Skip to content

AI Fundamentals for Marketers & Creators · Privacy, confidentiality and disclosure · lesson 7 of 16 · 11 min

What never to paste into an AI tool

Why this matters

When you type into an AI tool, your text is sent to the provider's servers. Depending on the provider, the plan and your settings, it may be stored, reviewed for safety, or used to improve future models. For a creator or agency, that creates three kinds of risk:

  1. Privacy risk: exposing personal data about customers, followers or colleagues.
  2. Confidentiality risk: leaking a client's unreleased product, pricing or strategy, often in breach of an NDA.
  3. Security risk: exposing passwords, API keys or account access.

Data protection laws such as the EU and UK GDPR, Pakistan's emerging data protection framework, the UAE's federal data protection law, Saudi Arabia's Personal Data Protection Law (PDPL) and US state privacy laws share common principles: collect only what you need, use it only for the stated purpose, keep it secure, and be transparent. Pasting customer lists into a consumer chatbot can conflict with all of them.

The red list: never paste

  • Passwords, API keys, one-time codes, recovery phrases or login links.
  • Payment data: card numbers, bank details, IBANs, wallet credentials.
  • Government identifiers: CNIC, Emirates ID, Iqama, passport or national insurance numbers.
  • Health, religious, biometric or other sensitive personal data about identifiable people.
  • Customer lists with names, phone numbers or emails, such as your WhatsApp broadcast list or CRM export.
  • Client material under NDA unless the client has agreed and you are using an approved tool with suitable terms.
  • Private messages and DMs from followers that identify them.

The amber list: only in approved tools, minimized

  • Internal strategy documents and pricing
  • Sales call recordings and transcripts
  • Unreleased campaign creative
  • Performance data tied to named accounts

For amber material, use a business tier or company-approved tool where inputs are not used for training, share only the parts you need, and follow your agency's or client's rules.

The green list: generally fine

  • Public information (your published posts, public web pages)
  • Your own drafts and ideas
  • Anonymized or aggregated data ("32 comments asked about delivery times")
  • Generic questions ("How do I structure a product launch email?")

Anonymize before you paste

Often you can get the same value without the risk:

  • Replace names with roles: "Customer A", "Client (skincare brand)".
  • Replace exact figures with ranges, or remove them if they are not needed.
  • Strip phone numbers, emails, order IDs and addresses.
  • Summarize instead of pasting raw: "Three customers complained about late delivery in Jeddah" instead of the full chat logs.

Hands-on: anonymize a real brief in five minutes

Take a real client brief or customer message you'd like AI help with and run it through this before pasting anything.

Replace before pasting:
- People's names            -> roles ("the founder", "Customer A")
- Company names (if NDA)    -> "the client", "Competitor 1"
- Phone, email, address     -> delete
- Order / account / ID nos. -> delete
- Exact prices or revenue   -> ranges or "[PRICE]" unless essential and approved
- Health or other sensitive details -> generalize ("a skin concern")
- Unreleased product names  -> "Product X"

Before (unsafe): "Fatima Khan (+971 50 xxx xxxx) from Noor Clinic says her Botox patient had bruising after the 12 March session. Write a reply."

After (safe): "A clinic client received a complaint from a patient about bruising after a cosmetic treatment. Write a calm, empathetic reply from the clinic that apologizes for the experience, invites the patient to call the clinic directly, and makes no medical claims."

The second version gives the AI everything it needs to help, and nothing it doesn't.

Settings that change the risk

  • Consumer plans (ChatGPT, Claude and the Gemini app on personal accounts): check the model-training setting ("Improve the model for everyone", "Help improve Claude", Gemini's "Keep Activity"), use temporary or incognito chats for one-off questions, review memory and shared links.
  • Business plans (ChatGPT Business/Enterprise, Claude Team/Enterprise, Microsoft 365 Copilot, Gemini in Workspace business editions): business terms apply and providers state they don't train on business content by default; admins control retention, connectors and sharing.
  • Connectors and agents widen the risk: an assistant connected to your Gmail or CRM can read much more than you paste. Connect only what the task needs.

Temporary chats reduce history and training use; they are not permission to paste red-list data.

A worked example

Omar, an account manager at a small Dubai agency, wants help replying to an angry client email. The email includes the client's unreleased launch date, the CEO's mobile number, and a budget figure.

Unsafe: paste the whole email into a personal free chatbot account.

Safe: use the agency's approved AI tool, or rewrite the context: "A client is upset that a campaign deliverable was two days late. They are worried about an upcoming launch. Draft a calm, accountable reply that apologizes, explains the fix and confirms the new timeline." Same useful output, none of the sensitive details.

Video lecture: What never to paste into an AI tool

Lecture coming soon · 10 chapters · about 8 minutes. Read the full transcript below.

  1. What never to paste
  2. Why it matters
  3. The traffic light
  4. The postcard analogy
  5. Anonymize before you paste
  6. Example 1: the WhatsApp complaints
  7. Example 2: the leaked pricing
  8. Watch me do it
  9. Settings that change the risk
  10. Recap and try this now

Lecture transcript

What never to paste

Here's a thirty second habit that could save your agency's reputation. Before you paste anything into an AI tool, you glance at it and ask one question: which color is this? Red, amber or green. That's it. Most AI privacy disasters in marketing don't come from hackers. They come from a busy person pasting a customer list, a client's unreleased pricing, or a patient's message into a chatbot to save ten minutes. In this lecture you'll learn the red, amber and green lists, a five minute anonymization method, and the settings that change your risk. You'll see two examples and watch me anonymize a real style brief.

Why it matters

When you type into an AI tool, your text goes to the provider's servers. Depending on the provider, the plan and your settings, it may be stored, reviewed for safety, or used to improve future models. For a creator or agency, that creates three risks. Privacy risk: exposing personal data about customers, followers or colleagues. Confidentiality risk: leaking a client's unreleased product, pricing or strategy, often in breach of an NDA. And security risk: exposing passwords, API keys or account access. Data protection laws, like the EU and UK GDPR, the UAE's federal data protection law, Saudi Arabia's Personal Data Protection Law and US state privacy laws, share common principles. Collect only what you need, use it only for its purpose, keep it secure.

The traffic light

Here's the traffic light. Red means never paste, in any general AI tool: passwords, API keys and one time codes; card numbers and bank details; government IDs like CNIC, Emirates ID, Iqama or passport numbers; health, religious, biometric or other sensitive personal data about identifiable people; customer lists with names, phones or emails; and client material under NDA unless the client has approved that tool. Amber means only in approved business tools, and minimized: client briefs, internal plans, campaign results, unreleased creative. Green means generally fine: public information, your own non sensitive drafts, general questions. When in doubt, treat it as the next color up.

The postcard analogy

Here's a simple test. Would you write it on a postcard? A postcard passes through many hands on its way, and anyone along the route could read it. Pasting into a consumer AI tool, with default settings, is a bit like that. It's not that anyone will read it, but you've given up control over where it goes and how long it stays. For amber material, you want an envelope and an approved route: a business plan your organization has vetted, with the right terms. And for red material, you don't send it at all. You describe the situation without the sensitive details, which, as you'll see, works just as well.

Anonymize before you paste

The technique that makes most red and amber data safe to work with is anonymization. Before pasting, replace people's names with roles: the founder, customer A. Replace company names under NDA with the client or competitor one. Delete phone numbers, emails, addresses, order numbers and ID numbers. Turn exact prices or revenue into ranges or a placeholder, unless they're essential and the tool is approved. Generalize sensitive details, so a medical condition becomes a skin concern. And rename unreleased products to product X. Here's the key idea. The AI needs the situation, not the identities. You'll be surprised how rarely the output gets worse.

Example 1: the WhatsApp complaints

A simple example. A home bakery in Lahore wants to understand complaints from its WhatsApp Business chats. The owner's first instinct is to export the chats and paste them straight in. But the export has customer names and phone numbers, which is red list material. So she does it properly. She copies only the complaint messages, deletes names and numbers, and replaces them with customer one, customer two. Then she asks her assistant to group the complaints into themes and suggest one fix for each. The insight is identical: most complaints are about late deliveries on Fridays. And no customer's personal data went anywhere it shouldn't.

Example 2: the leaked pricing

Now a realistic business case, with illustrative details. At a small agency in Karachi, a junior designer pastes a client's unreleased price list into a personal free chatbot to make a comparison table. A colleague notices. What should happen? No blame game, but fast action. Delete the conversation. Check the account's training and memory settings, and delete any memory entry. Tell the account lead, who checks the client contract and informs the client if the contract requires it. Then fix the system: the agency adds a one page AI policy, sets up an approved business workspace, and runs a thirty minute team session on the traffic light. The designer wasn't careless. The agency hadn't given them a safe route.

Watch me do it

Let me anonymize one. Here's a complaint a clinic client forwarded to our agency. It says: Fatima Khan, with her mobile number, from Noor Clinic, says her patient had bruising after a treatment on the twelfth of March. Write a reply. Line by line. Fatima Khan becomes the clinic. The mobile number goes. The date goes. The specific treatment becomes a cosmetic treatment. The patient stays the patient. Now my prompt reads: a clinic client received a complaint from a patient about bruising after a cosmetic treatment. Write a calm, empathetic reply that apologizes for the experience, invites the patient to call the clinic directly, and makes no medical claims. The reply is excellent. And the clinic's own team will review it before anything is sent, because it's a health matter.

Settings that change the risk

Settings matter too. On personal plans, check the training setting: improve the model for everyone in ChatGPT, help improve Claude in Claude, and Keep Activity in Gemini. Use temporary or incognito chats for one off questions, and review memory and shared links regularly. But remember, a temporary chat is not a license to paste red list data. On business plans, like ChatGPT Business or Enterprise, Claude Team or Enterprise, Microsoft 365 Copilot, or Gemini in a Workspace business edition, business terms apply and admins control retention and connectors. And connectors widen the risk, because an assistant connected to your Gmail or CRM can see far more than you paste. Connect only what the task needs.

Recap and try this now

Let's recap. Before you paste, classify: red never, amber only in approved business tools and minimized, green generally fine. Anonymize by turning names into roles, deleting contacts and IDs, and generalizing sensitive details. The AI needs the situation, not the identities. And set your training, memory, temporary chat and connector settings on purpose. Here's your try this now. Take one real brief or customer message you'd like help with this week. Run it through the anonymization checklist in the lesson text, then prompt. Afterward, spend five minutes checking your assistant's settings. Two small habits, and most of the risk is gone.

Key takeaways

  • Red list (never paste): credentials, payment data, government IDs, sensitive personal data, customer lists, NDA material in unapproved tools.
  • Amber list: client and internal material only in approved business tools, minimized.
  • Anonymize before you paste: roles instead of names, delete contact and ID details, generalize sensitive facts.
  • Check training, memory, temporary-chat and connector settings; temporary chats are not a license for red-list data.

Try it

Audit the privacy and training settings of every AI tool you use, and write down one change you made.