AI image generation is already impressive — but who decides if AI's output looks good?

The answer is: still humans. That's why Design Arena could raise funding — on August 3, the AI aesthetic evaluation company announced a $7.9M seed round, with 5.3 million global users on the platform doing one thing every day: rating AI-generated designs. Which one looks better? Which is ugly? Which color scheme feels right? Which layout feels premium? Pure human subjective judgment, with no standard answer key.

AI's Achilles' Heel: Taste

You've probably had the experience: using Midjourney or GPT Image, tweaking every parameter perfectly, but the output still feels like it's "missing something." What's missing? It's not that details aren't realistic enough, or composition isn't standard enough — it's that the "vibe" is off. AI-generated designs are often technically flawless, but lack that "spark" and "taste" human designers can feel.

This isn't AI's fault. Today's LLMs are fundamentally statistical — having seen millions of excellent designs, they learn what pixel arrangements are statistically likely to be "good." But taste is often anti-statistical: the most classic designs frequently break conventions; the most tasteful choices are often not the "safest" ones. You can't learn to break conventions from statistics.

That's the problem Design Arena is solving. It's not a design tool — it's an "aesthetic data labeling platform", but far more sophisticated than traditional labeling. Traditional labeling pays people to do mechanical judgments (is this a cat? is this a car?); Design Arena has real designers and creatives making subjective calls (which is more premium, A or B? does this logo convey the right brand feeling?). These judgments are collected and used to train models' "aesthetic capability."

The Era of Vertical RLHF

Design Arena's funding actually foreshadows a trend: RLHF (Reinforcement Learning from Human Feedback) is moving from general to vertical.

GPT-4-era RLHF was general — getting a crowd to answer "which response is better" to train models for general conversation. But in vertical domains like AI design, AI coding, AI medicine, AI law, general feedback isn't enough. You don't want average people's judgments — you want professionals' judgments. Judging whether code is good takes a programmer; judging whether design looks good takes a designer; judging whether a diagnosis is correct takes a doctor.

This means a new layer is emerging in the AI industry chain: the vertical feedback layer. Just as the internet era needed content moderation and data labeling, the AI era needs professionals across fields to continuously provide high-quality feedback to models. Design Arena is taking the design vertical; similar platforms will inevitably emerge for coding, writing, music, medicine, and more.

Of course, this has a controversial side. If AI aesthetics are trained entirely on human designers' collective judgments, will AI-generated designs converge toward "average good design" and lose genuine creativity? After all, history's greatest designs were often misunderstood by most people in their time. But at least at this stage, getting AI to reach average human aesthetic levels is already a big enough market.

AI can generate a million images, but picking which one is usable — that still takes a human. That might be designers' strongest moat in the AI era: it's not that you draw faster than AI, it's that you know what good is.

See you tomorrow.