Your data and AI: what providers do with your inputs

01Where your inputs go

An AI chat feels like a program on your computer but is a service: every input is sent to the provider’s servers and processed there. Depending on your account, the provider may additionally store these inputs and use them as training material for future models. Once something has landed in training, you will practically never get it back. No reason to panic, but a good reason to look closely.

02Consumer account or business plan

The big providers now draw a clear line between personal accounts and business customers. The basic rule: whoever shows up as a business and pays gets the better commitments. As of August 2026 the picture looks like this at most providers:

• Free and personal accounts: training on your inputs is often active by default, and can be switched off in the privacy settings.
• Business and enterprise plans: training is off by default, with contractual commitments and a data processing agreement on top.
• API access: no training on your data in the default setting either.
• Careful with feedback buttons: thumbs up or down can release the conversation for review.
• Settings change again and again. Check the current state directly in your provider’s privacy documentation.

03What never belongs in a public chat

In a public chat on a personal account you have to assume that your inputs can be stored and analysed. Four categories have no business being there:

• Customer data: names, addresses, order histories, mail threads with real senders.
• Credentials: passwords, API keys, payment data, in any context.
• Trade secrets: calculations, supplier terms, unpublished plans.
• Health data and other specially protected information, including about yourself.

04The GDPR still applies

As soon as personal data flows into an AI tool, you need a legal basis, just like with any other service provider. If the provider processes the data on your behalf, a data processing agreement belongs in place. And because many providers sit in the US, third-country transfer comes into play: it needs its own basis, such as a certification of the provider under the current EU-US framework or standard contractual clauses. The situation keeps changing here, so check it when you sign and not just once.

05How to protect yourself in practice

The most effective step is switching to a business plan with a data processing agreement. The second is a simple habit: anonymise before you paste. Replace names with placeholders, strip addresses, shorten identifiers. For most tasks the AI does not need the real data at all, a "customer A" works just as well. Both together, the right plan and frugal inputs, solve most of the problem.

My rule of thumb: never type anything into an AI chat that you would not send to an external contractor without a contract.