Which account are you using, and what it changes

The same window, three different legal situations

Open a chatbot at home and open the same brand at work and you may see an identical box. What sits behind it can be three quite different things, and the difference is not a feature list. It is who has promised what to whom.

The consumer tier. You agreed to the provider's terms personally. On most consumer products, conversations are retained for a period and human reviewers may read some of them for safety and quality work. Several providers train on consumer conversations unless you turn it off. Your employer has no contract here, no visibility, and no ability to retrieve or delete anything. If a regulator asks where a client's data went, the honest answer is "to a company we have no agreement with, under terms an employee accepted on a phone".

The business or team tier. A contract exists between your organisation and the provider. Training on your content is normally off by default and stated in the terms. There is an administrator, retention settings, and usually an audit trail.

The enterprise tier or your own cloud tenant. The strongest position: contractual commitments about training, region, retention and deletion; single sign-on so leavers lose access; logging; and in some deployments the model runs inside your organisation's own cloud account so the content never reaches the provider as a customer conversation at all.

The models available on these tiers are broadly the same. What differs is the paperwork, and the paperwork is the thing your professional duty actually turns on.

The settings worth reading once

Whatever tier you are on, go and look. It takes ten minutes and most people have never done it.

• Training or "improve the model for everyone". Find it. Read whether it is on. On consumer tiers it is frequently on by default, and turning it off is one click.
• Chat history and retention. Some products keep conversations indefinitely. Some offer temporary chats that are not retained. Know which you are in.
• Connected apps. Any connector to your mail, drive or calendar widens what the tool can read enormously. Grant these deliberately, not while dismissing a dialog.
• Who else can see it. In a shared workspace, "shared" links and team projects may be visible more widely than you think.

Two things these settings do not do, and it is important not to over-read them. Turning training off does not mean nothing is stored — retention for abuse monitoring usually continues, often for a stated number of days. And a setting on your account does nothing about your duty to the person whose information it was, which is the point of the next lesson.

Anonymising is harder than it looks

The standard workaround is to strip names. It helps, and it is not enough on its own, because identity survives in detail. A "42-year-old male sub-postmaster in a town of 8,000 with a 2019 conviction" is one person. A case description is often unique even without a name — that is precisely what makes it a case.

Use it as a discipline rather than a licence:

• Replace names with roles, not initials. Initials re-identify.
• Round or band anything that is unusual. "Early forties", not 42.
• Remove the combination, not just the fields. Three ordinary details can be jointly unique.
• Cut anything that is not needed for the question. Most pasted documents contain far more than the question requires, and the surplus is pure risk.
• Then ask yourself the test from the previous module: if this exact text appeared in a newspaper, would anyone know who it was?

Shadow accounts and why they happen

People do not use unapproved tools out of recklessness. They use them because the approved tool is slower, or because there is no approved tool and there is a deadline. Punishing that behaviour does not remove it; it removes your visibility of it.

If you are the one deciding: provide a good approved tool quickly, because the gap between what people need and what they have is exactly the size of your shadow-AI problem. If you are the one using it: ask for the tool in writing. "Is there an approved AI tool for this?" is a one-line email, it moves the risk to the person whose job it is to carry it, and in every organisation that later has an incident, the people who asked in writing are in a different position from the people who did not.

The one sentence to remember

The question is never "is this model any good". It is "whose contract governs this text once I press enter".