Designing AI Interfaces

Foreword

One of the most persistent myths in AI right now is that design doesn’t matter when the AI is powerful enough. The argument goes something like this: if the model is smart enough, it doesn’t need a good interface. Users will just talk to it, it’ll understand them, and everyone goes home happy. You can tell it what you want, and it does it! UI is seen as old-fashioned, just like Windows 95.

I’ve been arguing against this for years, not because I’m a designer defending my turf (although let’s be real, I am), but because the argument is empirically wrong. And now, thankfully, Louise Macfadyen has written the book that proves it.

The book you’re holding makes a case I find deeply satisfying, not because it vindicates a position I already held (although, again, it does), but because it goes further than my ranting and arrives somewhere much more useful and practical. Macfadyen takes a deceptively simple framework (input-computation-output) and shows that different species of potential failure haunt each stage. Her argument isn’t that design should be layered on top of AI (Lipstick on a Pig 3.0 anyone?); it’s that design is the discipline that determines whether an AI system works at all, for actual people, in the actual world.

We’ve been in this situation before, and we’ve been wrong in the same ways. When ELIZA arrived in 1966, people knew it was a program. MIT professor Joseph Weizenbaum built it as a “trivial” demonstration of natural language processing. He didn’t expect ...