Designing for AI Optionality

Karl Koch came into design sideways: a fine art degree, a stint in account management that he calls getting lost, and then a startup job where he ended up doing design engineering before that was a title anyone used. He learned Vue, then SwiftUI the year it was announced, and has spent the last three and a half years at DuckDuckGo as their first design engineer, working on Search Assist — the company's answer to AI overviews.

The governing principle there is optionality, and it predates the current backlash. Search Assist can be set to always, sometimes, on-click-only, or never, and never sticks. There's also noai.duckduckgo.com, a full opt-out where you won't even see promos for AI features. Karl's read on the traffic spike is that nothing changed on their end — people are only now noticing. He's blunt about why competitors can't follow: money that large has to be recouped by putting AI in front of everyone, opt-out or not. On accuracy, which he calls the biggest problem in the category, the team spends a few extra milliseconds vetting every response before it renders, maintains hand-scored datasets offline, and re-verifies everything each time the model changes. The design job in that is escape hatches — error states for when the answer isn't good enough to show.

He also thinks the industry is swinging back toward generalists, closer to the early web than to the specialist handoff chains that came after. The danger is offloading: building with Claude and learning nothing, which is how broken apps ship with API keys in the code. He doesn't think every designer needs to code, but he wants enough fundamentals to judge what the tools return — and he's sharp about the company version of the mistake, where one person plus an AI is expected to ship everything. Everything an AI does is the best average of everything, and nobody moves the world forward shipping averages.