Protect a file before sharing

This is general information, not legal advice. Your employer's policy, your contracts, and the law that applies to you decide what you may share. When they conflict with this article, they win. See our terms of service.

What data should not go into an external AI tool

A practical guide to the kinds of data that need a decision before they are pasted or uploaded to an AI assistant, what happens to what you submit, and what to do instead.

Published · Updated

An AI assistant is useful because you can give it real material: a table, a contract, a block of code. The same convenience makes it easy to hand over something you should not. The question is not whether the assistant is trustworthy. The question is what you are entitled to give to a third party, and what that party is allowed to do with it.

Key terms

External AI tool
an AI service run by another organization, reached through a website, app, or programming interface, where your input is processed on their systems.
Personal data
any information about an identified or identifiable person, such as a name, an email address, an employee number, or a combination of facts that points to one person.
Special category data
personal data that the EU's General Data Protection Regulation (GDPR) treats with extra care. It covers racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic or biometric data used to identify a person, health, and sex life or sexual orientation. The rules for it are stricter.
Retention
how long the provider keeps what you submitted, including after you delete a conversation.
Training
using submitted content to improve the provider's models. Providers often treat consumer and business products differently.

What happens to what you paste?

It depends on the product, the plan, and the settings, and all three change. As a rule of thumb, read four things in the provider's current terms for the exact product you use. The four are whether your content is used for training, how long it is kept, whether people can read it (for safety review or support), and where it is processed.

"No training" does not mean "not kept"

Two providers' own pages, read on 28 October 2026, show the pattern. OpenAI's data controls page says that content in personal ChatGPT accounts can be used to improve its models unless the setting "Improve the model for everyone" is switched off. It says that business, enterprise, and education workspaces do not have their content used for training by default, and that "temporary" chats may be kept for up to 30 days for safety purposes. Anthropic's announcement of 28 August 2025 says that for its Free, Pro, and Max plans the person chooses whether chats are used for training. It keeps them for five years if the person agrees and for the existing 30 days if not. Its commercial plans and its programming interface are not affected. But "not used for training" is a narrower claim than "not kept." Content can still be retained for a period, and in some cases it can be viewed by administrators of your own workspace or reviewed for safety.

Turning a training setting off may affect only future chats: Anthropic says its change applies to new or resumed chats. OpenAI warns that a thumbs-up or thumbs-down on a reply may allow the whole conversation to be used even after you opt out. Treat these as examples of what to look for, and check the provider's own pages, because terms change without much notice.

Which kinds of data need a decision first?

Kind of data Why it needs care What to do instead
Passwords, keys, access codes, connection strings Anyone who sees them can act as you. Deleting a conversation does not revoke them Never paste them. If one was pasted, treat it as exposed and replace it
Personal data of customers, staff, or the public You may have no legal basis or permission to disclose it to this recipient Remove or replace the personal values first, or use an approved service under an agreement
Special category data (health, beliefs, union membership, biometric, and similar) Stricter rules apply and harm from a leak is greater Do not use an external tool without a specific legal assessment
Payment card numbers The card industry's security standard restricts where card data may be stored and processed Mask or remove the numbers; ask for help with the pattern, not the data
Other people's confidential information (client files, contracts under non-disclosure) Your duty of confidence is to them, not to you Get permission, or anonymize and generalize
Non-public company results and plans Disclosure can breach policy and, for listed companies, securities rules Use dummy figures with the same shape
Source code and trade secrets You may lose the ability to keep them secret Share a minimal example without secrets, or use an approved enterprise product

Has this gone wrong for anyone?

Yes, in two different ways.

Staff sharing company data. In May 2023 Bloomberg reported that Samsung Electronics had found that employees uploaded sensitive code to ChatGPT. Samsung told staff in one of its largest divisions not to use ChatGPT and other generative AI tools, saying that data sent to such services is stored on servers run by the providers and is hard to retrieve or delete.

Regulators scrutinizing the providers. In December 2024 Italy's data protection authority, the Garante, announced a €15 million fine on OpenAI, finding among other things that it had trained its model on personal data without an adequate legal basis and had breached transparency duties. OpenAI said it would appeal, and the fine was suspended in March 2025. On 18 March 2026 the Court of Rome annulled it. As reported, the court decided on jurisdiction, not on the merits. OpenAI had set up an Irish company, which made the Irish authority the lead regulator, so the Italian one could not impose the fine alone. The court did not rule on whether OpenAI had broken the rules. Whether the Garante will appeal to Italy's highest court was not known on 20 October 2026. The case concerns the provider's own obligations. It shows two things: regulators do examine how AI services handle personal data, and which authority is responsible can be as contested as the rules themselves.

How do you decide, in a few seconds?

What are the better options?

  • Ask for the method, not the analysis. Describe the columns and ask for a formula, a query, or a script. Run it yourself on the real data.
  • Use a small, invented sample with the same structure.
  • Protect the real file first. Replace names with stand-ins, round amounts, shift dates, remove free-text columns, and look at the result before you share it (see how to remove personal details from a spreadsheet).
  • Use an approved service. An enterprise agreement can give contractual limits on retention and training and can name where data is processed. Your organization may already have one.
  • Use a model that runs on your own machine for material that cannot leave it. Quality and speed differ.

How does DataAttendant fit in?

DataAttendant prepares a CSV or Excel file for this situation. In your browser, it finds sensitive values, lets you decide what happens to each kind, previews the result, and lists what remains in the protected copy. It cannot decide whether you may share something, and it cannot find every sensitive value. The decision about the recipient stays with you. How DataAttendant works goes through the steps.

Try DataAttendant

Sources

  • Regulation (EU) 2016/679 (GDPR): Article 4(1) (personal data), Article 9 (special categories of personal data).
  • Bloomberg, "Samsung Bans Staff's AI Use After Spotting ChatGPT Data Leak," 2 May 2023; Forbes Australia, "Samsung bans ChatGPT among employees after sensitive code leak." Read 19 October 2026.
  • Tribunale di Roma, judgment no. 4785 of 18 March 2026, as reported by diritto.it; PPC Land, "How Europe's AI privacy rules are being gamed" (court's full reasoning published 28 May 2026). Read 20 October 2026.
  • Euronews, "Italy's privacy watchdog fines OpenAI 15 million after probe into ChatGPT data collection," 20 December 2024. Read 19 October 2026.
  • Reports of the annulment by the Court of Rome (18 March 2026; jurisdiction, not the merits): Reuters via Freevacy, 20 March 2026; gdprfine.com case summary; dirittodiinternet.it, June 2026. These are secondary sources; the judgment itself has not been read. Read 19 October 2026.
  • OpenAI Help Center, "Data controls FAQ," help.openai.com/en/articles/7730893. Read 28 October 2026.
  • Anthropic, "Updates to Consumer Terms and Privacy Policy," 28 August 2025, anthropic.com/news/updates-to-our-consumer-terms. Read 28 October 2026.

Where next?