Study: Designer Knowledge Significantly Improves AI-Generated UIs

Anyone who has ever asked an AI to suggest a user interface for an app has likely thrown their hands up in despair. Although there is certainly no shortage of training material, Large Language Models do not have a knack for generating something aesthetic and at the same time easily usable from it. However, if you feed in the feedback of professional designers, the situation looks quite different. Apple has investigated this in a study and published the results on its Machine Learning Blog.

The problem with user interfaces is that the classic training of AI is apparently completely insufficient for design matters. In the conventional method of reinforcing learning from human feedback, a person assigns scores to the AI's outputs. With this method, which is applied during finetuning in the training phase of a new model, the AI is supposed to be able to distinguish between good and bad. But in reality, it still doesn't work that well for designs.

Apple's Trick: Not Just Good or Bad

Instead, the Apple researchers used a more complex method that also corresponds to the feedback process in design departments. With the help of 21 professional designers, designs were annotated with comments, marked with line drawings, or directly modified. The team fed this training material – a total of 1500 annotations – into the training data. This way, the AI knew not only whether something was good or bad, but also why. With visible success, as the Apple researchers explain in the paper: The AI's results were subjected to a professional evaluation again. Compared to the AI-generated designs of other models, including OpenAI's GPT-5, the specially trained model performed best.

What Apple intends to use the insights gained for remains open, as is usual with these research papers. In addition to pure basic research, it is conceivable, for example, that Apple may want to support its design team with AI assistance in the future. In the area of tools for developers, an AI model that supports inexperienced developers in creating suitable user interfaces for their apps would also be a great help.

The study used Qwen2.5-Coder as the base model. The 21 professional designers had between two and 30 years of professional experience and came from the fields of UI/UX, product design, and service design.

(mki)