Can Humans Read Your AI Chats? Copilot, ChatGPT, Gemini and Claude Privacy Compared

What "human review" does and doesn't mean​

None of the providers claims that people read every conversation. Reviews are described as targeted, and they are tied to specific purposes:

• Improving models.
• Investigating abuse or Terms of Service violations.
• Security.
• Support.
• Legal obligations.

Four controls get conflated, and they behave differently:

• Training opt-out: this governs whether your chats feed future models.
• Human review: this can continue for safety and abuse purposes.
• History: this controls what you can see.
• Deletion: this removes the chat from your view, and sometimes from backend systems, on a timetable that varies by provider.

Microsoft Copilot: the part Windows users should read twice​

Microsoft's Copilot Privacy FAQ says some Copilot conversations are subject to both automated and human review for product improvement and digital safety. I'm paraphrasing the FAQ here, since the search results I pulled only covered Anthropic. Microsoft also says conversations may be reviewed when a Code of Conduct violation is suspected. It says limited human review is a required part of that investigation, so there is no opt-out.

Microsoft also says trained AI experts may review conversations to build and evaluate its models. It says it partners with external research organizations that can review conversation logs for safety research. One example it gives is understanding how people seek extremist content.

The training opt-out is narrower than a general privacy switch:

• If you're signed in with a Microsoft account or third-party authentication, you can stop future conversations from being used to train Microsoft's generative AI models.
• That choice doesn't remove the review described above.
• Consumer conversation activity is retained for 18 months by default, and you can delete individual conversations or your history.
• If you aren't signed in, Microsoft says it doesn't train on your conversations.

Work and school accounts are a different story​

Microsoft says the consumer training setting does not apply to the following:

• Copilot users signed in with organizational Entra ID accounts.
• Microsoft 365 consumer users.
• Copilot conversations inside Word, Excel, PowerPoint or Outlook.

Those users don't see the setting, and their conversations aren't used to train the models in the FAQ. Enterprise Data Protection is a separate arrangement. It means your employer's tenant boundary applies. It does not mean the data is invisible. As general IT knowledge, administrators in a managed tenant can typically search and audit logged prompts and responses.

Google Gemini: three years and a deletion catch​

Google's Gemini Apps Privacy Hub says a subset of chats is reviewed by human reviewers, including trained service providers. The stated purposes are:

• Assessing whether responses were low-quality, inaccurate or harmful.
• Suggesting better responses.
• Keeping Gemini safe.
• Addressing violations of the Terms of Service.

Reviewed chats are retained for up to three years. The deletion detail matters most. Google says reviewed chats, plus related data such as language, device type, location and feedback, are not deleted when you delete your activity. They are no longer connected to your Google Account.

This is an exception for chats already reviewed. It is not a blanket three-year hold on everything you've deleted.

Google's controls:

• Keep Activity off: this stops future chats from being reviewed to improve Google services. Even so, Google says chats are still saved for 72 hours to respond to you and protect users, including with human reviewers.
• Temporary chats: these are not used to improve Google AI with human reviewers, but the same protective processing applies.
• Auto-delete: it defaults to 18 months and can be changed to 3 months, 36 months or indefinite.
• Work and school accounts: these may fall under different Workspace terms.

OpenAI ChatGPT: limited access, with a 30-day safety copy​

OpenAI's help center says a limited number of authorized personnel and trusted service providers may access content only as needed for four reasons:

• Investigating abuse or a security incident.
• Supporting you.
• Handling legal matters.
• Improving model performance, unless you've opted out.

OpenAI says access is need-to-know, monitored and logged, and staff must complete security and privacy training.

Model-improvement preferences live under Settings, then Data controls. Temporary Chats stay out of history, don't create or update memories, and aren't used to train models while they remain temporary. OpenAI may still keep a copy for up to 30 days for safety. If you save a temporary chat, it becomes a regular chat under your usual settings.

Cleared chats are deleted from OpenAI's systems within 30 days. The exceptions are chats that were already de-identified and disassociated from your account, or that must be kept for security or legal reasons.

By default, OpenAI doesn't use content from ChatGPT Business, Enterprise, Edu or Healthcare workspaces, or from the API Platform, to improve models. That default is not a promise of zero retention or zero access.

Anthropic Claude: the claim needs trimming​

The TechGig piece says Claude chats are reviewed by a small number of personnel involved in model training. The sources I could verify don't support that framing.

What Anthropic's documentation does show:

• Deletion: consumer deletion guidance says conversations are removed from history immediately and deleted from the backend within 30 days. Exceptions apply for enforcing the Usage Policy, addressing violations, or complying with law.
• Training opt-in: Anthropic's August 2025 announcement said Claude Free, Pro, and Max users can now opt in to having conversations used to improve models, with a five-year retention period for participants. Turning the setting off stops new chats from being used for future training.
• Flagged chats: third-party summaries of the Consumer Terms (for example, a legal-sector blog) say that even if you opt out, Anthropic will still use inputs and outputs for model improvement when a conversation is flagged for safety review or when you report something through the feedback tools. I'd treat that as secondary reporting and check Anthropic's current terms.
• Incognito chats: these are described as not used to improve Claude, even with model improvement enabled.

Anthropic's Clio research paper says a small number of employees can access conversation data for limited business purposes, such as enforcing its Usage Policy. I couldn't open it, so that comes from an evidence brief rather than a document I read. Neither source shows that ordinary chats are routinely read by human reviewers for training.

Be careful with retention figures. Third-party sources describe 30 days with training off and five years with it on. One commentator notes the 30-day figure no longer appears in the current policy text. Check your own account's setting rather than relying on a blog.

Side-by-side comparison​

Where the TechGig summary overreaches​

• Identical policies: the four companies don't share one policy. Purposes, retention and controls all differ.
• Menu paths: the article lists settings paths for each service. Providers change these, and what you see depends on your app version and account type, so treat them as a starting point.
• Paid plans: the article says paid subscriptions don't improve privacy. That may hold for consumer plans, but Microsoft's FAQ doesn't make that claim.
• Enterprise protection: business tiers typically add contractual protections, but that is not the same as no access or no retention.

What to do about it​

1. Know which account you're using. Consumer Copilot, work Copilot Chat, Gemini on a personal account and Gemini on a Workspace account are different products with different rules.
2. Don't paste secrets into consumer chatbots. That means passwords, API keys, customer data and financial details. Both Google and OpenAI explicitly say not to enter anything you wouldn't want a reviewer to see.
3. Use placeholders. Swapping real names and identifiers for stand-ins reduces exposure. It doesn't guarantee anonymity.
4. Don't treat temporary modes as invisible. They limit history and training, but safety processing and short retention can remain.
5. Follow your organization's policy. For work use, stick to approved services and ask IT about logging, retention and audit access.
6. Consider local models for sensitive work. TechGig suggests tools such as Ollama and LM Studio so that data isn't uploaded to cloud providers. That's reasonable general advice, though you take on setup and maintenance.

The takeaway​

"Private" in an AI chat window means the chat is private from other users. It does not mean private from the company. Review is limited and purpose-bound, and the details vary by provider. Policies also change, so recheck your settings periodically.