Conventions: Don't Let AI Invent New Interactions

Chapter 13 of About Face 4 splits interfaces into three approaches: grown from the program's internals (implementation-centric), drawn to look like real objects (metaphor), and learned (conventions / idioms). Cooper's stance is blunt: modern GUIs mostly run on the third. Windows, scrollbars, dropdown menus, hyperlinks — none of them are intuitive; users learned them one by one.

His example is the mouse: nobody glances at a mouse and knows it drives the on-screen cursor — someone has to teach you, or you fiddle for half a minute. Once learned, you never forget. Scroll, double-click, drag-and-drop, pull-to-refresh, swipe-left-to-delete — same story: learn once, use everywhere.

For product people that's a free gift: the world's products already taught your users — reuse existing conventions and learning cost is zero. The snag is AI often ignores that favor. It stitches patterns from training data and invents a control nobody has ever learned.

While we're here, bury the second approach too. In the early GUI boom, credit went to "looking like the real thing": desktops drawn as desks, folders as folders, users reading pictures. Cooper doesn't buy it in Chapter 13: metaphor interfaces make users play connect-the-dots between icon and function. Simple features maybe; anything abstract can't be drawn. "Batch-rename 50 files" — what would you draw?

Worse, metaphor overstays. Once users know the basics, leather textures and page-flip animations stop helping and start dragging — they eat space, constrain interaction, and drag mechanical-era rules along. Apple admitted it: the leather calendar in iOS 6 was wiped clean by iOS 7.

So conventions win by traveling light: they don't ask you to guess — they ask you to learn, and learning is fast. Next, how this vocabulary is built.

Chapter 13 has an inverted pyramid explaining why conventions are easy to learn: the whole GUI interaction vocabulary sits in three layers — fewer and simpler as you go down. Tap each layer to see what it holds.

Users can carry this vocabulary from other products into yours via mental models: they assume your UI obeys the same rules as the ten thousand they've used. How lost they get when models mismatch was covered in the Psychology part — here's a jump card, no re-teaching.

Below is an AI-generated cloud-drive settings page. Features work, the UI isn't ugly — but it hides three clever inventions, each breaking a well-learned convention. Tap them one by one; hits get a verdict.

Why does AI do this? Its training data mixes everything: real products, award-winning concepts, portfolio showpieces from design communities. Ask for "a settings page with design sense" and it blends patterns from several sources — grammatically fine, rule-wise fighting each other.

One thing it can never learn: what your users already know. Conventions live in user experience. AI sees how a control looks, not thirty years of teaching history. So this gate is yours: any control that needs a tiny hint line to explain how to use it deserves suspicion.

The second word comes from Norman's The Design of Everyday Things: affordance — how a thing looks tells you how to use it. Cooper sharpens it in Chapter 13: affordance is about what you think it can do, counted separately from what it actually does. A small round button by a door reads as a doorbell to everyone — even if pressing it does something else, its affordance is still doorbell.

Classic case: the door handle. A horizontal metal bar shouts "pull me" in shape and height — but the door pushes. The shop sticks a "Push" label on. Useless; people still pull. Signals the hand can read speak louder than words.

On screen it becomes a contract: draw a button and you promise it can be clicked, reacts when clicked, and does what the label says. Controls that break the deal Cooper calls bait-and-switch affordance.

He also calls out flat design: early UIs used shadow and depth to shout "press me"; flat style scraped that away. Visually clean — affordance took the hit. So clickable elements in flat UIs need stricter discipline: one color, one shape, reserved for things that click. No guest appearances.

Two membership centers below — same upgrade card, two drawings. Tap the one you think keeps the contract.

Affordance has a zero-cost test: send a screenshot to someone who's never used it, give 5 seconds, ask them to point at what's clickable. Every miss is an affordance accident. Especially useful on AI-generated UIs — AI never asked that question while drawing.

Easy wrong conclusion here: always copy existing interactions — innovate and die. Cooper's ruler is finer:

Navigation, save, forms — users rehearse these hundreds of times a day in other products. Changing the pattern forces them to relearn walking, with no real payoff — use the most common patterns here. Spend the innovation budget on your core differentiator: the capability unique to your product, where no convention exists to copy. Nail it, and others copy you.

Measure one question with that ruler. AI added four "highlight designs" to a batch file manager. Which innovation spent the budget right?

One last cheat sheet. Users already know these conventions cold from other products — copying them is free usability. Check AI output against this list; the right column pastes straight into a prompt.

Nothing in the table asks users to learn — that's the value. The learning budget you save, by the last section's ruler, goes to your core differentiator.

Source: Original to Xiaoshan Academy's Interaction Engineering series; some interaction principles adapted from About Face 4: The Essentials of Interaction Design; affordance concept from Don Norman's The Design of Everyday Things.