pages read. In your region, in the United States or in any country where Microsoft, its affiliates or its service providers have facilities. Personalization: memory, shared experience with Bing, Edge and MSN, web search, personalized ads; deleting the history.

The old app followed other rules (training unless you opt out, human review with no way to opt out); its conversations not moved to the new one by 31 December 2026 will be deleted.

Checked on


Microsoft 365 Personal, Family, Premium or Pro (Copilot in Word, Excel, Outlook…)
No: prompts, responses and file contents are not used to train the foundation models. History visible in some apps (Word, Excel, PowerPoint), exportable from the privacy dashboard. Period not stated in the pages read. In your region, in the United States or in any country where Microsoft, its affiliates or its service providers have facilities. “Experiences that analyze content” and “All connected experiences”: turning either off turns Copilot off in Word, Excel, PowerPoint, Outlook and OneNote.

Checked on


Microsoft 365 Copilot (work or school account)
No: prompts, responses and Microsoft Graph data are not used to train the models. Prompts and responses stored, encrypted; period set by the administrator; each user can delete their history. Users in the European Union: data kept within the EU Data Boundary, except for models provided by Anthropic. Administrators: retention policies (Microsoft Purview).

Checked on

Sam
ChatGPT, Claude, Gemini, Vibe, Copilot… which one does what with my conversations?

The table above answers, service by service, with its sources. Here is how to read it.

Four columns, four questions

  • Training

    Are my conversations used to improve the models? Can I opt out?

  • Retention

    How long are they kept? Deleted from the screen, a conversation does not always disappear from the servers straight away.

  • Hosting

    Where are they stored? The provider does not always say for personal accounts.

  • Settings

    What can I, or my organisation’s administrator, change?

The type of account matters more than the service

Personal account, free or paid

Training possible in most cases: it is up to you to opt out.

With Claude, you choose, with the “Model Improvement” setting.

Paying for a personal plan usually changes nothing.

At OpenAI, Plus and Pro follow the rules of the free plan; at Anthropic, Pro and Max follow those of Free. Exception: Microsoft does not train its foundation models on the conversations of its new Copilot app, nor on those of Copilot in Microsoft 365 Personal or Family.

Plan for organisations

No training by default, for all five services: ChatGPT Business, Enterprise and Edu, Claude Team and Enterprise, Gemini in Google Workspace, Vibe Enterprise, Microsoft 365 Copilot.

At Mistral, on the Teams plan, the administrator can turn training off for the whole organisation.

Opting out of training does not settle everything

Training: off. Usually for your future conversations only.

Retention period: unchanged.

Human review: unchanged, depending on the service.

  • At Google, even with the setting that keeps your activity turned off, your chats are used to answer you and for safety, with the help of human reviewers; a chat already reviewed is kept for up to 3 years.
  • At OpenAI, rating an answer thumbs up or down can make the whole conversation available for training, even after opting out, on individual plans.

Hosting does not tell the whole story

  • Mistral

    By default in the European Union, with temporary transfers possible to its subprocessors.

  • Anthropic

    Business plans: data stored in the United States.

  • Google

    Says it has servers around the world.

  • Microsoft 365 Copilot

    European users: within the EU Data Boundary, except for models provided by Anthropic.

Useful, but: where data is stored says nothing about how long it is kept, human review or training. The other columns still need reading.

What the table does not say

  • Your contract. For an organisation, the contract and its data annexes are what count.
  • Changes. In 2026, Mistral’s Le Chat became Vibe, and Microsoft released a new Copilot app. Hence the check date on each row.
  • Your country. Google, for example, gives part of its Gemini guide to users under European Union or United Kingdom law.

Check for yourself, in three steps

  1. Your settings. Under privacy or data controls: note the state of the training setting.
  2. The provider’s help page on data use. The links are in the table.
  3. At work, ask which plan your organisation uses, and whether a contract covers the data.

There are no silly questions

At the office, they tell me “no plan, no contract”. So what?
Then your personal account is not a suitable work tool for your clients’ data.

Your turn: true or false?

I pay for a personal plan, so my conversations are not used for training.

False, as a rule. A personal plan usually follows the rules of the free plan.

My data is hosted in Europe, so nobody reads it.

False. Where data is stored says nothing about human review, retention or training.

Key points

  • A personal account, free or paid, generally does not offer the guarantees of a business plan.
  • Opting out of training removes neither retention nor, depending on the service, human review.
  • Hosting tells you where data is stored, not for how long or who can read it.
  • At work, what counts is the plan and the contract your organisation chose, not your personal account.

Sources

  1. OpenAI, Chat and file retention in ChatGPT, consulted .
  2. OpenAI, How your data is used to improve model performance, consulted .
  3. Anthropic, Is my data used for model training? (commercial plans), consulted .
  4. Google, Guide sur la confidentialité dans les applications Gemini (in French), consulted .
  5. Google, Generative AI in Google Workspace Privacy Hub, consulted .
  6. Mistral AI, Le Chat is now Vibe, consulted .
  7. Microsoft, Privacy FAQ for Microsoft Copilot, consulted .
  8. Microsoft, Data, Privacy, and Security for Microsoft 365 Copilot, consulted .
  9. OpenAI, Europe privacy policy, consulted .
  10. OpenAI, Expanding data residency access to business customers worldwide, consulted .
  11. Anthropic, Configure custom data retention controls for Enterprise plans, consulted .
  12. Google, Google Workspace Data Regions, consulted .
  13. Microsoft, Microsoft Copilot for individuals: your data, privacy, and responsible AI, consulted .
  14. Microsoft, Microsoft Copilot for individuals: your privacy controls and choices, consulted .
  15. Microsoft, Copilot in Microsoft 365 apps for home: your data and privacy, consulted .
  16. Microsoft, Microsoft Privacy Statement, consulted .

This page is for information only. It is not legal advice and does not replace a professional’s view of your situation. Rules change: each source is dated with the day it was consulted.