DesignArena began with a practical problem: AI models could generate working games, but the results were not necessarily enjoyable. For Grace Li and a small group of college friends, that raised a harder question than whether the software functioned. It asked how anyone could measure taste, fun and preference at scale.
The answer became DesignArena, a product now used by 5.3 million people around the world. On Monday, the company behind it, Intelligence, announced a $7.9 million seed round led by Index Ventures, with participation from Conviction (Sarah Guo and Mike Vernal), A*, Valkyrie, and others.
From game feedback to AI evaluation
Li says the company started a few weeks before graduation in 2025. The early team was working on an AI game engine, and the technology could produce functional games. The problem was that functionality did not equal fun.
That gap pushed the founders toward human judgment. If a model can produce a game, image, website or other visual asset, the remaining question is whether people actually like what it made. Automated tests can help in some settings, but the source article makes clear that the team saw no replacement for direct human feedback when evaluating design quality.
DesignArena grew out of that insight. Instead of asking users to score an output in isolation, the tool asks them to compare options. That comparison-based approach gives AI companies a clearer signal about which generated result a person prefers.
Li described the problem this way: “It was the missing bottleneck for a lot of these models to make improvements in the design space,” adding that “About a week later, we closed our first major deal with a frontier lab, and the rest is kind of history.”
How DesignArena works for everyday users
For non-enterprise users, DesignArena resembles a sophisticated model router. The interface includes a Chat-GPT-style prompt window, with dropdowns for websites, images and a dozen other visual formats. Users enter a request, choose a format and style, and then review competing outputs.
The ranking process is built around repeated “A vs. B” choices. Users compare generated results until they have ordered a small set of outputs from best to worst. The person using the product gets a better result, while the platform collects structured preference data from those decisions.
That matters because users often do not care which model created a particular output. As Li puts it, they want the best output they can get. This makes their rankings useful because the feedback is attached to preference, not model loyalty.
In plain terms, DesignArena turns ordinary usage into a continuous evaluation loop. A person asks for a visual result, chooses what looks better, and that judgment becomes a data point about what users value in generated media.
Why AI labs are paying attention
The enterprise side is where the source article says the platform’s real value sits. Participating models can use DesignArena as a stream of instant feedback for media-generating models. For frontier labs trying to improve visual output, that kind of human-led evaluation can be commercially important.
Li says the site is currently generating $60 million in ARR. That figure places Intelligence in a notable position among companies supplying human evaluation data to the AI industry.
The value is not only in which output wins a comparison. Because users must log in to get their output, Intelligence can also track how preferences shift across different continents and over time. Li notes, for example, that web dashboards in Asia tend to have a more maximalist design style.
Those differences matter for AI models that generate design work. A visual result that looks clean, useful or polished to one group of users may not map perfectly onto another group’s expectations. DesignArena’s model is built around gathering those preference signals directly from people making choices.
Human taste versus automated benchmarks
The source article frames human feedback as a complement to automated benchmarks. Automated systems can operate at a greater scale, but they can also be gamed or manipulated. The Hugging Face breach last week is cited as a dramatic example of that vulnerability.
DesignArena’s approach is different because it relies on repeated user judgments. That does not make it effortless or automatically dominant, but it gives AI companies another way to understand quality in areas where clear right answers are hard to define.
Design, images, websites and games often involve subjective choices. A model output may satisfy a prompt and still feel wrong to a user. It may include the requested elements but fail to communicate the intended style. Human comparison data can expose those gaps in a way that purely automated evaluation may miss.
A promising market with real risks
The source also points to a clear warning: crowdsourced human feedback is not guaranteed to become a durable business. Yupp shuttered its doors earlier this year, less than a year after launching, after raising $33M from a16z crypto’s Chris Dixon. It also had frontier models as customers and said it had over 1.3 million users, but it still could not build a sustainable long-term business.
At the same time, other companies built around human evaluation appear to be growing. LM Arena, which applies a similar idea to text-based responses, raised $150 million in a Series A in January, four months after formally launching its paid product.
That contrast is important. Human feedback can be valuable, but value alone does not settle the business model. The companies in this space still have to turn user activity, enterprise demand and model evaluation into a product that customers keep paying for.
For Intelligence, the new $7.9 million seed round gives DesignArena more room to prove that its feedback loop can become a lasting part of how AI models improve visual output. The bet is straightforward: as AI systems generate more media, the industry will need better ways to measure not just whether an output works, but whether people actually prefer it.