Figma explains how its AI cloned Apple's Weather app

Figma recently pulled its shiny new generative AI tool, Make Designs, after users noticed that asking it for a weather app yielded something that looked almost identical to Apple’s iOS weather app. Now, the company is finally explaining how that happened.

In a detailed post, Figma VP of product design Noah Levin admitted the company failed to properly vet assets that were rushed into the tool just days before its annual Config conference.

The context window did the copying

When the Apple clone first surfaced on X, critics assumed Figma had trained its models on copyrighted iOS screenshots. Figma executives pushed back, confirming the tool relies on third-party models, specifically OpenAI’s GPT-4o and Amazon’s Titan Image Generator G1, with zero custom training on Figma user data.

Instead, Make Designs works through in-context assembly. Figma commissioned two bespoke design systems, one for mobile and one for desktop, featuring hundreds of handcrafted UI components and layout examples.

When a user enters a prompt, Figma feeds the prompt alongside metadata from these pre-built components into the model's context window. GPT-4o acts as an assembler, picking pieces and arranging them into an editable file, while Amazon Titan fills in the imagery.

Basically, the model was not hallucinating Apple’s interface out of thin air. It was simply following example screens that Figma itself fed into the prompt.

A last-minute push for the keynote

According to Levin, the core design systems were reviewed carefully during early private testing. But the wheels came off in the final sprint toward Config.

In the week leading up to the conference, the team added new components and example screens that were not vetted closely enough. A few of those assets closely mirrored existing real-world apps, causing the model to reproduce them when prompted.

Figma CEO Dylan Field took the blame for pushing the team to hit the Config deadline. Figma has since scrubbed the problematic assets and kept Make Designs disabled while building a stricter QA process, with no confirmed return date.

Why it matters

We think Figma’s architectural approach is genuinely clever, but this incident exposes a dangerous blind spot in prompt-based UI generation.

By having GPT-4o assemble vector design systems rather than generating flat bitmap images, Figma built a tool that produces genuinely useful, editable layers. That is miles ahead of pure diffusion models for real product design workflows.

However, leaning on human-crafted context examples means your output is only as safe as your weakest reference asset. Rushing third-party design components into production without visual similarity checks turned a flagship feature into an immediate copyright liability for users.

For teams building generative workflow tools, this is a clear warning. You do not need to train a model on proprietary data to create legal trouble; an unvetted context prompt can turn your AI into a high-speed copyright copier just as easily.