How to stop employees leaking data to ChatGPT

If your company uses computers, some of your data is probably already going into ChatGPT. Not because people are careless, but because it is genuinely useful, and the quickest way to get an answer is to paste in the thing you are working on. The question is not whether it happens. It is how much of what goes in is sensitive, and what you can do about it without taking a useful tool away from people.

The leak is quieter than you think

The security firm Cyberhaven has measured what employees actually paste into ChatGPT across a large set of companies. In their analysis, about 11% of that pasted content was confidential: source code, client records, regulated data. The average company was leaking material like this hundreds of times a week. And it was not spread evenly. Fewer than 1% of employees accounted for roughly 80% of the leaks.

Those numbers change the shape of the problem. This is not a company-wide discipline issue. It is a small number of heavy users, doing normal work, who happen to move a lot of sensitive text through a chatbot. You do not need to police everyone. You need to catch the sensitive data, wherever it comes from.

Why blocking usually backfires

The instinct is to block ChatGPT on the corporate network and call it done. Samsung did a version of this in 2023, after engineers pasted source code into ChatGPT, and banned generative AI on company devices. Blocking has a predictable side effect, though. People who find the tool useful do not stop using it. They move to a personal laptop, a phone, or a home account, where you have no visibility at all. Most employees are already using AI tools their company never approved, and they pick them up faster than IT can vet them. Security teams call this shadow AI.

So a hard block often trades a visible, manageable risk for an invisible one. The goal is not to stop people using AI. It is to stop the sensitive data from leaving with them.

What actually helps

A few things work better together than any one of them on its own.

1. Give people a sanctioned tool. A lot of the risk comes from the free tier, where conversations can be used for training. A paid Team or Enterprise plan does not train on your content by default, so putting one in front of people removes a big reason to reach for a personal account.
2. Write a short, plain policy. Many companies have no written rule on AI at all, so staff genuinely do not know that pasting client data into a chatbot is a problem. One page that says what is fine, what is not, and which tool to use beats a long document nobody reads.
3. Catch it at the point of paste. The moment that matters is the send. A check in the browser can spot a key, a customer record, or a block of source code as someone is about to submit it, and hold it back before it leaves. This layer works regardless of which site the person is on.
4. Coach the few, rather than punishing everyone. Since a small group causes most of the exposure, a quiet word with those users does more than a company-wide crackdown. Most of them are not trying to do anything wrong.

Keep the check on the device

One trap when you add controls is routing everything people type through a monitoring service so it can be scanned. That fixes a leak by creating a second copy of the same sensitive data somewhere else. It is worth doing the check on the device instead, so the text is inspected where it is typed and only the decision leaves, not the content.

Where Riidact fits

Riidact is built for that last layer. It sits in the browser, watches the boxes people type into on ChatGPT, Claude, Gemini, and any other site with a text field, and checks each message for keys, credentials, personal and tax IDs, confidential markers, and rules you define for your own business. When it finds something, it flags it before the message is sent. The checking happens on the device, so the sensitive text never passes through us. People keep using the tools they like, and the data that should not leave, does not.

You are not going to talk your way out of shadow AI, and blocking your way out tends to make it worse. Give people a safe tool, tell them plainly how to use it, and put a check at the point where data actually leaves. That version holds up, because it works with how people already behave instead of against it.