Five million people rank AI designs so machines can learn what looks good

Key points

• Intelligence, the company behind DesignArena, closed a $7.9 million seed round led by Index Ventures in 2025.
• DesignArena has 5.3 million registered users worldwide who vote on AI-generated visual outputs.
• The platform is generating $60 million in annualised recurring revenue from AI companies that buy the feedback data.
• Rival platform Yupp shut down earlier in 2025 after raising $33 million, showing this market isn't guaranteed.
• LM Arena, a similar platform focused on text rather than visuals, raised $150 million in a Series A in January 2025.

Co-founder Grace Li and her college friends built an AI game engine in 2025. The models produced working games. None were any fun to play.

That gap pointed toward something bigger. Automated tests can measure whether a game runs correctly; they can't tell you whether anyone will enjoy it. Only real humans could answer that at scale, the team decided, and they started looking for ways to collect honest human feedback across millions of users.

What they built became DesignArena. Using it feels like asking a friend to pick between two logos. You type a request, choose a format such as a website layout or a marketing image, then work through side-by-side pairs and pick the better one. The platform scores those votes, ranks the outputs, and hands back a winner.

So who is actually paying for this?

The paying customers are AI laboratories that build image and video generators. Those labs need to know whether their models produce output that real people find attractive, and automated scoring often misses the point.

"It was the missing bottleneck for a lot of these models to make improvements in the design space," Li says. "About a week later, we closed our first major deal with a frontier lab, and the rest is kind of history."

Because users must log in to retrieve results, Intelligence can also track how preferences shift across regions and across time. Li has noted, for instance, that web dashboards produced for Asian markets tend toward a denser, more information-heavy visual style than those aimed at Western audiences. That kind of nuance is exactly what automated benchmarks, which test AI models against fixed question sets, struggle to capture. Benchmarks can also be gamed: developers can quietly tune a model to score well on a test without actually improving what it produces.

The seed round, first reported by TechCrunch AI, included Conviction, A* and Valkyrie alongside lead investor Index Ventures.

Index Ventures' Neil Rimer has spoken publicly about where AI value flows and who captures it. We reported on his thinking in our July story on AI wealth and redistribution.

Should users be worried about how their votes are used?

Not urgently, but you should know what you're signing up for. Your preferences become data that AI companies purchase. You get a free tool to compare outputs; companies get insight into what millions of real people find appealing. A login is required, so your regional preferences and how they change over time are part of what Intelligence sells.

The market carries genuine risk. Yupp, a competitor with a similar model, closed earlier in 2025 after raising $33 million. It had more than 1.3 million users and paying clients among large AI labs but still couldn't build a sustainable business. Intelligence is betting that a tight focus on visual design, rather than text, gives it a position Yupp never had.

Common questions

What does DesignArena actually do for a regular user?

You describe what you want, pick a format such as a website or an image, and vote on which AI output looks better in a series of head-to-head comparisons. The platform uses your votes to deliver the best result and, with your permission, sells the aggregated data to AI companies.

Why can't AI companies just test their own models without paying for this?

They can, and they do, but automated tests tend to reward outputs that score well on narrow metrics rather than outputs that humans genuinely prefer. Collecting honest votes from millions of real users is hard to replicate internally, which is why companies pay for it.

Is $60 million in annual revenue unusual for a startup this young?

For a company that began a few weeks before its founders graduated in 2025, it's a striking figure. LM Arena, which applies a nearly identical approach to written text, raised $150 million in January 2025, just four months after launching its paid product. The comparison matters: if the text-evaluation market can support a $150 million raise, the visual equivalent may be larger still, given how much harder aesthetics are to benchmark automatically.