Design Arena's Human Voting Became a $60M AI Business
A platform that ranks AI-generated designs by blind pairwise human votes raised $7.9 million on August 3, 2026, at more than five million users and $60 million in annual revenue — evidence that frontier labs will pay for the preference data an automated benchmark cannot produce.
Design Arena, a platform where people vote on anonymous AI-generated designs, raised $7.9 million in seed funding on August 3, 2026, led by Index Ventures, with Conviction, A*, and Valkyrie also participating. The round landed at a company that had already reached more than five million users worldwide and an annual recurring revenue of about $60 million, up from roughly $5 million six months earlier.
The company behind it, called Intelligence, started in 2025 as a side effect of a different project. Co-founders Grace Li and Kamryn Ohly, then computer science students at Harvard, were building an AI game engine that never shipped. The comparison tool they wrote to judge the engine's output turned out to be the more useful thing, and they kept building that instead.
The mechanism is a blind pairwise vote. The same prompt goes to several AI models at once, their identities hidden until after a person picks a preferred output. A single prompt runs through multiple rounds of comparisons, including a winner-versus-winner and a loser-versus-loser pass, before Design Arena's Bradley-Terry model converts the results into an Elo-style score. The categories run past a dozen: websites, mobile apps, games, UI components, data visualization, 3D assets, logos, and slide decks each get their own leaderboard.
Voting itself is free. The revenue comes from the other side of the same data: AI labs building models that generate images, code, and other visual media pay for the resulting preference records as evaluation feedback. Li has described it as the missing input those models needed to keep improving — a rendered webpage or a logo does not reduce to a formula the way a chess result does, so judging whether one is actually good has to come from people with no stake in which model wins.
That is also why the format sits next to something like Arena, formerly LMArena, rather than apart from it. Arena already runs blind pairwise voting for text and code; Design Arena applies the same structure to visual and generative-design output, a category where automated grading is weaker and a human glance still catches what a metric misses.
The commercial case moved fast once it existed: according to Li's account to TechCrunch, the company closed its first paying deal with a frontier lab within about a week of offering the enterprise product. Whether that pace holds as more evaluation platforms chase the same labs is the open question — but for now, the data shows AI labs treat outside human judgment on design as worth buying rather than building in-house.