Company Apple slowly Apple continues a relatively quiet but systematic exploration of how generative artificial intelligence could impact app development. The latest study, published by researchers at Apple This time, it's not just about code generation, but about user interface design. In other words, AI isn't just learning to "write applications," it's also starting to learn how they should look. But this is a much more sensitive area and it's worth dwelling on for a moment.
You might be interested in
Apple He got to this point gradually. A few months ago introduced the UICoder project, an open family of models whose goal was to generate functional UI code based on textual input. The main emphasis back then was not on aesthetics, but on practicality – that the code compile, match the description, and roughly meet expectations regarding the layout and behavior of the interface. Design as such was more of a side effect.
A new study titled Improving User Interface Generation Models from Designer Feedback takes a different approach to the problem. The authors openly state that commonly used methods of learning from human feedback (RLHF), typically based on simple evaluation or comparison variant, are not ideal for the field of UI design. According to them, they do not correspond to the real way designers work and ignore the arguments, context, and gradual adjustments that are essential for interface design.
Humans vs. AI
Instead Apple took a more practical approach. The research involved 21 professional designers with varying lengths of experience—from about two years to more than three decades—and from a variety of fields, from UI/UX to product to service design. These experts were not simply asked to “choose better variantu", but directly modified the AI-generated designs. They used written comments, sketches, and direct interventions in the interface. It was these specific changes that then became the training data.
In total, the researchers collected 1,460 annotations, from which they created pairs of the original design and its designer-modified version. Based on them, a so-called reward model was created, which evaluates the quality of the interface based on the rendered preview and text description and assigns it a numerical score. The better the design matches human designer judgment, the higher the score it receives. This model is then used to further fine-tune the UI generator.
As a basic generator Apple used the Qwen2.5-Coder model and subsequently applied the same reward model to smaller and newer ones varianty Qwen to verify that the approach is transferable across different model sizes. The results are impressive from a technical perspective. Models trained using designer feedback, especially sketches and direct edits, achieved noticeably better quality than both the baseline version and models trained using only classical evaluation. The top-rated model even outperformed in internal tests GPT-5, using a relatively small amount of high-quality feedback.
Gallery
AI only nestachi
The authors of the study also openly point out the limits of the whole approach. Design is a highly subjective discipline, and agreement on what constitutes a “better interface” is not self-evident. When researchers independently evaluated the same designs that designers agreed on, they agreed on their choice only about 49% of the time. Significantly better agreement occurred when designers actively edited their designs – agreement increased to 63,6% for sketches and even 76,1% for direct edits. The study indirectly shows that showing a concrete solution is significantly more beneficial for AI learning than abstract evaluation.
In a broader context, it is interesting to compare this research with the approach of other technology companies. For example studio Rockstar Games recently announced that they don't use generative artificial intelligence at all when developing GTA 6This can be a conscious choice, especially for projects where atmosphere, detail and handwork play a key role.
Apple is thus walking a fine line. It is clear that AI can be a very powerful tool for rapid prototyping, finding alternative solutions or supporting less experienced developers and designers. At the same time, the study itself suggests that without human judgment and the ability to explain specific decisions, it is difficult to create quality design. Rather than replacing designers, the research shows the direction in which their work could change in the future - with AI as an assisting tool, not as the main author of the user interface.
