AI Privacy at Work: What You Can and Can't Paste

What happens to text you paste into an AI tool at work depends on three things: whether the service may use it to train its models, how long it keeps your chats, and what agreement your company has with the vendor. The business versions of the big tools (ChatGPT Business and Enterprise, Claude for Work, Gemini in Google Workspace, Microsoft's work Copilot) say they do not train on your data by default. The personal versions of ChatGPT, Claude, Gemini and Copilot can use your chats to improve their models, depending on a setting you control.

The everyday risk is often not training at all. It is that your pasted customer list sits in a chat history for months, where anyone who gets into your account can read it, and where a policy you never read decides when it goes away. Knowing where your data goes after you press send is the difference between a helpful assistant and a compliance problem.

In No-Code AI Automation: Your First Real Workflow, you built your first automation. Now you are ready to feed AI real work data, and before you paste that spreadsheet, you need to know where it goes. Every policy below was checked against the vendor's own pages in October 2026. Policies change often, so check the links before you rely on them.

Rule zero: your company's AI policy comes first

Before any of the details below, find your company's AI policy and its list of approved tools. It may live on the intranet or with IT. It overrides everything in this article. If your company says "use only our Copilot" or "no customer data in any AI tool", that is the rule, even where a vendor's terms would allow more. If there is no policy, ask IT or your manager before you paste anything that is not already public.

A simple way to think about AI data handling

Think of pasting into an AI like handing a note to a colleague. The colleague reads it, uses what they already know to respond, and then something happens to the note. They might shred it soon after. They might file it in a locked cabinet for a set time. Or they might also use it to train new colleagues, unless you tell them not to.

Different AI tools handle your note in different ways. The tool's policy, not its intelligence, decides which path your note takes.

What happens the moment you paste?

The AI does not read your text the way a person does. It splits your words into small pieces called tokens (you met them in Tokens: The Currency of AI) and feeds them to a model that writes a response one token at a time.

All of this happens inside the model's context window, the amount of text it can look at in one go (Context Windows: The Model's Working Memory explains it). Your pasted text and the conversation so far must fit inside it. The model's knowledge, stored in numbers called weights, does not change while it answers you. Answering is a read only operation.

So pasting does not, by itself, teach the model anything. Your text shapes this one answer, and the model's permanent knowledge stays the same. There is one exception to keep in mind. Many chat apps now have a memory feature, and when it is switched on, the app may save facts from your chat and use them in later conversations. OpenAI's page on Memory in ChatGPT describes how this works and how to turn it off. That is the app keeping notes about you rather than the model learning, but for your privacy the effect is similar.

Where do logs and histories fit in?

While the model forgets your text after answering, the service around it keeps records: so you can scroll back through your history, to check for abuse, and on personal plans sometimes to improve future models. How long varies a lot. Here is what the vendors say, as of October 2026.

• ChatGPT, personal plans. Chats stay until you delete them; a deleted chat is then removed within 30 days, unless OpenAI must keep it for security or legal reasons (chat and file retention). A Temporary Chat stays out of your history but may be kept up to 30 days for safety checks.
• Claude, personal plans. A deleted chat leaves Anthropic's storage within 30 days. If you allow Anthropic to use your chats to improve Claude, it may keep them, de-identified, for up to five years (How long do you store my data?).
• Gemini app, personal account. With Keep Activity on, chats are deleted after 18 months by default (you can pick 3 or 36 months, or never). Chats read by human reviewers are kept up to three years, even if you delete your activity. With Keep Activity off, new chats are still kept for 72 hours (Gemini Apps Privacy Notice).
• Gemini, work account. In the Gemini app, your admin deletes chats after 3, 18 or 36 months, with 18 as the default (Generative AI in Google Workspace Privacy Hub). The Gemini side panel in Gmail, Docs and the other Workspace apps keeps its own history, kept forever by default unless your admin sets an auto delete period (Gemini in Workspace conversation history).
• APIs (application programming interfaces, the doors developers use to build AI into other software). OpenAI may keep API data for up to 30 days to spot abuse (Enterprise privacy at OpenAI); Anthropic deletes it within 30 days (How long do you store my organization's data?).

The four numbers below come from the same pages by OpenAI, Anthropic and Google.

This is the split that matters. The model itself does not learn from your paste. But the service may store it, and on personal plans it may later use it to train a new model. That is where the real privacy risk lives.

How do business plans protect your data?

Business and enterprise versions add contracts and controls that personal accounts lack. They treat your prompts and responses as your company's data, under terms like those that cover your email and documents.

• Google Workspace. Google says your Gemini content is not reviewed by humans or used to train generative AI models outside your organization without permission, and that prompts count as customer data under its Cloud Data Processing Addendum (Generative AI in Google Workspace Privacy Hub).
• Microsoft. Microsoft 365 Copilot, which Microsoft now calls Microsoft Copilot, answers from your emails, documents and calendar. Microsoft says prompts, responses and that data are not used to train its foundation models (Data, Privacy, and Security for Microsoft Copilot). This applies when you sign in with your work account. Your prompts and Copilot's answers are stored with your company's other Microsoft 365 content, where admins can apply retention policies and search them with compliance tools such as Microsoft Purview. Copilot with a personal Microsoft account is a consumer product: Microsoft uses those conversations for AI training unless you opt out, and keeps them 18 months by default (Privacy FAQ for Microsoft Copilot).
• ChatGPT Business, Enterprise and Edu. OpenAI does not train on content from these workspaces by default, and admins control how long data is kept (Enterprise privacy at OpenAI). ChatGPT Business was called ChatGPT Team until August 2025 (rename FAQ).
• Claude for Work (Team and Enterprise) and the Anthropic API. By default, Anthropic does not use inputs or outputs from its commercial products to train its models (Is my data used for model training?).
• Slack. Slack will not use customer data to train generative AI models unless the customer opts in, and its AI runs on models hosted inside Slack's own cloud setup, out of the model providers' reach. Its older, non generative features (channel and emoji suggestions, search ranking) do learn from workspace data unless a workspace owner emails Slack to opt out (Slack's privacy principles).
• Apple. Apple Intelligence handles many requests on the device and sends bigger ones to its Private Cloud Compute servers, where Apple says the data is not stored or made accessible to Apple (Apple Intelligence and privacy). The exception is the optional ChatGPT extension: Apple asks before sending a request to ChatGPT, and if you are signed in to a ChatGPT account, OpenAI's policies and your ChatGPT settings apply (ChatGPT Extension and Privacy).

The common thread: business plans promise no training by default, retention your company controls, and access that follows your existing permissions. They are not magic. They are bound by stronger rules, and only while you are actually using the business version.

Turn training off on a personal account

If you use a personal account for anything, even public material, switch training off. The settings as of October 2026:

• ChatGPT: Settings > Data controls > Improve the model for everyone (Data controls in ChatGPT). Temporary Chat is not used for training.
• Claude: Settings > Privacy, then turn off the setting that lets Anthropic use your chats to improve its models (how to change it). New users make this choice when they sign up. Incognito chats are not used for training.
• Gemini: open Gemini Apps Activity and turn Keep Activity off, or use a temporary chat.
• Microsoft Copilot with a personal account: opt out of model training in Copilot's privacy settings.

Two catches apply almost everywhere. Turning training off covers new chats; it does not delete old ones, so delete those separately. And rating an answer with a thumbs up or down can send that whole conversation for review or training even when training is off. OpenAI, Anthropic and Google all say so on the pages above.

What data is safe to paste where?

Alex, our operations manager, has a spreadsheet of customer orders. It includes names, email addresses, and order totals. She wants the AI to summarize it. Can she paste it safely?

The answer depends on which tool she uses and what her company has set up. Here is a practical guide, after rule zero.

Public or already published information is safe with almost any AI tool. Press releases, public product descriptions, and industry reports carry little privacy risk.

Internal but not sensitive material, like meeting notes or draft reports, can go into the business tools your company has approved, since those do not train on your data by default.

Personal data needs more care. Alex's spreadsheet holds customer names and email addresses, and those count as personal data under privacy laws such as the EU's GDPR (General Data Protection Regulation), where the European Commission's own examples of personal data include a name and an email address, and Canada's PIPEDA (Personal Information Protection and Electronic Documents Act). That does not mean she can never use AI on it. It means she should use only a company approved business tool and strip out anything the task does not need. To total orders by region, the AI does not need names at all: she can delete those columns or replace the names with "Customer 1, Customer 2" before pasting.

Confidential intellectual property and sensitive personal data, such as health records or financial account numbers, need the tightest controls. Use only systems your company has approved for that kind of data, covered by a data processing agreement (DPA, the contract that limits what the vendor may do with your data). Never paste this into a personal account.

Secrets like passwords, API keys, or private encryption keys should never be pasted into any online AI service. Even business tools can expose them in logs, error messages, or support tickets. Treat every AI input box like a public channel when it comes to secrets.

Quick reference: major AI tools and their privacy stance

Checked in October 2026; check the links in Sources before you rely on a row.

The Gemini API row comes from Google's Gemini API terms, which ask you not to send sensitive, confidential or personal information to the free tier.

Frequently Asked Questions

Q: Can my company see what I paste into an AI tool? If you use a company provided tool, assume yes. OpenAI says ChatGPT Business admins can view, export and delete members' conversations, and ChatGPT Enterprise admins can get an audit log of conversations. Microsoft lets admins search Copilot activity with its compliance tools. Treat a work AI account like your work email.

Q: What if I accidentally paste a password? Change the password immediately. Then report the incident to your IT or security team. They may need to rotate related credentials and check if the tool's logs could expose the secret. Do not assume the AI "forgot" it.

Q: Does using incognito mode help? A private or incognito browser window only stops your browser from keeping local history. The AI service still receives your text and stores it under its own rules. The tools' own private modes are different: ChatGPT's Temporary Chat, Claude's incognito chats and Gemini's temporary chats are not used for training, though they may still be kept for a short time for safety checks (OpenAI says up to 30 days, Google says 72 hours).

Q: Are there AI tools that run entirely on my device? Yes. Some open models can run on a laptop, and Local Inference: Running Models on Your Own Machine shows how. These keep your data off remote servers, but they are usually less capable than the big cloud models. Apple Intelligence sits in between: it handles many requests on the device and sends bigger ones to Apple's own servers, which Apple says keep nothing after answering, while its optional ChatGPT extension sends requests to OpenAI.

Q: How do I know if my company has an enterprise agreement? Ask IT; that is the only reliable answer. Do not go by the login screen. Signing in with your work email does not prove you are on a business plan, because anyone can open a personal account with a work address. Even a sign in page that shows your company's name and logo proves only that your company runs your work login; the tool behind it can still be a personal account. Then check inside the tool. ChatGPT and Claude both let one person keep a personal account and a company workspace under the same email and switch between them from the account menu (ChatGPT Business FAQ, Claude's help page), so look at which one is selected before you paste. In Gemini and Copilot, check that you are signed in with your work account, not a personal Google or Microsoft account. If you are not sure, assume the personal rules apply.

Test yourself

Alex, the operations manager, needs to draft a performance review for a team member. The review contains feedback about specific projects, sales numbers, and a note about a medical leave the employee took. Alex uses Microsoft Copilot with her work account, which her company has configured with standard retention policies. She pastes the draft into Copilot to ask for phrasing suggestions. Is this safe? What should she check first?

Answer: First, her company's AI policy: performance reviews are HR records, and many companies have stricter rules for them. If the policy allows it, a work Copilot is the right kind of tool, because Microsoft does not use those prompts to train its models. But the whole review is personal information about a named employee, so the feedback and the sales numbers deserve care too. The medical leave note is the real problem. Health information is sensitive personal data under many privacy laws, and a phrasing task does not need it, so she should remove it before pasting. She should also know that nothing vanishes: her prompt and Copilot's answer are stored with the company's other Microsoft 365 content, where retention policies apply and compliance tools can find them. The safest version is to strip the health detail, use a role instead of the name, then ask for phrasing help.

Next: Building Your Personal AI Stack. It pulls the series together into a set of tools you can actually use, from free to paid, with the privacy rules from this part built in.

If you want clear, no nonsense breakdowns of how workplace AI actually works, subscribe to Internals Decoded at internalsdecoded.com.

Sources

• Data controls in ChatGPT (OpenAI)
• How your data is used to improve model performance (OpenAI)
• Chat and file retention in ChatGPT (OpenAI)
• Memory in ChatGPT (OpenAI)
• Enterprise privacy at OpenAI
• ChatGPT Business rename FAQ (OpenAI)
• ChatGPT Business general FAQ (OpenAI)
• Is my data used for model training? Consumer plans (Anthropic)
• How long do you store my data? (Anthropic)
• How do I change my model improvement privacy settings? (Anthropic)
• Is my data used for model training? Commercial products (Anthropic)
• How long do you store my organization's data? (Anthropic)
• Move your personal Claude account to a Team or Enterprise organization (Anthropic)
• Gemini Apps Privacy Notice (Google)
• Generative AI in Google Workspace Privacy Hub (Google)
• Find and manage your Gemini in Workspace conversation history (Google)
• Gemini API Additional Terms of Service (Google)
• Data, Privacy, and Security for Microsoft Copilot (Microsoft)
• Privacy FAQ for Microsoft Copilot, personal accounts (Microsoft)
• Privacy Principles: Search, Learning and Artificial Intelligence (Slack)
• Apple Intelligence and privacy on iPhone (Apple)
• ChatGPT Extension and Privacy (Apple)
• Data protection explained (European Commission)
• PIPEDA in brief (Office of the Privacy Commissioner of Canada)