Apple's FlowEval Uses AI to Grade AI-Generated Interfaces

Apple ML Research has introduced FlowEval, an AI system that checks the quality of interfaces generated by language models and AI agents by comparing navigation trajectories on real websites and on their AI-generated copies.

Why is evaluating AI interfaces so hard?

Because existing approaches sit between two extremes, and each is inconvenient in its own way. The first is bringing in experts: they accurately test critical user scenarios and catch subtle usability problems, but the method is slow and expensive. Scaling it to hundreds of iterated UI versions is unrealistic. The second is automated evaluators: fast and scalable, but less accurate and often opaque, leaving the developer unsure on what basis a score was given.

FlowEval takes a position between these poles, aiming to combine the scalability of automated methods with the accuracy of expert review.

How reference-based evaluation works

The framework's key idea is a reference-based approach: real websites are the starting point. FlowEval records navigation trajectories on the original web pages, then matches them against the interaction paths that the AI-generated interface supports.

What the system measures:

• Support for realistic user flows, not just visual similarity to the original
• How closely the generated UI's navigation trajectories match the real reference routes
• Specific points of deviation: where exactly the AI reproduced the functional structure incorrectly

The logic is simple: if the navigation flows in the generated interface are close to the originals, the AI has reproduced not only the look but also the functional structure of the page. Where the flows diverge, the system points to the specific component or step that caused the deviation.

This gives developers an objective, measurable signal of UI quality without having to bring in specialists for every iteration. The evaluation is also reproducible and more transparent than the "black boxes" of most existing automated evaluators.

Why does the industry need this?

The problem FlowEval addresses will only get worse. As coding AI agents move into mass production, in IDEs, standalone services and AI pipelines, the gap between a "pretty screenshot" and a "working interface" becomes critical. Many existing UI-generation benchmarks measure visual similarity or syntactic correctness of markup, but do not answer the key question: does navigation work, do forms get filled in, do buttons lead where the user expects?

FlowEval shifts the emphasis from "looks similar" to "works as it should." Notably, the tool is published by Apple ML Research, a company that actively develops on-device AI but opens up its methods less often than other major labs. That suggests the problem is acute enough for Apple to decide to share its approach with the academic community.

What this means

FlowEval offers a methodological bridge between costly expert review and opaque automated evaluators. For developers and researchers who use LLMs and AI agents to build interfaces, it means a way to systematically measure UI-generation quality and speed up iteration, without losing evaluation reliability.

How does AI check interface quality?

FlowEval compares navigation routes on real websites with routes in AI-generated copies and determines how functionally accurate the result is.

Why is evaluating AI interfaces harder than ordinary testing?

Existing approaches either require expensive experts for accurate testing, or rely on automated checks that miss subtle usability problems.