The AI Setting That’s On By Default, And Quietly Training On Your Customers

By Konstantinos Karamatzianis · Co-Founder, Dabster Labs

Published 9 July 2026

Here’s a setting most businesses running an AI tool have never looked at, because it’s buried two menus deep and switched on by the vendor: use my data to improve the model. On a lot of popular AI platforms, on the plans small businesses actually buy, that toggle is on by default, which means the conversations your customers have with your bot may be quietly feeding someone else’s AI.

We found it on our own stack while tightening up a client’s assistant, and it’s the kind of thing that turns a compliant setup into a problem without anyone deciding it should.

What the toggle actually does

When it’s on, the content flowing through the service, the questions, the transcripts, sometimes the recordings, can be retained and used to train or improve the vendor’s models. For a business that means customer personal data leaving your control for a purpose your customer never agreed to, and one you probably never mentioned in your privacy policy. Under GDPR that’s a purpose-limitation problem: you collected the data to answer an enquiry, not to train a US company’s AI.

Why the default catches people out

Enterprise tiers usually ship with training switched off. The cheaper tiers, the ones a small business signs up for, often ship with it on, and leave it to you to find the opt-out. Worse, the opt-out is frequently forward-looking only: it stops future data being used, but anything already submitted may already be in the training set. So every week you leave it on is a week of customer conversations you can’t pull back.

The fix takes two minutes

You don’t need to upgrade your plan or hire anyone. On most platforms it’s a single switch in the account’s data or privacy settings, turn off “use my data for training,” and while you’re there, check the data-retention period and make sure there’s a data-processing agreement on file (that part is usually a one-click accept, not a negotiation). Do it once, write down that you did it, and the exposure stops.

The wider point

Compliance rarely fails at the dramatic, obvious step. It fails at a default nobody changed. When we set up an AI assistant for a client, the boring switches, training off, retention capped, recordings deleted on a schedule, a real record of what’s processed and why, get done on day one, because they’re the ones that quietly go wrong later. The clever part of the assistant is easy to see. The compliant part is the part you have to go looking for.

The takeaway

If you’re running any AI tool that touches customer data, go and look at its data-use settings today. There’s a good chance something is on that shouldn’t be. It’s a two-minute fix that most businesses never make, until someone asks them to prove they didn’t.

Not sure what your AI setup is exposing? Get in touch and we’ll take a look.