Picture a small business owner pasting a customer email into ChatGPT to draft a reply: name, address, order and invoice amount, all straight into the box. Before hitting enter, they need to check which data is necessary, whether their organization approves this use and which data terms apply to the account. The plan’s price alone answers none of those questions.
The last post in this series told you to paste the real thing in, because real material beats describing it. That is still true. This post is the other half of that advice: there is a line, and crossing it is the one beginner mistake that can actually cost you. The good news is that staying on the right side of it takes about half a second once it becomes a habit.
Why This Is Worth Two Minutes#
The issue is not limited to free accounts. Cyberhaven’s February 2026 AI Adoption and Risk Report describes substantial use of personal accounts at work. Such observations are a reason to check access and data approvals. They do not prove every personal chat becomes public or that paying alone protects confidential material.
The input box is not a private notebook. On personal ChatGPT accounts, including Plus and Pro, model training depends on data controls. Free users can opt out too. Business, Enterprise and Edu data is not used for training by default. This is a training rule, not a promise that data is never stored, processed or reviewed where permitted.
The Postcard Rule#
The postcard rule is a cautious starting point: until you have checked approval and the service’s data terms, enter only material that this recipient is allowed to receive. That applies to free and paid accounts. The analogy helps with the first decision; it does not mean your chats are actually public.
Three Buckets#
Sort what you want to paste into three buckets and the decision gets easy.
Green, usually straightforward. General questions, entirely invented examples and your own public copy without additional confidential or personal information. Public data can still involve other people’s rights. For third-party texts and real people, check whether you may use them for the task.
Yellow, edit and check first. Customer emails, quotes and contract clauses often contain more than names and numbers. Remove unnecessary identifiers, metadata and details that could identify the person or company indirectly. Placeholders reduce the data supplied; they do not guarantee full anonymization.
Red, only through an explicitly approved process. Personal customer lists, health or financial records, confidential source code and material under confidentiality obligations need an appropriate legal basis, data minimization and reviewed contract and security terms. A Business subscription alone does not authorize processing. Passwords, API keys and other credentials do not belong in its chat text either.
The One Trick: Anonymize, Don't Omit#
You can often do the same writing task with less data. Replace identifying parts with placeholders, send only the required excerpt and put real details back yourself afterwards. Check the result and the remaining material. Depending on the content, this is reduction or pseudonymization, not guaranteed anonymization.
Take a follow-up email as the example. The basics are in How to Prompt Claude. Instead of "write a follow-up to Frau Berger about her four thousand euro patio quote," you write "write a follow-up to [CUSTOMER] about her [AMOUNT] [PROJECT] quote, warm tone, under a hundred words." The AI writes an equally good email. You paste her real name and the real number back in before you send. This reduces the data transmitted. Still check whether the project description, unusual circumstances or hidden metadata could identify someone. Quality and remaining risk depend on the actual material.
Two Settings on Day One#
Two settings help, but do not replace approval for the actual use.
First, check the training choice in Settings → Data Controls. In ChatGPT, turning off “Improve the model for everyone” prevents new conversations from being used for that purpose. Chat history is still retained. Temporary chats have different retention rules and may be kept for safety for up to 30 days; they do not promise immediate deletion.
Second, use only the service and contract your organization has reviewed for business data. OpenAI does not train on Business, Enterprise and Edu workspace data by default. Also review processing, retention, access and any required data processing agreement. The plan label does not replace checking the actual content or the necessary legal basis.
The Mistakes Beginners Make#
Uploading the whole document when you needed two paragraphs. The more you put in, the more there is to leak, and the model rarely needs the rest.
Pasting a screenshot with data still visible in a corner. The AI reads the entire image, not just the part you were thinking about.
Dropping in passwords, API keys, or logins to "let it help with the setup." Never. Those belong in a password manager, not a chat window.
Believing training choices and storage are the same. They are not. Check history, deletion, retention and permitted review, including on a paid account.
Make It a Habit, Not a Worry#
The point of all this is not to make you nervous about a tool you just started to like. It is to turn one instinct into a reflex, the same way you already lower your voice when you say a customer's name out loud in a busy café. Reduce the yellow material, use an approved and reviewed process for red material, and supply only what the task needs. If you want the deeper, legal side of this, our guide to using AI in a GDPR-compliant way goes there. The next post in this series gets practical again: a plain list of the everyday jobs where this all starts paying off, the dozen tasks worth trying first thing Monday morning.
Sources and product status#
OpenAI data controls / Codex access and data terms
Cyberhaven: February 2026 study / EDPB: anonymization and identifiability
