Before we press send
Privacy information checked on 9 October 2026. This article is a practical comparison of providers' published policies, not an independent technical audit or individual legal advice.
Uploading an invoice, asking for a clearer version of a letter or summarising a lengthy email: these tasks can seem harmless at first. After all, we are simply trying to save ourselves some work.
Yet an invoice may contain a private address and bank details. A letter might discuss an illness, a legal dispute or a family's financial situation. An email may include information its sender never intended to share with an AI provider.
That leaves me with a question which is easy to overlook when we are pleased with a useful answer: what happens to the information after I press “Send”?
I have looked at the published privacy information from OpenAI, Anthropic, Google, Microsoft, Mistral, Meta and DeepSeek. This comparison covers their cloud services, with particular attention to personal accounts. Business services are considered separately where different rules apply. The basis is what the providers say, not an independent technical inspection of their systems.
Not used for training does not mean deleted
Several different processes need to be kept separate when discussing privacy.
A service first processes an input to produce an answer. There is then the question of whether the conversation is stored and can be opened again later. Using it to improve or train models is another matter.
Turning off training permission does not automatically stop storage. ChatGPT conversations, for example, can remain in history even when their use for model training has been disabled. Temporary chats have separate retention rules. [1]
For me, “We do not train on your data” is therefore not enough to assess a service. I also want to know how long content is retained, who may access it and under what circumstances, and what actually disappears when I delete it.
The key differences at a glance
This overview reflects the provider information checked on 9 October 2026. It is deliberately not a ranking: a service can be cautious about model training while still retaining extensive histories or additional information.
OpenAI – ChatGPT: content from personal services may be used for model training, which can be switched off. Business, Enterprise, Edu and the API do not use content for training by default without permission. [2]
Anthropic – Claude: for personal services, training permission matters. Additional rules apply to feedback and safety reviews. [3]
Google – Gemini: the “Keep activity” setting affects training use. Certain processing and temporary retention continue even when it is turned off. [4]
Microsoft – Copilot: for the new consumer app available since 18 August 2026, Microsoft says prompts, responses and file contents are not used to train foundation models. Older versions need to be considered separately. [5]
Mistral – Vibe: inputs and outputs in the standard Vibe service are used for training by default unless the user opts out. Enterprise customers are opted out by default. [6]
Meta – Meta AI: Meta identifies interactions with its AI features and public content shared by adults as training sources. People in the EU can object. [7]
DeepSeek: its privacy policy identifies model improvement as a purpose and describes rights to object. According to the provider, personal data is processed and stored in China. [8]
ChatGPT: Settings help, but do not replace a deletion plan
The OpenAI guidance checked on 9 October 2026 says training use can be turned off for new conversations through the data controls. Voluntary feedback is an important exception: rating an answer may make the associated conversation available for model improvement even when the general training setting is off. [2]
Retention deserves a closer look. Regular and archived chats generally remain until they are deleted or a workspace retention rule applies. After deletion, permanent removal is scheduled within 30 days, subject to exceptions including content already de-identified and disassociated from the account, security needs and legal obligations. Temporary chats may also be retained for safety purposes for up to 30 days. The commonly cited 30-day period is therefore not a general maximum retention period for every ChatGPT conversation. [9]
There is a practical detail too: files stored separately in Library do not automatically disappear when the chat is deleted. The Library copy also needs to be deleted. [10]
My conclusion is that training permission, conversation history and stored files belong together in a privacy review. Simply changing one switch and then uploading every document would not be enough for me.
Claude: Permission can have a lasting effect
In its guidance dated 16 March 2026, Anthropic describes when content from personal Claude accounts may be used for improvement. Alongside permission, it explicitly identifies safety reviews and voluntary feedback. According to the provider, Incognito chats are not used to improve Claude. [3]
The retention rule dated 1 July 2026 is particularly revealing. If chats are made available for model improvement, the relevant data may be retained in de-identified form for up to five years in training pipelines. Otherwise, deleted conversations are due to be removed from back-end systems within 30 days. Exceptions apply for safety cases and legal obligations. Changing the setting later does not undo training already completed; data may also remain in training runs already under way. [11]
I would consider that before agreeing. Permission to improve a service is not simply an informal expression of support. It can materially change how conversation content is used afterwards.
For commercial Claude services and the API, Anthropic's notice dated 18 August 2026 says inputs and outputs are not used for training by default. Explicitly shared content and feedback are again treated separately. [12]
Gemini: Distinguish history, review and connected services
Google's Privacy Hub was last updated on 24 September 2026. It describes several retention periods for personal Gemini use. When activity is kept, activities are automatically deleted after 18 months by default, although this period can be changed. Conversations reviewed for improvement purposes can be retained separately from the account for up to three years. [4]
When “Keep activity” is turned off, Google says future chats are not used to improve its AI models unless feedback is submitted. Retention for 72 hours still applies. Temporary chats are also stored temporarily, and processing for safety can continue. [4]
Google also distinguishes personal use from certain Workspace services. Eligible Workspace licences which include Gemini as a core service provide additional data protections. A business-sounding account name is not enough: the actual licence matters. [13]
For a service used for many different tasks, I would not treat the account type as a minor detail.
Copilot: Older privacy comparisons can now be misleading
Dates are particularly important for Microsoft. Its official guidance distinguishes the older Copilot app from a new version available for web, desktop and mobile devices since 18 August 2026.
For this new app used with a personal Microsoft account, Microsoft explicitly says prompts, responses and file contents are not used to train the underlying foundation models. The accompanying help page also warns against applying these statements to older app versions or other Copilot services without checking. [5]
That does not rule out every other use of data. Microsoft describes separate controls for memory, conversation history, cross-service personalisation and personalised advertising. Turning memory off, for example, does not automatically delete memories already saved. [14]
Work or school organisational accounts have their own protections and contractual conditions. Microsoft also says prompts and responses in those services are not used to train foundation models. [15]
Anyone assessing Copilot should therefore first establish which product they are using. A general statement about “Microsoft AI” would be too imprecise.
Mistral: A European provider is not automatically training-free
Mistral's former assistant Le Chat is now called Vibe. The company explains this change in a notice dated 12 August 2026. [16]
Its privacy guidance dated 31 August 2026 says inputs and outputs in the standard Vibe service are used for training by default. Users can opt out. Enterprise customers are opted out by default, while voluntary feedback is handled separately. [6]
The reviewed version of Mistral's privacy policy has been effective since 3 September 2026. The French company also describes possible transfers to service providers outside the EU. Vibe inputs and outputs are generally retained until the conversation or account is deleted; other purposes may justify further retention. [17]
For me, this is a useful example of why a company's location alone is not sufficient to make a decision. A European provider may be appealing. I would still read its training rules, service-provider arrangements and retention conditions as carefully as those of an American or Chinese company.
Meta AI: Public posts also deserve attention
On 14 April 2025, Meta announced plans to use public content from adults in the EU and interactions with Meta AI for training. Its published statement was amended on 27 March 2026 and describes a way for people in the EU to object. [7]
An important distinction is that, according to this statement, private messages with friends and family are not used for training unless someone in the conversation chooses to share those messages with AI features. A conversation between people is therefore not the same as handing that content to Meta AI. [7]
I would check not only the AI assistant's settings but also which of my own posts are public. Nor would I copy somebody else's messages into an AI chat without asking, simply because it makes a summary convenient.
DeepSeek: Check storage location and retention rules
DeepSeek's privacy policy is dated 10 February 2026. It identifies text inputs, files, voice inputs and chat histories among the content collected. Improving and training the technology is also listed as a purpose. The policy describes rights to object to the use of personal data for training, depending on the applicable rights. [8]
The provider also says it processes and stores personal data in the People's Republic of China. It does not promise one short retention period for all content. Instead, it refers to different purposes, the lifetime of the account, and business and legal requirements. [8]
For confidential information, this general description alone would not be a sufficient basis for me. Before using it, I would want to understand the specific agreement and how deletion and individual rights work in practice.
This review concerns the service operated by the provider. A separately operated model running entirely offline is a different situation: without an external transfer, its inputs do not reach the cloud provider.
Connections to other services bring in more parties
Not every relevant transfer of data is visible in the chat window.
Microsoft's documentation for business Copilot Chat provides a concrete example. For a web search, terms generated from a request and potentially from supplied content may be sent to Bing. The search service operates separately from Microsoft 365 and is subject in part to different privacy conditions. Microsoft says generated queries do not contain user or tenant identifiers or entire uploaded files; they are not a wholesale transfer of the conversation. [18]
For confidential tasks, I would therefore ask about more than the main provider. Connected apps, external search features and additional tools also belong in the review.
An assistant with access to email, files and contacts may save me work. I still want to understand which permissions are genuinely necessary. I would not grant access to an entire mailbox merely to summarise a single message.
Small everyday changes that make a difference
I have the most control before sending anything. Information I do not share does not subsequently need to be protected by the provider on my behalf.
For help drafting a payment reminder, for example, I would start with a general description of the task:
“Please draft a friendly reminder for an overdue invoice. Use placeholders for the name, invoice number, amount and payment deadline.”
I would then add the actual customer details myself. There is no need to upload a complete invoice containing an address and bank details.
I would approach other tasks in the same way: first decide what information the result requires, then share only that part. This is also consistent with data minimisation, as described in the European Commission's GDPR guidance checked on 9 October 2026. Personal data should be limited to what is necessary for the purpose. [19]
There are also three things I would check regularly:
Settings: training permissions, memories and connected services should fit how the service is being used.
Stored content: I would deliberately remove chats and separately stored files that are no longer needed.
Permissions: I would grant access to other accounts or applications only when a specific task requires it.
I would be especially cautious with identity documents, medical records, login credentials and confidential information about other people. General guidance often does not require the complete personal file.
A personal subscription does not prove suitability for customer or employee data
When a company processes personal data, the issue is no longer just an individual's willingness to share information. The European Commission's guidance says the purpose, lawful processing, necessary amount of data and appropriate safeguards must be considered, among other things. [19]
The AI provider's role also needs to be established. Where it acts as a processor, its obligations towards the company acting as controller must be set out in a contract or another binding legal act. The Commission explicitly explains this distinction for IT and cloud services too. [20]
A paid account is therefore not proof that every business use is lawful or sufficiently protected.
For a business, I would decide in advance which tasks are allowed, which data must not be entered and which accounts are intended for use. That is more reliable than leaving each employee to decide when the next urgent assignment arrives.
What I would look for when choosing a service
I do not want to declare an overall winner from this comparison. Products, accounts and intended tasks differ too much for that.
For straightforward wording help, I would primarily share information sparingly and set the privacy controls appropriately. For confidential work documents, contractual terms, access rights, retention periods and actual data routes would matter most to me.
If information must never leave a particular computer or controlled environment, I would choose a solution designed and verified to operate on that basis.
A provider can say it takes privacy seriously. To make a decision, I need more: understandable rules, sensible defaults and a way to meaningfully limit permission once it has been granted.
I want to use AI without casually revealing more about myself and other people than the task requires whenever I receive a helpful answer.
Sources and further reading
The following original sources were read on 9 October 2026. Provider rules may change; stated publication dates are included with each source.
[1] OpenAI: Data controls in ChatGPT
Read source ↗
[2] OpenAI: How your data is used to improve model performance
Read source ↗
[3] Anthropic: Is my data used for model training? — consumer services. 16 March 2026
Read source ↗
[4] Google: Gemini Apps Privacy Hub. 24 September 2026
Read source ↗
[5] Microsoft: Copilot for individuals — activity history. New app available 18 August 2026
Read source ↗
[6] Mistral: Training controls. 31 August 2026
Read source ↗
[7] Meta: Making AI Work Harder for Europeans. 14 April 2025; amended 27 March 2026
Read source ↗
[8] DeepSeek: Privacy Policy. 10 February 2026
Read source ↗
[9] OpenAI: Chat and file retention in ChatGPT
Read source ↗
[10] OpenAI: Using Library to manage files in ChatGPT
Read source ↗
[11] Anthropic: How long do you store my data? 1 July 2026
Read source ↗
[12] Anthropic: Is my data used for model training? — commercial services. 18 August 2026
Read source ↗
[13] Google: Gemini Apps with a work or school account
Read source ↗
[14] Microsoft: Copilot for individuals — privacy controls and choices
Read source ↗
[15] Microsoft: Data protection for work or school Copilot Chat
Read source ↗
[16] Mistral: Le Chat is now Vibe. 12 August 2026
Read source ↗
[17] Mistral: Privacy Policy. Effective 3 September 2026
Read source ↗
[18] Microsoft Learn: Copilot Chat Privacy and Protections
Read source ↗
[19] European Commission: Principles of the GDPR
Read source ↗
[20] European Commission: Application of the GDPR
Read source ↗