Our responsible AI article covers how we build AI features: advisory, failing closed, showing the evidence. This one is about a smaller question that comes up the moment such a feature meets the public. Do you tell people, and how? The answer is yes, in plain language, and the rest of this article is about what that one sentence has to carry.
Why plain language matters
People mind discovering that the letter they thought a nurse wrote was drafted by software and signed without a second look, or that their request sat at the bottom of the queue because of a model's guess. The test in our responsible AI assessment is blunt: would people who believe a human is doing this work feel deceived to learn otherwise? If yes, silence is a design choice, and the wrong one. A second test is more personal: would you hesitate to explain this use of AI on your own homepage? If you would, the problem sits upstream of the wording.
Plain language means the sentence says what the software does, with a verb a person outside your organisation understands: drafts, sorts, flags, suggests. It also says what a person still does.
Errors have an owner
An AI-assisted step will be wrong sometimes. Our own general terms put it plainly: that output is an aid, not a decision, it can contain errors, and you check it before you build on it. Disclosure without ownership is a disclaimer, and a disclaimer protects the organisation while the person affected is left with the error. So before you publish the sentence, answer three questions in writing.
- Who catches the mistakes? A named role reviews the output before it reaches the people affected, wherever a mistake would hurt. If review is sampled, know the sample and write it down.
- Who corrects them? Correcting means the record, the letter or the decision changes, and the correction reaches everyone who saw the wrong version.
- Who informs the person affected? If a sorting step delayed someone's request or a draft went out with a wrong figure, that person hears it from a human, with the correction, without having to ask.
The answers belong in the same document as the sentence on the homepage. If any of the three has no name next to it, the feature is not ready.
What the provider stores is a data-processing question
Most AI features send text to a model that runs somewhere else. That text may contain a name, a diagnosis, a complaint. Whatever the provider keeps, and for how long, is a data-processing questionUnder the GDPR every use of personal data needs a purpose, a legal basis and a record of who processes it on your behalf. Sending text to a model provider is such a use. with a real answer: a legal basis for the personal data involved, a processing agreement and a decision on where the data sits. Our own processor list states that our data sits within the European Economic Area, and that where a supplier also operates outside it, we select the European region and record that contractually. Whatever you decide, decide it before launch, send the minimum the task needs and strip identifying details when the task allows it.
Three sentences you can adapt
The wording depends on what the software does. Three shapes cover most cases. Replace the role and the subject with your own and keep the verb.
- For drafting: "Software drafts the first version of this letter. A member of our care team reads and edits every letter before it is sent, and signs it."
- For sorting or flagging: "Software sorts incoming requests so urgent ones are seen first. It refuses and approves nothing; a person reads every request, and you can ask us to review the order."
- For a second opinion on data: "Software compares these results with earlier ones and flags values that look unusual. The flag is a prompt for the person reviewing your file; that person decides."
Each of them names what the software does, what a person does and what the reader can do about it. Put the sentence where the people affected read: the page with the form, the letter itself, the footer of the message. A link to a policy page comes second.
What not to write
- A badge. "Powered by AI" or "AI-enhanced" tells the reader nothing about what happened to their data or their request. It is marketing in the place where information should be.
- A blame clause. "The AI may make mistakes" without an owner moves the risk to the reader. The errors are yours; say what you do about them.
- More than you can keep true. A sentence that promises "every output is reviewed" becomes a liability the day review is sampled. Write what actually happens.
Where this sits in the decision
Step two of our responsible AI assessment lists the conditions before AI touches real people or real data. Two of them are this article: errors have an owner, and the people affected know AI is involved, in plain language. The others, sending the minimum data, a person reviewing where mistakes hurt, stopping when uncertain, measured quality and a budget for monitoring, belong on the same launch checklist. Step one comes before all of it and asks whether AI is the right tool at all. Often it is not, and then there is no sentence to write. In care, where the decision belongs to the clinician, the sentence has to say that too.
Keep reading
We build AI features for a living, which is exactly why we'll tell you when you don't need one.
AI that assists is useful. AI that decides alone is a liability. The difference is a few concrete rules we never bend.
Past the hype and the fear, AI is quietly reshaping diagnostics, admin, and how patients follow their own care.
Facing something similar?
This is how we work. If the challenge in this article sounds familiar, tell us about yours and we'll give you a straight, useful answer.
Talk to us