The analogy
Entrusting sensitive data to a consumer AI is like handing confidential documents to a public counter without knowing where the copies end up. Maybe they stay in a secure drawer. Maybe they're used to train the counter's staff. Maybe, one day, someone finds them again. The point is that you don't decide it: once you've handed over the sheet, you've lost control of the copy.
Sensitive data works like this: the damage isn't the use you make of it, it's the loss of control. That's why the defense isn't "trusting the right counter," but deciding upstream what leaves your hands and what doesn't. What you don't hand over can't be lost.
How it really works
What makes a piece of data "sensitive"
In practical terms, it's any data that identifies a person or that, if exposed, harms them: name and contacts, health data, financial data, orientations, information about minors, but also company secrets and data entrusted by customers. For many of these, data-protection laws impose precise obligations on whoever handles them. Inputting them into a tool whose handling you don't control may violate those obligations regardless of your intentions.
Where the risk arises: retention and reuse
When you paste a text into a consumer tool, two things can happen. The conversation is retained on the service's servers. And, on several services, the inputted content is used to improve the model, unless you've explicitly turned off this option. It means that a customer's data could enter the training material. You don't "find it again" published, but you've lost control over where it sits and how it's used: and that's exactly what privacy rules forbid.
The risk multiplies with third-party data
Exposing a piece of your own personal data is your choice. Exposing the data of a customer, a patient, an employee, is a violation toward them, and it triggers liability and notification obligations. Here good faith isn't enough: what counts is what got out and whose it was.
Regulated sectors: the risk is double
Healthcare, legal, finance and public administration have internal and sector rules that often forbid or limit the use of external tools precisely for this reason. In these fields, using a consumer AI on work data can violate two levels at once: the privacy law and the sector policy.
What you can do in practice
- Turn off the reuse of data. In the settings look for the item on data control or on use for training and turn it off. Use temporary chats, which don't stay in the history, for delicate content.
- Anonymize before pasting. Replace identifying data with placeholders. The operational syntax:
In the text that follows I've replaced personal data with placeholders such as
[NAME], [CODE], [DIAGNOSIS]. Work only on the placeholders, don't ask me for the real values
and don't try to guess them.
- For work, use a version with guarantees. Business or enterprise versions often declare, contractually, that they don't use your data for training and that they meet protection standards. For third-party data it's the minimum condition.
- Verify the legal basis. If you handle other people's data, you must have a basis to do so (consent, contract, obligation). Inputting them into an external tool is a new processing: it must fall within that basis.
- Keep track of what you input. A log of which data passes through which tools is useful for answering in case of an audit and for noticing an error quickly.
When it does NOT work (and how to fix it)
If you've already pasted a third party's sensitive data
Delete the conversation, verify that the reuse was turned off, and assess the notification obligations of your organization and of the law: for certain violations, reporting to the authority and to the data subject is due within tight deadlines. Acting quickly reduces the consequences; hiding worsens them.
If you don't know whether the tool reuses your data
Treat the doubt as a yes: don't input sensitive data until you've verified the service's terms and turned off the reuse. The absence of information isn't a guarantee.
If your organization doesn't have a policy
Propose a minimum rule before you find yourself in the situation: no customer or personal data in consumer tools, a business version for work on third-party data, anonymization by default. A written policy protects you and the organization.
A tip from someone who really uses it
Keep two accounts or two habits separate: a "free" tool for brainstorming on your own, non-sensitive content, and a controlled path (business version, anonymized data) for everything that touches real people. The physical separation of habits avoids the most common error, which is pasting in the wrong place out of haste.
Frequently asked questions
Can I use AI to summarize my customers' emails?
Only with caution: those emails contain third-party data. Use a version with contractual guarantees and, where you can, anonymize the names. With a consumer tool with reuse active, you're exposing other people's data, and it's the riskiest case.
If I delete the chat, is the data safe?
Deleting removes the conversation from your history, but it doesn't guarantee the data wasn't already retained or used elsewhere if the reuse was active. Deletion is a remedy after the error, not a protection: the protection is not inputting the data.
Aren't all AIs by now compliant with privacy rules?
No, and it's the costliest misunderstanding. The provider's compliance and your compliant use are different things: even a compliant tool becomes a violation if you input third-party data into it without a basis or with the reuse active. The responsibility for how you use the tool stays yours, it doesn't transfer to the certified provider.