ChatGPT for work: how to use it safely and not share unnecessary data

ChatGPT is a draft assistant, not a responsible person

The best work scenarios have two characteristics: the result is easy to verify, and an error does not trigger an irreversible action. This includes preparing headline options, converting notes into an outline, explaining complex text in simple words, creating a table template, or finding questions worth asking at a meeting.

Caution is required where an answer becomes the basis for payment, dismissal, medical treatment, a legal position, a credit decision, or a public statement. The model can convincingly formulate an inaccuracy, invent a source, or miss an important exception. Therefore, the final decision, access to the system, and responsibility remain with the human.

Data traffic light: green, yellow, red

Before every request, classify not the document as a whole, but specifically the fragment you plan to send. If in doubt, move to a stricter level and ask the data owner or security officer.

Anonymization is not just replacing a name with "Client A". Combining a job title, city, amount, and date can still allow a person to be identified. Remove unnecessary columns, round values, replace specific events with neutral labels, and check if data remains in the filename, comments, or metadata.

Personal ChatGPT and workspace are not the same

In personal OpenAI services, content may be used to improve models depending on user settings. In the official section Data controls you can turn it off Improve the model for everyone; after which new conversations are not used for training. This does not delete old chats and does not make what an employer prohibits permissible.

For ChatGPT Business, Enterprise, Edu, Healthcare, and API, OpenAI states that workspace data is not used for model training by default. However, the organization's administrator can manage access, retention periods, and auditing. The service plan is only one limit: contracts, company policy, law, and data owner permission also apply.

What to check in settings

1. Account type. Make sure you are in the correct personal or managed workspace, and not accidentally in a private profile.
2. Data controls. In a personal account, open Settings → Data controls and consciously choose whether to allow the use of new conversations to improve models.
3. Memory and Personalization. Memory and training are different settings. If context is not needed, turn off personalization or start a non-personalized temporary chat.
4. Temporary Chat. Such a conversation does not appear in history, does not create memory, and is not used to improve models as long as it remains temporary. OpenAI may keep a copy for up to 30 days for security.
5. GPTs and actions. If a custom GPT passes data to a third-party service via an action, that service's policy applies; the retention period may be longer.

Temporary chat reduces the footprint, but does not turn prohibited data into permitted data. For the red zone, the correct decision is not "another mode," but not transmitting data without an approved environment.

Secure workflow in seven steps

1. Formulate the result without data. Write what you need to obtain: a plan, a criteria table, a draft response, or a list of risks.
2. Check permission. Do company policies, contracts, and the employee's role allow using AI for this task?
3. Minimize context. Provide only the fragment needed for the answer. Remove names, contacts, contract numbers, secrets, and unnecessary attachments.
4. Choose an environment. Use an approved workspace; for non-secret one-off tasks, use appropriate personal account settings or Temporary Chat.
5. Set the boundaries. Specify the audience, format, a prohibition on inventing facts, and a rule to flag uncertainty.
6. Verify independently. Cross-check names, dates, numbers, quotes, and references with the original or primary source. For calculations, repeat them in a table or code.
7. Approve and remove the trace. The responsible person makes the final decision. Do not copy secrets back into the chat, do not save the result in a public folder, and delete unnecessary files according to organizational policy.

Three ready-to-use prompts without extra data

First, replace square brackets with neutral descriptions. Do not insert real names, addresses, or contract numbers unless they are necessary for the task.

How to verify an answer before use

In its generative AI risk profile, NIST recommends documenting fact-checking methods and verifying the accuracy and truthfulness of generated information. In practice, verification depends on the type of output.

Asking the model to "check itself" is useful for finding weak spots, but it is not an independent verification. The same model may repeat the same error in different words. For the full process with a statement log and primary sources, see the separate guide how to check AI answers and find hallucinated facts.

Common mistakes

• Copying the entire document "for context". First, try an excerpt or a schema without real data.
• Confusing disabling training with deletion. History, memory, storage, and use for model improvement are different things.
• Using a personal account for internal work. This may violate organization policy even with training disabled.
• Believe a confident tone. Response style is not proof of accuracy.
• Uploading a secret to mask it. Cleanse data before sending.
• Automatically send the result. An email, publication, decision, or system change must have a human checkpoint.

Checklist before pressing Enter

• I know exactly what result I want to get and how I will verify it.
• Company rules allow this task and the chosen tool.
• The request contains no passwords, keys, payment, or unnecessary personal data.
• I provide a minimal fragment, not the whole file "just in case."
• Anonymization truly makes simple re-identification impossible.
• Verified account type, Data controls, Memory, and chat mode.
• Facts, dates, numbers, quotes, and links will be verified independently.
• An external action requires approval from a responsible person.

Quick answers

Can you upload work documents to ChatGPT?

Yes — only when allowed by organization rules, workspace type, and document purpose. Remove unnecessary personal and confidential data before uploading.

Does Temporary Chat make a conversation completely anonymous?

No. It does not enter the history and is not used to improve models as long as it remains temporary, but a copy may be kept for up to 30 days for security.

Is it enough to turn off 'Improve the model for everyone'?

No. This is an important personal account setting, but it does not delete history, override employer policy, or guarantee response accuracy.

Can links and numbers in the response be trusted?

Only after verification. Open the primary source, check the context, and repeat the calculation independently.

Read about how we select sources and correct inaccuracies on the "About the project" page. If you want to discuss a secure workflow without publishing private data, join the Ukrainian community.