• Regulation & Compliance

Every day millions of people type something into an AI tool and press enter, and there is a chance that is exactly what the employees at your company are doing right now. A question, a paragraph to tidy up, a spreadsheet to summarise. It feels like thinking out loud. It is not. It is a transmission of data, and sometimes of your intellectual property. The moment they press enter, their words leave the building and travel to a computer owned by someone else, usually in another country, where they are processed, possibly logged, and governed by rules nobody in your organisation has read. So here is a question worth a minute of any leader’s time. When your staff type a prompt, where does it go, who can see it, who can demand it, and who can use it as training data? And who owns it afterwards, your company or the AI company?

The journey is simple enough. The prompt travels encrypted to the provider’s servers, where it is decrypted so the model can read it and produce an answer. What happens next varies. Depending on the provider and the settings, that prompt may be deleted in seconds, kept for thirty days, held for years in logs, or used to train the next version of the model. These are the default settings, and they differ between the free tools your staff signed up for on their own and the business versions nobody got round to buying. And if the prompt contained a customer’s details, it is still personal data and you are still responsible for it, wherever it went.

So, who can see it? First, the provider, whose staff can under certain conditions review conversations, usually for abuse monitoring. Second, the chain of sub processors, cloud hosts and support services behind every AI tool, each listed in a document nobody reads. Third, and this is the one that surprises people, governments.

This is where most organisations have an out of date mental model. The American CLOUD Act of 2018 lets a US court order any company under US jurisdiction to hand over data it holds, wherever in the world that data is stored. The obligation follows the company, not the data centre. Choosing a European server region from an American provider changes where the data sits. It does not change who can demand it.

What fewer people know is that the picture changed again in October 2022, when the UK US Data Access Agreement came into force. It was the first of its kind. It lets American law enforcement serve orders directly on British companies in serious crime investigations, and British police do the same to American ones, with no court or minister in the receiving country reviewing the request first. Each side agreed not to target the other’s citizens, but the check on that is done by the requesting country. The practical effect is that switching to a British provider no longer puts your data out of reach. It changes which route an order arrives by, not whether it arrives.

GDPR does not help much here either. It was written to govern how organisations use personal data, not to shield it from governments, and it allows disclosures that the law requires. When a lawful order arrives through an agreement like this one, the order usually wins.

Then there is ownership. Most providers say you own what you type in and what comes out. Read the terms and you have usually granted them a licence to use it to run and sometimes improve the service. Ownership is not control. You can own a prompt that is sitting in a log on another continent, and control is the question the marketing pages answer least clearly.

None of this means avoid AI. Used well it is a real advantage, and the organisations that ban it simply push it onto personal phones where they have no visibility at all. Your people are going to use AI. The only choice you have is whether you see it.

What it does mean is asking three plain questions of any provider before you sign. Who actually holds the data, which country’s courts can compel it, and will you be told if that happens? Then sort your data into two piles, ordinary and sensitive. The ordinary can go to the best tool available. The sensitive, customer records, contracts, anything a regulator cares about, you may choose to keep in a tool you control. Most organisations end up running both, and that is not a compromise. It is the right answer.

Most organisations could not say today where their prompts went or who could demand them. That is the gap to close.

---

Read More Regulation

Comments are closed.