Millions of people share personal questions, financial concerns, health information, and confidential documents with AI assistants. But how private are those conversations—and what can you do to protect them?
Imagine telling an artificial intelligence assistant something you would hesitate to share with another person. Perhaps you are worried about a medical diagnosis. You want advice about a troubled relationship. You are reviewing your finances, preparing a confidential business proposal, or asking for help with a sensitive family matter.
The conversation feels private. There is no human being sitting across from you, no judgmental expression, and no apparent audience. Just you and a machine. But the apparent intimacy of that interaction can create a misleading impression.
When you type a message into a cloud-based AI service, the information generally travels to computers operated by the provider, where it is processed to generate a response. Depending on the service, your settings, and the circumstances, conversations may be stored, analyzed by automated systems, or reviewed by authorized personnel.
Some conversations may also be used to improve AI models. That does not mean employees are routinely reading every message. Nor does it mean your conversations are publicly accessible.
It does mean that a chatbot conversation should not automatically be treated as confidential in the same way as a private conversation protected by professional secrecy. Understanding that distinction is increasingly important as AI becomes part of everyday life.
Who Can Actually Read Your Conversations?
The answer depends on which service you use and how you use it. AI companies generally process conversations automatically. Their systems need access to the text you submit to understand the request and generate an answer. But automated processing is not the same as human review.
Companies may authorize limited access to conversations for purposes such as investigating misuse, responding to security incidents, evaluating model performance, or improving the service.
Google, for example, explicitly states that a subset of Gemini conversations may be reviewed by trained human reviewers, including external service providers.
Microsoft acknowledges that some consumer Copilot conversations undergo automated or human review for product improvement and safety.
Anthropic says certain Claude conversations may be analyzed when flagged for safety review, even if the user has not otherwise agreed to contribute conversations to model improvement.
OpenAI similarly describes controls governing model improvement, retention, and safety-related handling of ChatGPT conversations.
These practices differ, but they illustrate an important point: the fact that a conversation takes place through a chatbot interface does not guarantee that its contents will never be accessible to a human reviewer.
The Difference Between Storing a Chat and Training an AI
Two concepts are frequently confused: conversation history and model training.
A service may store a conversation so you can return to it later. That does not necessarily mean the conversation will be used to train future AI models. Conversely, deleting a conversation from your visible history does not necessarily reverse processing that already occurred.
Training involves using data to improve how AI systems perform. Depending on the company and product, eligible conversations may contribute to that process. Privacy settings can limit this use, but they do not always eliminate other forms of processing.
For example, a company may still need to process a conversation to answer the user’s question, detect misuse, or comply with legal obligations. That is why turning off model training should not be confused with making a conversation invisible to the service provider.
It is an important privacy protection, but not an absolute guarantee of confidentiality.
ChatGPT: How to Limit the Use of Your Conversations
OpenAI provides several privacy controls for personal ChatGPT accounts. In personal Free, Plus, and Pro workspaces, the setting that allows conversations to help improve models is enabled by default, according to OpenAI’s documentation.
Users can disable it without losing their ordinary conversation history. To change the setting:
- Open ChatGPT and select Settings.
- Choose Data controls.
- Turn off Improve the model for everyone.
OpenAI says new conversations will not be used to train its models after this setting is disabled. Users can also start a Temporary Chat.
Temporary chats do not appear in the normal conversation history unless the user chooses to save them, and they are not used for model improvement while they remain temporary.
OpenAI may retain a copy for up to 30 days for safety purposes.
Another consideration is memory and personalization. Depending on the options selected, ChatGPT may use information from previous interactions to make future conversations more relevant.
Users who prefer less personalization can review the available controls under Settings → Personalization.
For a particularly sensitive question, choosing a temporary, unpersonalized conversation can reduce the amount of information carried into the interaction.
However, temporary mode should not be interpreted as an invitation to share passwords, confidential client records, or other information that should not be uploaded to an online service.
Google Gemini: Pay Attention to Human Review
Google’s privacy disclosures contain an especially important warning.
When Gemini’s Keep Activity setting is enabled, Google may use conversations and other submitted information to improve its services, including AI models. A subset of conversations may be reviewed by trained human reviewers.
Google says conversations sent to external reviewers are disconnected from the user’s account. However, reviewed conversations may be retained for up to three years, even if the user subsequently deletes the associated activity.
This is a significant distinction between deleting a visible chat and removing information that has already entered a separate review process.
Users can limit future data use by changing their activity settings.
To do so:
- Open Gemini.
- Access Settings and find Activity or Gemini Apps Activity.
- Turn off Keep Activity.
Google says future chats made with this setting disabled generally will not be used to train its AI models unless the user submits feedback that permits additional processing.
Those conversations may still be retained for 72 hours to operate and protect the service.
Gemini also offers temporary chats, which are not used for ordinary model training and are retained for a limited period.
Google cautions that certain connected features may become unavailable when activity storage is disabled.
For consumers, the practical lesson is that privacy settings can involve trade-offs between personalization, functionality, and data retention.
Claude: Training Choices and Safety Exceptions
Anthropic, the company behind Claude, also distinguishes between conversations used for ordinary service delivery and those used to improve its models.
According to Anthropic’s consumer privacy documentation, users can choose whether to allow their chats and coding sessions to contribute to model improvement.
The company says conversations may also be used or analyzed when flagged for safety review or when users participate in certain voluntary testing programs.
Users can review the relevant preference in Claude’s privacy settings.
Anthropic says deleted conversations are ordinarily removed from its back-end storage within 30 days, subject to applicable exceptions. There is an important retention difference for users who agree to contribute conversations to model training. Anthropic states that eligible data may be retained in de-identified form for up to five years in its training systems.
That does not mean every Claude conversation is kept for five years. The extended period applies to the training-related circumstances described in its policy.
The lesson is that users should review not only whether a service offers an opt-out, but also what happens to data they have already agreed to share.
Microsoft Copilot: Training and Personalization Are Separate
Microsoft allows users of its consumer Copilot service to control whether their conversations may be used to train generative AI models.
Users can also manage personalization separately. This distinction matters because a person may want an assistant to remember useful preferences without allowing their conversations to contribute to general model training.
According to Microsoft’s privacy documentation, consumer Copilot conversations are stored for 18 months by default, although users can delete individual conversations or their entire history.
Microsoft also acknowledges that some conversations may undergo human review for safety or product-improvement purposes.
To review your settings:
- Open Microsoft Copilot and sign in.
- Go to Settings.
- Review the controls for model training and personalization.
- Disable either feature according to your preferences.
Microsoft notes that consumer Copilot and organizational Microsoft 365 Copilot experiences may have different privacy and training rules.
This distinction is especially important for people who use AI both at home and at work.
Four AI Assistants, Four Sets of Rules
Although the services offer broadly similar capabilities, their privacy policies differ in important ways.
| AI service | Model-training control | Important privacy consideration |
|---|---|---|
| ChatGPT | Personal users can turn off model improvement. | Temporary chats may be retained for up to 30 days for safety. |
| Google Gemini | Users can turn off Keep Activity. | Previously human-reviewed chats may remain for up to three years. |
| Claude | Users can control whether eligible chats improve Claude. | Training data may be retained in de-identified form for up to five years when users opt in. |
| Microsoft Copilot | Consumer users can opt out of model training. | Conversations are stored for 18 months by default; limited human review may occur. |
These descriptions concern consumer services. Business, enterprise, educational, and specialized accounts may have different contractual protections and retention policies.
Privacy settings can also change as products evolve.
Users should consult the current documentation for the specific service and account they use.
The Risks Are Not Limited to Chat History
Even when an AI company offers strong privacy controls, other risks remain. Consider someone who uploads a medical report containing a full name, date of birth, address, and patient identification number.
The assistant may need the medical details to explain the report. But it does not necessarily need the person’s identity or administrative information. A safer approach would be to remove unnecessary identifying details before uploading the document.
The same principle applies to tax returns, bank statements, employment records, legal correspondence, and business documents. There is also the question of connected applications. Some AI assistants can interact with email, calendars, cloud storage, websites, and third-party services.
When a user authorizes these connections, information may be transferred to another company or processed under a separate privacy policy. The privacy controls governing the chatbot itself may not apply in the same way to every connected service. Users should therefore review which applications have access to their accounts and disconnect services they no longer need.
Seven Ways to Protect Your Privacy When Using AI
The good news is that users do not need advanced technical knowledge to reduce many of these risks.
1. Turn off model training if you do not want your conversations used for that purpose. Review the setting separately in every AI service you use. Do not assume that changing it in one application affects another.
2. Use temporary conversations for questions you do not want saved in ordinary chat history. Remember that temporary does not always mean immediately deleted or inaccessible for legitimate safety purposes.
3. Remove identifying details before uploading documents. Names, addresses, account numbers, signatures, and identification codes are often unnecessary for the task you want AI to perform.
4. Never share passwords, authentication codes, or private security credentials. An AI assistant generally does not need this information to explain a problem or provide guidance.
5. Review memory and personalization settings. An assistant that remembers preferences can be convenient, but not every detail needs to become part of a continuing profile.
6. Check connected applications and sharing permissions. Email, cloud storage, workplace systems
