by Figma has taken down its “Make Designs” generative AI tool after it was found producing designs that closely resembled Apple’s weather app. This incident raised concerns about potential legal issues for users and suggested that Figma’s AI might have been trained using Apple’s designs. In response, Figma CEO Dylan Field clarified that the AI was not trained on Figma content or app designs, and the company has now provided a detailed explanation in a blog post.
The Explanation: Figma’s VP of Product Design, Noah Levin, explained that while the company “carefully reviewed” the underlying design systems of Make Designs during development and its private beta, new components and example screens added shortly before the Config event were not vetted thoroughly. Some of these new assets inadvertently resembled real-world applications and showed up in the AI-generated designs when specific prompts were used.
Once Figma discovered the issue, the problematic assets were removed from the design system, and the Make Designs feature was disabled. Levin stated that the company is now working on an “improved QA process” before reintroducing the tool, although no specific timeline was provided. In a recent interview, Figma CTO Kris Rasmussen mentioned that the feature is expected to return “soon.”
The Incident: The issue came to light after Figma launched Make Designs in a limited beta during its Config event. Shortly after, users on X (formerly Twitter) posted examples of the tool generating designs that closely mirrored Apple’s weather app. Figma quickly pulled the feature, with CEO Dylan Field taking responsibility for rushing the team to meet the Config event deadline.
Technical Details: In the blog post, Levin elaborated on the design systems powering the tool. To ensure the model could generate designs across various domains, Figma commissioned two comprehensive design systems—one for mobile and one for desktop. These systems included hundreds of components and examples of different ways to assemble them. The metadata from these components, along with user prompts, were fed into the model’s context window, allowing it to compose designs from a subset of these components. Amazon’s Titan, a diffusion model, then generated the necessary images for the designs. Essentially, the AI aids in identifying, arranging, and theming small composable templates to create a starting point for users.
Levin did not disclose who Figma commissioned to create these systems, and the company declined to comment further.
This incident highlights the challenges and risks associated with integrating AI into design tools, particularly regarding originality and intellectual property. Figma’s prompt response and transparency about their QA process improvements indicate their commitment to addressing these issues and preventing similar occurrences in the future.
