You might tell ChatGPT things you’d never share with a stranger. Maybe you’ve asked it about a medical concern, argued with your spouse in a conversation about relationship advice or uploaded a document containing financial information.
But before your next conversation, here's something to know: Real people at AI companies may be able to read some of what you share.
OpenAI, Google, Microsoft and Anthropic all have policies that allow certain conversations to be reviewed by humans, whether to improve their AI systems, investigate potential safety violations or troubleshoot problems.
That doesn’t mean someone is sitting behind a screen reading every message you send, but the assumption that your AI conversations are strictly between you and the chatbot is not accurate. Not to mention, just deleting a conversation doesn’t always mean the company deletes every copy immediately. In Google’s case, Gemini chats reviewed by humans can remain in company systems for up to three years, even after you delete your activity.
Here’s what actually happens to your conversations, who might be able to read them and what you can do to protect your privacy.
Yes, humans can read some AI conversations
When you type a question into ChatGPT, Gemini or Claude, your message is typically sent to the company’s servers, where an AI model processes it and generates a response.
Most of that process is automated. But AI companies also use human reviewers to evaluate certain conversations. Several reasons might prompt a review; safety is a big one. AI companies use automated systems to detect potentially harmful requests, policy violations and suspicious activity. Some interactions may be escalated for human review.
Another reason conversations are reviewed is for improving the AI itself. Depending on your settings and the company’s policies, selected conversations may be used to evaluate responses, identify mistakes and help improve future models.
Companies generally describe measures intended to protect users during these reviews, including limiting access and removing identifying information. But removing a person’s name or account details doesn’t necessarily remove every sensitive fact contained in a conversation.
For example, a chat about a medical diagnosis, workplace dispute or financial problem could contain personal information even without the user’s name attached.
The important difference is that human review doesn’t happen to every conversation, and the rules vary depending on which chatbot you’re using.
What big tech might do with your conversations
ChatGPT
OpenAI allows consumer ChatGPT conversations to be used to improve its models unless users opt out. You can disable this by opening Settings > Data Controls and turning off Improve the model for everyone.
But disabling training doesn’t mean every conversation becomes inaccessible to human reviewers.
OpenAI’s Data Controls FAQ explains that Temporary Chats aren’t used for training but may be reviewed to monitor abuse. Temporary Chats also don’t appear in your regular chat history and are generally deleted from OpenAI’s systems after 30 days.
For ordinary saved conversations, OpenAI says deleting a chat removes it from your account immediately and schedules it for permanent deletion within 30 days, unless the information has already been de-identified or must be retained for legal or security reasons.
Google Gemini
Google is particularly explicit about how human reviewers may access conversations.
According to Google’s Gemini Apps Privacy Hub, some conversations are reviewed by humans to improve its services and help protect users.
Google says conversations sent to service providers for review are disconnected from users’ accounts before being examined.
But there’s a significant catch: Chats reviewed by humans can be retained for up to three years, and deleting your Gemini Apps activity doesn’t remove those reviewed conversations.
Google also allows users to turn off its Keep Activity setting, which prevents future conversations from being used to train its AI models unless users submit feedback.
Even with that setting disabled, conversations may be retained for up to 72 hours to provide the service and for safety-related purposes. Google says some human review can still occur to help protect its systems and users. Essentially, disabling activity history reduces how your conversations are used, but it doesn’t guarantee that no human will ever see them.
Microsoft Copilot
Microsoft is unusually direct about human review. Its Copilot Privacy FAQ confirms that some conversations undergo automated and human review for product improvement and digital safety.
Conversations suspected of violating Microsoft’s Code of Conduct may also be examined. Microsoft says users cannot completely opt out of human review because limited review is necessary when investigating suspected violations.
You can opt out of having future conversations used to train Microsoft’s generative AI models, but the company explains that this doesn’t prevent other uses, including security, compliance and certain product improvements.
That means switching off model training doesn’t provide a blanket guarantee against human access.
Anthropic Claude
Anthropic also allows certain conversations to be reviewed, although it places restrictions on how that information is accessed. According to Anthropic’s explanation of who can view Claude conversations, conversations used for model improvement are separated from identifying account information before review.
Access is limited to a small number of personnel involved in model training, and Anthropic says it uses tools to filter or obscure sensitive information.
Users can control whether eligible conversations are used for model improvement through Claude’s privacy settings.
However, Anthropic also makes an exception for safety-related processing. Conversations flagged by its safety systems may still be used to detect harmful content, enforce policies and improve safety systems.
Claude’s Incognito Chats aren’t used for general model training and aren’t saved to chat history, but they’re still retained for a limited period, generally 30 days.
As with the other platforms, opting out of model training and preventing all possible review are two different things.
Deleting your AI chats doesn’t always erase them
One of the biggest misconceptions about chatbot privacy is that deleting a conversation means it disappears completely. In reality, deleting a chat from your visible history and permanently removing it from company systems can be separate processes.
OpenAI generally schedules deleted ChatGPT conversations for permanent removal within 30 days, subject to exceptions for legal or security obligations and previously de-identified information.
Google’s policy is particularly worth clarifying because Gemini conversations that have been reviewed by humans can remain stored for up to three years, even after the user deletes their Gemini Apps activity.
This doesn’t mean every deleted conversation is kept for three years. The policy applies to conversations that have undergone human review. But it does highlight why users shouldn’t assume deleting a chat is equivalent to immediately erasing every copy of the information.
It’s also worth remembering that turning off model training usually affects how conversations are used, not necessarily how long they’re stored or whether they can be accessed for safety reasons.
Upgrading your AI doesn't mean more privacy
Paying for ChatGPT Plus or another premium chatbot doesn’t guarantee privacy. It’s easy to assume that it might, that paying for a subscription gives you stronger privacy protections, but subscribing to ChatGPT Plus, Claude Pro or a paid Gemini plan doesn’t automatically make your conversations confidential or exempt them from the provider’s consumer privacy policies.
Products such as ChatGPT Enterprise and business-focused AI services often provide additional contractual protections, administrative controls and restrictions on the use of customer data for model training.
But those protections vary by provider and product. They shouldn’t be mistaken for a universal guarantee that no authorized person can ever access customer information.
For everyday users, the safest assumption is that paying for an AI chatbot primarily provides access to additional features and capabilities, not the same confidentiality protections you’d expect from a doctor, lawyer or other professional relationship.
How to stop your AI chats from being used for training
The good news is that all four major platforms provide controls that can reduce how your conversations are used.
Here’s where to find them:
- ChatGPT: Open Settings > Data Controls and turn off Improve the model for everyone. You can also use Temporary Chat for conversations you don’t want saved in your regular history or used for model training.
- Gemini: Open Settings & help > Activity and turn off Keep Activity. This prevents future conversations from being used to train Google’s AI models unless you provide feedback. You can also use Temporary Chats.
- Claude: Open Settings > Privacy and review Help improve our AI models. You can also use Incognito Chats, which aren’t used for general model training.
- Microsoft Copilot: Open your profile settings and navigate to Privacy > Training on conversation activity. Turn that setting off to prevent future eligible conversations from being used to train Microsoft’s generative AI models.
These settings can make a meaningful difference, but none should be interpreted as a promise that conversations can never be reviewed for safety, security or legal reasons.
Five ways to keep your AI conversations more private
If you’re using AI regularly, a few precautions can significantly reduce the amount of personal information you expose.
1. Don’t share anything you wouldn’t want someone else to see. Avoid entering passwords, Social Security numbers, banking credentials, sensitive medical records or confidential workplace documents into consumer AI chatbots. Even when a company has privacy protections, minimizing what you share is the simplest safeguard.
2. Remove identifying details before submitting information. If you’re asking AI to analyze a document or help with a sensitive situation, consider replacing real names, addresses, account numbers and other identifying details with placeholders. Be mindful that context can sometimes reveal someone’s identity even after obvious identifiers have been removed.
3. Turn off model training. Check your chatbot’s privacy settings rather than assuming the default configuration is the most private. The option to prevent conversations from being used for model improvement is separate from deleting chats or disabling memory.
4. Use temporary or incognito conversations when appropriate. ChatGPT, Gemini and Claude offer options designed to limit saved history and model training. These modes don’t necessarily prevent short-term retention or safety-related review, but they can reduce unnecessary exposure.
5. Consider running AI locally for particularly sensitive work. Tools such as Ollama and LM Studio can run supported AI models directly on your computer. When properly configured to process information locally without sending it to external services, they can reduce the need to upload sensitive information to a cloud provider. Just remember that local AI applications may still have optional online features or network connections.
Final thoughts
The companies behind ChatGPT, Gemini, Claude and Copilot have different policies, but they share one important reality: Some conversations may be accessible to human reviewers under certain circumstances.
Understanding the privacy settings, thinking carefully about what information you share and recognizing that deleting a conversation or paying for a subscription doesn’t necessarily make it completely confidential. The safest approach is to treat AI chatbots as useful assistants, not a private journal.
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds.
More from Tom's Guide
- I asked ChatGPT, Gemini, Claude and Perplexity if there really is an AI bubble and when it’s going to burst — and one compared it to the dot-com crash
- ChatGPT has a clever new way to fight back against election deepfakes — I tried it and it actually works
- A few tech leaders are rebranding AI as ‘superintelligence’ — calling it that doesn’t make it accurate
