Turning off training doesn't mean your AI tool keeps nothing.

Turning off training in an AI tool doesn’t mean it keeps nothing.

When Claude Code reads a file, what it reads goes to Anthropic as part of the conversation. Often that’s the whole file. If it only searches, it’s the matching lines. The docs say it sends all prompts and model outputs over the network. The same is true of anything you paste or upload into any cloud AI tool: a customer list, a contract, a board deck, a guest database. A copy now sits on someone else’s servers. How long it stays there depends on the account you use and a couple of settings most people never look at.

Claude is my tool of choice, so I’ll use that as the example. On the personal plans (Free, Pro and Max), if you’ve opted out of training, data is kept for 30 days, and for up to five years if you haven’t. On the business plans and the API, Claude Code data is kept for 30 days as standard. Zero data retention (ZDR) exists, but only on Enterprise and only once Anthropic’s account team switches it on. And a feedback or bug report sent from Claude Code includes the conversation, with any code and file contents Claude read, and can cover a week of sessions on the same project. It’s kept for five years whatever your other settings.

This isn’t an argument that any vendor is insecure. The data is encrypted. The point is that the copy exists, outside your control, for a period set by your account type and a few clicks. Most NDAs and client contracts were written before anyone imagined that.

Personal data adds a second problem. The business plans and the API come with a data processing agreement, which makes the AI provider a processor under GDPR. The personal plans don’t. With Claude Code it depends on how you sign in: a personal login means consumer terms, an API key means commercial ones. A manager who pastes customer details into their own account to tidy up a spreadsheet has handed personal data to a third party with no processor agreement in place.

That isn’t a technicality. Under UK GDPR, the most serious breaches carry fines of up to £17.5 million, or 4% of global turnover for the largest companies. A missing processor agreement on its own sits in a lower tier, up to £8.7 million. Those are ceilings, not typical fines. But it’s not a conversation you want with the ICO, or with the customers whose data it was.

For Claude Code you can also check what actually left the machine. Each session is saved as a transcript locally, including every file Claude read. Those transcripts are deleted after 30 days by default, so look soon after the work is done, or extend the period before you start.

So before anyone points an AI tool at your code or your data, whether that’s an auditor, a contractor or your own team, ask four questions. Which account type is it running on? What’s the retention period? Is feedback switched off? Does that fit your client contracts and your data protection obligations?

None of this is a reason to stop using AI. It’s a reason to treat it like any other supplier that holds your data.