Is It Safe to Put Work Information Into ChatGPT, Claude, Copilot or Gemini?

It can be, if you use the right account and keep certain things out of it. What happens to what you type depends on three things: the vendor, the kind of account you're signed in with (personal or work), and a few settings.

This is for small offices where people already use ChatGPT, Claude, Copilot or Gemini for work, often on a free personal account, and there's no AI policy yet. It covers what never goes in, how personal and work accounts differ, the settings each vendor documents, and a short policy you can adopt. It doesn't rank the four assistants. They're listed alphabetically.

What never goes into an AI chatbot#

Whatever the account or settings, keep these out:

• Passwords, MFA codes and recovery codes. No assistant needs them to help you.
• Payment card and bank details.
• Health information about anyone.
• Customer and employee personal data: names with addresses, dates of birth, Social Security numbers, account numbers.
• Anything under an NDA or a contract that limits who can see it.
• Keys and tokens that give access to a system, such as API keys.

There are two reasons. First, some vendors have people read a sample of chats. Google's Gemini privacy page says human reviewers see some chats, and asks you not to enter "confidential information that you wouldn't want a reviewer to see". Second, on some personal accounts, what you type can be used to train future models unless you change a setting.

If you need help with a document that contains any of this, take those parts out first. Swap in made-up values, such as "Jane Doe" and an example.com address, and put the real ones back yourself afterwards.

Personal and work accounts are treated differently#

The same vendor often makes different promises for a work account than for a personal one. Microsoft's pages show the difference most clearly, so they're the worked example here.

Microsoft's work accounts get "enterprise data protection". When you sign in to Copilot Chat with a work or school account (Microsoft calls these Microsoft Entra accounts), Microsoft applies what it calls enterprise data protection (EDP). That's Microsoft's name for the commitments in its Data Protection Addendum and Product Terms. As of October 29, 2026, Microsoft's EDP page says:

• Prompts and responses are covered by the same contractual terms as your email in Exchange and your files in SharePoint.
• Prompts, responses and the data Copilot reaches through Microsoft Graph "aren't used to train foundation models".
• Web searches Copilot makes go to Bing as a few words, with user and tenant identifiers removed. Bing handles them under separate terms.

A work account isn't private from your employer, though. Microsoft's work-or-school page says prompts and responses are logged, and your IT admin can view them with Microsoft's search and audit tools.

Microsoft's personal accounts don't get EDP. With a personal Microsoft account, Copilot is governed by the Microsoft Services Agreement and the Copilot Terms of Use, and the Microsoft Privacy Statement explains how your data is used (Microsoft Copilot for individuals).

The other three vendors draw a similar line, in their own words:

• Anthropic says it doesn't use inputs or outputs from its commercial products (Claude for Work and the API, among others) to train its models by default (Anthropic Privacy Center, commercial). The exception is feedback: if you report feedback or bugs, such as with the thumbs up or down buttons, Anthropic says it may use those chats for training. The consumer plans work differently (see the Claude section below).
• Google says a work or school Google Account may be under different data terms. Its Workspace privacy hub says your content "is not human reviewed or otherwise used for Generative AI model training outside your domain without permission".
• OpenAI's work admin FAQ says ChatGPT Enterprise's commitments include "no training on business data by default".

What this means in practice: if your business already pays for Microsoft 365 or Google Workspace, find out which AI assistant comes with those accounts and what it covers. Do that before anyone uses a personal account for work. That's a question of what you already have, not a recommendation of one vendor.

The privacy settings to check in each assistant#

These are the settings each vendor's own pages describe, as of October 29, 2026. Menu names are given exactly as the vendor writes them. If yours look different, the vendor's page is the place to check.

ChatGPT#

OpenAI's Data Controls FAQ and its privacy pages blocked automated access when this was checked on October 29, 2026, so this section doesn't describe them. Before you put work content into a personal ChatGPT account, see OpenAI's current guidance, the Data Controls FAQ.

What OpenAI's ChatGPT Learn pages do say: memories let ChatGPT carry useful context from earlier chats (Personalize ChatGPT), and you manage ChatGPT memory from Settings > Personalization (ChatGPT Learn: Memories). For ChatGPT Enterprise, see the line above. OpenAI's work admin FAQ says the commitments depend on the plan, and it doesn't say the same for other work plans, so check the terms for yours.

Claude#

This covers the consumer plans: Claude Free, Pro and Max.

• Model training. Anthropic uses your chats and coding sessions to improve its models if you allow it. The setting is under Settings > Privacy, labelled Help Improve our AI models (how to change it). Turning it off is a change: Anthropic says it stops using your chats in future training runs. Data already in a training run, or in a model already trained, stays there. While the setting is on, Anthropic says it may keep your data in a de-identified form for up to 5 years in its training pipelines.
• Safety review. Chats flagged by Anthropic's safety systems may still be used for its safety work, whatever the setting says (Anthropic: model training, consumers).
• Incognito chats aren't used to improve Claude, even if the training setting is on.
• Deleting a chat removes it from your history immediately and, as the standard timeframe, from Anthropic's back-end storage within 30 days (Anthropic: how long data is stored). There are exceptions: chats flagged as breaking Anthropic's Usage Policy can be kept for up to 2 years, and Anthropic says it may keep chats when the law requires it or to resolve disputes.
• Thumbs up or down sends feedback that stores the whole conversation for up to 5 years.

Copilot#

This covers Microsoft Copilot with a personal Microsoft account, in the updated app, available since August 18, 2026. Microsoft's privacy controls page puts these in Settings > Personalization:

• Saved memories. Copilot remembers details from your chats. Turning the toggle off is a change, and it doesn't delete memories already saved. To remove them, Microsoft says to use Manage, then Delete all memories.
• One shared experience. Your Copilot chats can personalize Bing, Edge and MSN, and the other way round. Turning it off is a change.
• Web search. Copilot can search the web for current information. Turning it off is a change.
• Allow ads personalization. Without a Microsoft 365 subscription you might see ads in Copilot. Microsoft says personalized ads can use your chat history and saved memories. Turning this off is a change.

Individual chats can be deleted from the chat list.

One more setting matters on a shared or work computer. Under Settings > Web browsing > Import browser data, the Bring over your browsing data from Microsoft Edge toggle copies data from Edge into Copilot. Microsoft says that can include cookies, history, payment info, passwords and autofill data, and that the setting is only available to Microsoft accounts based in the United States for now. Turning it on is a change. I'd leave it off on any computer used for work.

Model training: that privacy controls page lists no training setting for the updated app. Microsoft's Privacy Statement says Microsoft "may use your data to develop and train our AI models". The older app's privacy FAQ describes a training opt-out, but it says it applies only to that older version.

Gemini#

This covers the Gemini app with a personal Google Account. From Google's Gemini Apps Privacy Hub:

• Keep Activity is a setting in Gemini Apps Activity. While it's on, Google says a subset of chats is read by human reviewers to improve its services. Reviewed chats are kept for up to three years, even if you delete your activity.
• Turning Keep Activity off is a change. Google says your future chats then aren't used to train its models unless you send feedback. They're still kept for 72 hours.
• Temporary chats aren't used to train Google's AI models.
• Auto-delete. Activity is deleted after 18 months by default. Changing it to 3 months, 36 months or never is a change.
• Even with Keep Activity off, Google says it still uses chats to respond to you and to protect its users and the public, including with human reviewers.

A five-line AI policy for a small office#

A written rule beats everyone guessing. Here's a short one for a made-up business, Harbor Street Bookkeeping. Change the names and keep it to a page. It's a starting point for everyday use, not legal advice. If your business handles regulated data, check what those rules require too.

1. Use the AI assistant that comes with our work accounts, signed in with your work account. Don't put work content into a personal AI account.
2. Never paste passwords, MFA codes, payment details, health details, or client or staff personal data.
3. Before you paste a document, take out names, addresses and account numbers, or swap in made-up ones.
4. A person checks every AI answer before it goes to a client, into the books or into a file. These tools make mistakes; Google's own Gemini page says so.
5. Ask the office manager before connecting an AI tool to email, shared files or any other account.

Checklist#

• Know which account you're signed in with: personal or work.
• Nothing from the "never goes in" list, in any account.
• On a personal account used for work, check the training setting, where the vendor offers one.
• Know what the assistant keeps (memory, history, activity) and delete what you don't want kept.
• Check the answer before you use it.
• Put the policy in writing and share it with everyone who uses these tools.

A fuller framework, if you want one#

For a larger organization, or one with regulators to answer to, NIST (the US National Institute of Standards and Technology) publishes an AI Risk Management Framework. NIST says it's intended for voluntary use, and that version 1.0 is being revised. Most small offices won't need it to start with. The policy above covers the everyday risks.

Next steps#

AI chatbots are one more online service your business uses, and the same basics apply. The Small Business Cybersecurity Checklist covers MFA, password managers and vendor security, and there's a printable version for the office wall. If your office runs on Microsoft 365, Microsoft 365 for a Small Office covers the admin accounts and MFA that protect the work account your AI assistant signs in with.