Blind Test: AI Design Preferred Over Real

Subframe co-founder Adam Ginzberg launched Real Design Benchmark, where Claude Opus 5.5 and GPT-6.1 Sol design screens for real products based on a text brief and screenshots of other screens, and after several thousand blind responses, 57% of participants preferred the AI version, recognizing it in only 58% of cases.

Compiled by Sergey KostenchukPublished 2026-10-15Updated 2026-10-15

2026-10-15 Research Anthropic

🤖 Blind Test: AI Design Preferred Over Real

Subframe co-founder Adam Ginzberg launched Real Design Benchmark: Claude Opus 5.5 and GPT-6.1 Sol design screens for real products based on a brief and other people's screenshots, without seeing the original. Participants choose the best screen and guess where the AI's work is. According to data from October 12, 2026 (116 tasks from 52 products), 57% preferred the AI version, 33% — the real one, 10% saw no difference.

🌍 For the industry, this is a measurable argument for design automation and a new quality criterion for LLMs outside of code. Caveat: the test measures adherence to an existing design system, not new ideas, and 58% recognition against 50% random — a point estimate.

👤 You can play yourself at realdesignbenchmark.com, read the open methodology, and see how the models write HTML screens for real products.

Source 1: https://www.realdesignbenchmark.com Source 2: https://www.realdesignbenchmark.com/methodology

Sources