UX for AI features

AI features have a UX problem that ordinary software does not: they can be wrong, sometimes convincingly. The interface has to be designed for that.

Five principles:

1. Show the uncertainty where it exists — but not as a percentage the user cannot interpret. «I'm unsure about this — check it against the source» says more than «confidence 0.62».
2. Make correction cheap. Editable suggestions, not finished decisions. Undo should always be there.
3. Show the sources when the answer is built on retrieved material.
4. Set expectations early — what the feature is good and bad at, in the interface and not in the terms.
5. Let the user get on when the AI fails — a route to human help or manual input.

Automation bias is the underrated risk: people systematically trust automated suggestions too much, especially under time pressure.

Patterns that work:

Pattern	Why
A suggestion, not a decision	the user keeps the control and the responsibility
Streaming answers	perceived speed; it can be interrupted
Inline sources	checking is one click away
An «I don't know» mode	better than an invented answer — but it has to be rare enough to be taken seriously
A feedback button	gives data for evaluation, if it is actually read
Clear AI labelling	a requirement in the EU AI Act for synthetic content

Anti-patterns:

• Confidence figures with no explanation («87 % sure» of what?).
• Explanations that are post-hoc constructions — they increase trust without increasing reliability.
• Hiding that AI is being used.
• Making correcting harder than accepting.

For children and young people the requirements are stricter: it must be clear that it is a machine, no dark patterns encouraging longer use, and a visible route to an adult. That is both a design question and a requirement in the AI Act and data protection law.

Review an AI feature you use with six questions:

1. Is it clear that it is AI?
2. What happens when it is wrong — how easy is it to notice and correct?
3. Is the uncertainty shown in a way I can act on?
4. Can you see where the information comes from?
5. Is there a route to human help?
6. Am I encouraged to check, or to trust?

Apply it to AI-grafen's tutor:

• Labelling: yes, in the footer on every page.
• Errors: the pupil sees the node's source text alongside and can compare.
• Uncertainty: the tutor is instructed to say when it does not know.
• Sources: the node's sources are listed with their licences.
• Human help: the teacher sees the class's gaps; sharing with a mentor exists.
• Checking: the explanation encourages you to try it yourself in the exercises.

Do the same walkthrough for a feature you are building — the shortcomings become visible quickly.