CyphrQ
Menu

Is It Safe to Use ChatGPT at Work? 7 Things to Keep Out Before You Paste

Is it safe to use ChatGPT at work? Yes, with the right account and settings. Seven things to keep out, how to check data settings, and a team policy.

By CyphrQ.ai

Written with AI assistance and human editorial review. Every figure linked to its source.

Updated 11 min read

Is it safe to use ChatGPT at work? For most everyday tasks, yes, if you use an account your company has approved and you keep seven kinds of data out of the chat box. The real risk is rarely a hack. It's what you paste, which account you paste it into, and who can see it afterwards.

This guide covers what to keep out, why it matters, how to check the data settings in ChatGPT, Claude, Gemini and Microsoft Copilot, and a one-page policy you can hand to your team. For IT and security teams, there's a technical section on enforcing that policy.

What should you never paste into an AI chatbot at work?

  1. Customer personal data. Names, emails, phone numbers, ID numbers, order histories. If a real person could be identified from it, keep it out unless your company has cleared that tool for personal data.
  2. Passwords, API keys and access tokens. A key pasted into a chat is a key stored on someone else's server. If it happens, rotate the key. Deleting the chat is not enough.
  3. Unreleased financials. Results, forecasts, pricing and deal terms before they're public.
  4. Contracts and anything under NDA. Asking for "a quick summary" still means handing over the full text.
  5. Source code that contains secrets. Code in an approved tool is often fine. Code with hard-coded credentials, internal server names or customer data isn't. Strip those first.
  6. Health data. Medical notes, sick-leave reasons, insurance claims. The EU's GDPR treats health data as a special category with stricter rules (Article 9).
  7. HR and employee records. Performance reviews, salaries, disciplinary notes.

A quick test before you paste: would you be comfortable if this text turned up in a document outside the company? If not, remove the sensitive parts or don't paste it.

Why does it matter what you paste into a chatbot?

Four reasons, all documented by the vendors themselves.

  • Training. On consumer plans, your conversations can be used to improve the vendor's models unless you turn that off. ChatGPT, Claude, Gemini and Copilot all have a setting for it.
  • Retention. Opting out of training doesn't always mean your chats are deleted straight away. Anthropic keeps consumer Claude chats for 30 days if you opt out of training, and for five years if you opt in (Anthropic). Google keeps Gemini chats for 72 hours even with Keep Activity off (Google).
  • Human review. Google says a subset of Gemini chats is read by human reviewers, and reviewed chats can be kept for up to three years. Google's own advice is not to enter anything confidential that you wouldn't want a reviewer to see (Google).
  • Shared links. If you share a ChatGPT conversation from a personal account, anyone with the link can view it, and they can forward it. In Business, Enterprise and Edu workspaces, links only open for members of the same workspace (OpenAI).

In practice: in May 2023 Samsung temporarily restricted generative AI tools such as ChatGPT on company computers, after Bloomberg reported that some staff had uploaded sensitive code to ChatGPT. Samsung confirmed the restriction to CNBC. In a company survey the month before, 65% of respondents said they were concerned about the security risks of generative AI services (CNBC, via NBC News). No attacker was needed. The code left through the clipboard.

Is a business AI account safer than a personal one?

Usually, yes. The difference is the contract and the defaults.

Assistant Personal (consumer) account Work (business or enterprise) account
ChatGPT Can be used for training unless you turn off "Improve the model for everyone" ChatGPT Business and Enterprise data is not used for training by default (OpenAI)
Claude Free, Pro and Max users choose via the model training setting Claude for Work and the API run under Anthropic's Commercial Terms, which the consumer training change doesn't cover (Anthropic)
Gemini Keep Activity controls whether chats help improve Google AI With a Google Workspace work or school account, content isn't human reviewed or used to train models outside your organisation without permission (Google)
Microsoft Copilot "Training on conversation activity" and "Training on voice conversations" toggles With a work account, prompts, responses and your organisation's Microsoft 365 data aren't used to train foundation models (Microsoft)

Copilot's personal and work versions have different data terms. Check which account you're signed in with: that tells you which terms apply.

The breach data points the same way. Shadow AI means staff using AI tools the company hasn't approved and can't see. IBM's 2025 Cost of a Data Breach report covered 600 organisations with breaches between March 2024 and February 2025. It found that:

  • one in five organisations reported a breach due to shadow AI;
  • organisations with high levels of shadow AI saw breach costs an average of $670,000 higher;
  • of organisations whose AI models or applications were breached, 97% reported lacking proper AI access controls (IBM).

For scale, IBM's 2026 edition, a separate report, puts the global average cost of a breach at $4.99 million (IBM).

My view: a personal AI account used for company data is a blind spot. Your company can't manage it, can't set retention for it and can't delete it when you leave.

How do you check your data settings in ChatGPT, Claude, Gemini and Copilot?

These menus move. The steps below were checked against each vendor's help pages on 24 October 2026. If a menu has changed, the linked official page wins.

  • ChatGPT: Settings, then Data Controls, then "Improve the model for everyone". The setting applies to your whole account. For a one-off sensitive task, Temporary Chat isn't used for training, isn't saved to your history, and may be kept for up to 30 days, including for abuse review (OpenAI).
  • Claude: Privacy Settings at claude.ai/settings/data-privacy-controls, then the model training setting (Anthropic).
  • Gemini: Gemini Apps Activity, then Keep Activity. With it off, future chats aren't used to train models unless you send feedback (Google).
  • Microsoft Copilot (personal account): on copilot.com, select your profile, then Privacy. On Windows or macOS, go to Settings, then Privacy. Turn off both training toggles. Opting out only covers future conversations, and Microsoft may still use data for safety, security and some other purposes (Microsoft).

One habit covers all four: before you paste, check which account you're signed in with.

What do GDPR and the UAE PDPL have to do with AI at work?

Not legal advice, but this is why your legal team will care: pasting a customer's details into an outside AI tool counts as processing personal data under the EU's General Data Protection Regulation (GDPR) and the UAE's Personal Data Protection Law (PDPL), and usually means sharing it with another company. Under GDPR, a company that handles personal data on your behalf needs a contract with you (Article 28), and a personal chatbot account doesn't give you one.

I cover the UAE law in detail, including free zones, cross-border transfers and a checklist for business owners, in UAE Data Protection Law (PDPL) and AI Tools.

What should an AI-at-work policy include?

A good AI-at-work policy names the approved tools, lists what never goes in, and says who to tell when something slips.

Here's a one-page version you can copy. Adapt it, then have IT and legal check it.

[Company] AI assistant policy (version 1, [date])

  1. Approved tools: [list, for example "ChatGPT Enterprise via company sign-in"]. Don't use personal AI accounts for work data.
  2. Never enter: customer or employee personal data, passwords, API keys or tokens, unreleased financials, contracts or NDA material, source code containing secrets, health data. [Exceptions only where a tool is approved for that data type.]
  3. Strip before you paste: replace names with roles, real numbers with placeholders, and remove keys.
  4. Settings: model training off on any account used for work, unless it's an admin-managed business account.
  5. Sharing: no public share links for work chats. Delete links when you're done.
  6. Check the output: you're responsible for what you send. AI can be wrong.
  7. No unchecked actions: AI tools may not send, pay or delete anything without a person checking first.
  8. If something slips: tell [contact] within [X hours] and rotate any exposed key. Reporting quickly is never punished.
  9. Owner and review: [name] owns this policy and reviews it every [6] months.

Start with item 1 this week: pick an approved tool and tell people which one it is. Most of the other items get easier once that's done.

Going deeper: for security and IT teams

The threat model here is data egress, not intrusion. The actor is usually a well-meaning employee, the asset is whatever they paste, and the failure is that it lands in an account outside your retention, legal hold and offboarding. A written policy only holds if the approved path is the easy one and the unapproved path is visible.

Identity first. Put the approved assistant behind your identity provider with single sign-on (SSO). Create and remove accounts through SCIM or just-in-time provisioning, so leavers lose access when their directory account is disabled. Verify your email domain with the vendor so work addresses land in the company workspace, not a personal account. On Claude Enterprise, the audit log records SSO, JIT provisioning and domain verification changes, so you can see if someone weakens them (Anthropic).

Egress control at the endpoint and browser. Microsoft Purview documents this pattern well. Its DSPM for AI covers third-party sites such as ChatGPT, Google Gemini and DeepSeek. On Windows devices onboarded to Purview, Endpoint DLP can warn or block a user who pastes sensitive information, such as credit card numbers, into ChatGPT. Most features need the Purview browser extension. The alternatives are network data security through a SASE or SSE integration, or browser data security built into Edge. Ready-made policies include "DSPM for AI - Detect when users visit AI sites" and "Block elevated risk users from pasting or uploading sensitive info on AI sites". The second uses Adaptive Protection to give risky users a block-with-override in Edge, Chrome and Firefox (Microsoft Learn).

The same page lists two limits that affect budget and evidence. Third-party AI coverage requires pay-as-you-go billing. Retention and eDiscovery of prompts only work in Edge, and only for ChatGPT, Microsoft's consumer chat, Gemini and DeepSeek. If you run a different CASB or secure web gateway, the design carries over: categorise AI domains, allow the sanctioned tenant, coach or block the rest, and inspect paste and upload events for sensitive data types.

What the audit trail should capture. For each event, capture:

  • the user, as your identity provider knows them;
  • the timestamp and the device;
  • the app, and whether the session was the managed tenant or a personal account;
  • the data type that matched;
  • the action taken: allowed, warned, blocked or overridden.

Vendor logs cover the tenant side. On Claude Enterprise, Owners or Primary Owners export logs from Organization settings, then Data and Privacy, then Export logs. An export covers the past 180 days, and the emailed link works for 24 hours. Each entry carries the actor, event type, IP address, device ID and user agent. Chat titles and content are excluded. Organisations using customer-managed encryption keys can't use the export button. The same events are available through Anthropic's Compliance API (Anthropic).

In Purview, capturing prompts and responses from third-party AI apps needs a collection policy set to capture content. Reading them needs membership of the Content Explorer Content Viewer role group (Microsoft Learn). For ChatGPT Enterprise, ask OpenAI which compliance export your plan includes, which events it covers and how long logs are kept.

Detection and testing. Alert on DLP overrides, on repeated blocks for one user, and on a rising count in Purview's "Sensitive interactions per AI app" report. To test, run the policy in test mode. Paste a dummy card number from your payment provider's test set into ChatGPT on a managed device. Confirm what the user saw and that the event was logged. Reports can lag behind the event, so check again later before you call the test a failure.

Ask the vendor: Do audit logs hold content, or only metadata? Can we stream them to our SIEM, and how long are they kept? Which regions store chats and logs? Can an admin remove a leaver's content?

Is it safe to use ChatGPT at work? Key takeaways

  • Use your company's approved business account for work. A personal account has no contract between your company and the vendor.
  • Keep seven things out: customer personal data, credentials, unreleased financials, contracts, source code with secrets, health data, and HR records.
  • Turning off model training is not the same as deleting your data. Check retention as well.
  • Shared chat links can travel further than you planned.
  • Put the rules on one page so people actually read them.
  • For IT: enforce the policy with SSO, browser or endpoint DLP on AI sites, and an audit trail that shows managed versus personal sessions.

Sources