Telling users it is AI, and what you send about them

Say it where it appears

An AI-generated summary carries a label (“Generated by AI”) beside it, not in a settings page or a terms document. A chat assistant introduces itself as one and, asked directly, says so every time. A suggested reply is shown as a suggestion, not inserted as though the user wrote it.

This is partly a matter of law, which the end of this lesson comes back to, and mostly a matter of not lying to the people using your app. A user who believes a person wrote the reply will act on it differently, trust it differently, and feel differently when they learn otherwise. The label costs nothing and the absence of one costs the thing you cannot buy back.

Do not give the assistant a human name and face and a script for dodging the question. An illustrated portrait rather than a photograph is the right choice if there is a portrait at all (a photorealistic face asserts that a person exists), and the name, if any, is presented as the name of a tool. Addaly’s own tutors are drawn, labelled AI on every surface, and answer the question honestly, for exactly these reasons.

The loading state is part of honesty

Between the click and the answer there are seconds, and what fills them sets expectations. Streaming text reads as “being written”. A spinner with no words reads as “broken” after three seconds. A skeleton that suggests the shape of the answer, then fills, reads as “working”. Choose deliberately, and pair it with the sentence from the resilience lesson for when nothing comes.

Beside the output, one short line where it will be read: “May contain mistakes. Check dates and amounts.” Put it near the specific thing the model gets wrong for your feature, not as a general disclaimer at the bottom of the page where nobody looks.

What you are sending, and to whom

Every call sends the user’s text to a company that is not you. The provider’s terms say what they do with it: whether it is retained, for how long, whether it is used to train, whether it is reviewed by people, and in which country it is stored. These terms differ between providers and between the free and paid tiers of the same provider (free tiers more often retain and train), and they change. Read the current page for yours, and write down the date you read it.

Send only what the feature needs. A summary of one note does not need the user’s name, email or the rest of their notes. A categoriser does not need the receipt image if the amount and merchant are already text. Every field removed from the prompt is a field that cannot leak, cannot be retained, and costs nothing.

Local models change the answer entirely: text that goes to a model on your own server goes nowhere else. For a feature that touches health, finances or anything a user would not say aloud, that is worth the quality gap.

Your own logs are the same problem

The log line from the prompt lesson records inputs and outputs. That is a record of what users typed, held by you, searchable by anyone with access to the logs, for as long as the retention setting says. It is personal data. Decide how long to keep it, redact obvious identifiers before it is written, and put it behind the same access controls as the database. The evaluation set you build from real inputs is a copy with the controls removed; the same care applies.

The law

Disclosure rules exist and differ. The European Union’s AI Act includes transparency duties for systems that interact with people; some US states have laws on bots that must identify themselves in certain conversations; India has issued advisories on labelling AI-generated content; other countries have nothing yet, or something in draft. Data rules (where user text may be sent, whether consent is needed, what must be deletable) differ again: GDPR in Europe, the DPDP Act in India, a patchwork in the US.

This course cannot tell you what applies to your app, because it depends on where you are, where your users are, and what the feature does, and because these rules are being written now. What it can say is the shape: disclosure is expected almost everywhere and required in a growing number of places; sending personal data to a third party in another country is a regulated act in many; and the cheapest compliance is the honest design (label it, send less, keep less) which also happens to be the right thing to build.

When the app handles anything sensitive, or has users in more than one country, that is the point to ask somebody qualified in the relevant jurisdiction. Not a model, and not this lesson.