Sparkles, tiny icons, and chaos: every AI app has converged on the same visual language

Sparkles, tiny icons, and chaos: every AI app has converged on the same visual language

Open Claude, Codex, and Cursor side by side with Finder, Photos, and Mail on a Mac, and something odd happens: the three AI apps look more like each other than like the operating system they run on. Same tiny icons. Same shimmering "thinking" indicators. Same streaming text. And, of course, the sparkle emoji.

This visual convergence has crystallized in roughly three years, and it is now the dominant aesthetic of AI software. Designer Jim Nielsen, who documented the pattern in a blog post this week, published a screenshot where desktop AI apps share a noticeably smaller, thinner icon set than their native macOS counterparts. He observes that these AI apps clash with the system-level design of the applications around them 1.

The question is whether this convergence signals a design language maturing into permanence, or whether it reveals that nobody has figured out what AI software should actually feel like yet. Nielsen's essay suggests both. The distinction matters.

What will stick

Every computing era produces design idioms that either vanish as fads or harden into convention. The hamburger menu emerged from the screen-size constraints of early mobile and spread to nearly every corner of software 1. The sparkle emoji has undergone a similar transformation: before AI, it carried no established meaning. Now it is the universal shorthand for "this feature is AI-powered" 1.

Some AI-era patterns will likely follow the same arc. Streaming text was built for and refined by chat interfaces. Shimmering text, originally an AI "thinking" indicator, is already being repurposed to signal any asynchronous task: fetching, computing, loading 1. These are the patterns of a medium finding its footing. They solve real interaction problems that did not exist before, and they will probably outlast the novelty of chatbots.

What reveals the chaos

Then there are the patterns that do not solve problems so much as expose them.

Nielsen describes "whack-a-mole" UI controls in AI apps: you click a toggle, the entire interface repaints itself, and the toggle you were aiming at has moved somewhere else on the screen. He traces this to a deeper cause. "The non-determinism of AI's grain has seeped into its UI/X," he writes 1.

That sentence is more than a design critique. Large language models are probabilistic systems that produce different outputs for the same input. When that non-determinism leaks into the interface layer, into toggles that move and layouts that shift and responses that reorganize themselves between viewings, it creates a trust problem. Humans calibrate their expectations around consistency. An app that redraws itself unpredictably is asking you to trust software that does not hold still long enough to be trusted.

The tiny icons fit the same pattern. Nielsen notes that desktop AI apps, many built on Electron, use iconography that is noticeably smaller and thinner than what native macOS applications use 1. The icons look like they were designed for a different operating system because, in a sense, they were. The AI layer's visual language has diverged from the platform beneath it. Nielsen also identifies a cluster of other recurring elements: beige and cream color palettes, orange accents, and serif typefaces 1. Nielsen adds that he personally hopes tiny icons do not become the norm 1.

Mature or unsettled?

The hamburger menu survived because it solved a real problem well enough that designers stopped searching for alternatives. Streaming text and the sparkle emoji may travel the same path. They are functional shorthand for interactions that did not exist before AI.

But the whack-a-mole toggles and the interfaces that cannot hold still are not solved problems. They are symptoms of a deeper uncertainty about what AI software is supposed to be. Nielsen wonders how many of these aesthetics will "spread, take seed, and become part of common software interaction paradigms for years or decades to come" 1.

The honest answer is that the industry does not know yet. When every AI app converges on the same look in roughly three years, that convergence can mean two things. Either a mature design language has emerged, the way mobile settled on pull-to-refresh and the hamburger menu. Or nobody has found the right answer yet, and everyone is copying each other because copying is cheaper than inventing.

For now, the evidence points to the second. A medium that has found its footing does not produce interfaces that repaint themselves when you click a toggle. That kind of chaos is not a design language. It is the absence of one.