Teaching AI 'Good Enough' Design Taste Requires Expert Input
As AI-generated website design rapidly improves, the challenge shifts from basic functionality to sophisticated aesthetic judgment. Expert data is becoming crucial for training AI to understand and replicate high-quality design.
The quality of AI-generated website designs is rapidly improving, according to Ben Blumenrose, co-founder of venture capital firm Designer Fund and a designer with two decades of experience. While human designers can still identify AI-generated sites by certain stylistic cues, the output has become significantly better over the past year and is expected to continue improving.
The artificial intelligence industry is working to enhance creative output through refined evaluation metrics and automated tools. As AI models become more adept at objective tasks, the focus is shifting towards subjective qualities like 'taste' in creative fields such as writing and design. This involves teaching AI to understand aesthetic principles and produce work that appeals to expert human judgment.
Developing AI 'taste' is a complex challenge that requires specialized data. Unlike general-purpose models trained on broad internet scrapes, domain-specific models need to be trained on curated datasets of high-quality work. This data must reflect expert opinions and historical trends within specific creative disciplines.
Companies like Figma and Krea are actively collecting and utilizing expert-level design examples to fine-tune their AI models. Figma's AI Research Lead, Sumithra Bhakthavatsalam, notes that off-the-shelf models often struggle with design quality assessment. Krea's co-founder Diego Rodriguez emphasizes the need for professional-grade visual vocabulary, gathered from millions of references, to ensure their AI tools meet the demands of creative professionals.