Design Arena just raised $7.9 million, and human taste is becoming infrastructure for AI

Design Arena’s creators have raised $7.9 million in seed funding led by Index Ventures, with Conviction, A*, Valkyrie, and others participating. The company, now operating as Intelligence, says the product is already used by 5.3 million people and is generating $60 million in ARR from frontier labs and other customers paying for human feedback on AI-generated media.

That’s a serious number for a startup built around a simple premise: if you want models that make better images, websites, or games, people still have to tell you what looks good.

AI evaluation has long been treated like a niche research problem. It isn’t. As model output gets cleaner, the failures get harder to spot. A generated game can run fine and still be dull. A website can satisfy the prompt and still feel wrong. A product mockup can be technically coherent and visually dead. Benchmarks miss that. Static test sets miss it even more.

That’s where Design Arena sits.

Why taste is now a product category

Grace Li, who co-founded the company with a few college friends in 2025, says the original project started as an attempt to make an AI game engine work. The models could generate functioning games. They just weren’t fun. That distinction matters, and it’s becoming expensive.

AI labs are already good at training for correctness, format adherence, and basic user satisfaction. Those are measurable. Taste is messier. It shows up in preference data, pairwise ranking, and the kind of human judgment that doesn’t fit neatly into a loss function. The hard part is turning that into signal at scale.

Design Arena’s interface reflects that. For consumer users, it looks like a model router with a ChatGPT-style prompt box and dropdowns for websites, images, and other visual formats. You submit a request, choose a format and style, then rank outputs in head-to-head comparisons until the system builds a preference ordering.

Anyone who’s worked on RLHF, reward modeling, or preference optimization will recognize the setup. The difference is that Intelligence turns human evaluation into a public, high-volume feedback engine instead of keeping it inside a lab.

Pairwise comparisons are often more useful than absolute scores. Ask people to rate output from 1 to 10 and the scale drifts. Ask them which of two outputs is better and the data is usually cleaner. Enough of those comparisons can feed ranking models, reward models, or post-training pipelines aimed at what people actually prefer.

It’s not glamorous. It’s useful.

The enterprise business matters more

The consumer side gets attention because 5.3 million users sounds large. The money is on enterprise.

Li says frontier labs pay for the service because it gives them instant feedback on media-generating models. That makes sense, and it may be the strongest part of the business. Teams training or fine-tuning these systems do not need more synthetic benchmarks. They need messy human judgments that reflect how people react to outputs in practice.

The value is in the label economics. Traditional evaluation pipelines are slow, expensive, and too narrow in who gets to annotate. A product with a broad user base can generate more comparisons, faster, and with less setup. If it also handles routing, format selection, and ranking, it becomes a data flywheel. Users get generated content. The company gets training signal. Customers get evaluation data they’d have trouble collecting themselves.

The catch is that “human feedback at scale” can mean very different things depending on who is labeling and why.

Crowdsourced preference data is noisy. Sometimes that noise is fine. Sometimes it’s garbage. Low-effort judgments, prompt gaming, and demographic skew can all distort the result. A platform with millions of users still has to figure out which slices of that traffic are trustworthy for which jobs. A game studio doesn’t need the same signal as a design tool team. A frontier lab tuning a multimodal model cares about different failure modes than an ad-tech company generating landing pages.

The challenge is not just collecting votes. It’s weighting them.

Why investors keep funding this

Design Arena isn’t the first company to try this. Yupp raised $33 million from a16z crypto’s Chris Dixon, got some frontier models as customers, said it had more than 1.3 million users, then shut down earlier this year. That’s the warning sign sitting next to this round.

Human-evaluation products can look like a strong wedge and still fail as a business. Usage doesn’t automatically turn into durable revenue. Enterprise customers may pay for a while, then build something in-house, switch to a cheaper workflow, or only need the service for one training cycle. If the platform is too tied to one hot category, the revenue can swing hard when demand shifts.

LM Arena has done better by focusing on text-based response comparisons and raising a $150 million Series A in January, just four months after launching its paid product. Different modality, same basic insight: humans are still the best judge of certain model qualities, and companies will pay to outsource the machinery around that judgment.

Investors keep showing up because evaluation data becomes more valuable as models improve. Better models make evaluation harder, not easier. Once outputs are good enough to be plausible, benchmark scores stop telling the full story. Preference data starts to matter more.

That’s especially true in visual generation. A generated image can be technically clean and still feel off. A website can be valid HTML and still look like it was assembled by someone who’s never met a designer. A game can match the prompt and still feel like a demo rather than something anyone would play. Human preference data catches the failures automated metrics keep missing.

What engineers should watch

If you’re building model products, Design Arena points to a few practical things.

Evaluation is becoming part of the product loop, not a side task. Teams that treat human feedback as a one-off research exercise will miss where this is going. The better systems will mix offline benchmarks, synthetic tests, and live preference data from real users.

Modality matters. Text ranking is one problem. Images, websites, and games are messier because preference is more subjective and more context-dependent. A single winner from a pairwise vote doesn’t tell you whether the model is weak at composition, typography, interaction flow, or just this prompt. You need richer metadata and better task partitioning if the data is going to be useful downstream.

The feedback infrastructure has to be hardened too. If a platform is collecting user rankings at scale, it needs defenses against spam, adversarial prompting, and model-specific bias. Privacy handling matters as well. Users may think they’re playing with a public demo, but they’re often generating training data for paying customers. That line should be made clear.

There’s also the cost question. Human evaluation is expensive in a different way from GPU inference, but it’s still expensive. Paying users, filtering bad annotations, and supporting enterprise-grade data export all add overhead. If the company can’t keep the loop cheap enough, the economics break.

A decent business, with real limits

If Intelligence’s numbers hold up, it’s in a strong spot. $60 million in ARR from a category this new is real traction, not vanity usage. The funding round suggests investors think the company can keep turning consumer interaction into paid evaluation services.

Still, the market has a sharp edge. Human preference data is valuable, but it’s not infinitely defensible. If a customer can reproduce most of the workflow with an internal UI, a small annotation team, and a few hundred power users, the moat gets thinner. If the platform becomes too generic, it risks turning into a commodity ranking layer with a nicer interface.

The better version of this business is one that gets deeply embedded in model development. That means more than collecting votes. It means structuring data so customers can use it for post-training, reward modeling, regression testing, and product QA. It means enough scale to cover different modalities and enough discipline to keep the signal clean.

That’s a high bar. It’s also why this round matters. The AI industry keeps talking about bigger models, more tokens, and faster inference. One of the most useful things a startup can sell is a way to answer a much older question: did humans actually like the thing?

In 2026, that’s a real business.