A design tool rolled back its layout generator after launch

The process wall

On 26 June 2024, Figma announced Make Designs as part of Figma AI, described in the announcement as a way to answer the 'blank canvas problem': a prompt typed into the Actions panel would return UI layouts and component options a designer could pick apart. The change was procedural before it was aesthetic. It inserted a generated first draft ahead of the moment a designer opens an empty file, shifting the earliest working decision from what to draw to what to keep or discard from a machine's layout guess.

What the documents show

A later retrospective, posted 17 July 2024, states the feature ran on 'off-the-shelf models like OpenAI's GPT-4o and Amazon's Titan model' with 'no additional training or fine-tuning.' To give the model material to draw from, Figma says it 'commissioned two extensive design systems (one for mobile and one for desktop) with hundreds of components,' supplied to the model as context rather than as training weights. The retrospective records that designer Andy Allen, who spoke at Figma's Config conference, noticed that prompting Make Designs for a weather app produced a result 'very similar to Apple's first-party app.' Figma's own investigation found that a handful of the commissioned components were 'similar to aspects of real world applications' and had not been vetted before entering the library.

Authorship and limitation

The distinction the retrospective draws is between a model that memorised outside work through training and one handed a curated reference library at run time. Figma is explicit that no training or fine-tuning occurred; the resemblance travelled through context assets a commissioning team assembled and failed to check, not through learned weights copying a scraped app. That reframes the failure as a curation and quality-assurance lapse rather than a dataset dispute, and it is Figma's own account, not an independent audit, that draws that line. The retrospective says the company 'removed the assets that were the source of the similarities' and disabled the feature while it built 'an improved QA process,' which leaves the review standard for a commissioned design system as an internal, unaudited commitment.

What survives

What a maker keeps is the working habit the incident exposes: a generated layout is an assembly of reference material chosen by someone else, and its resemblance to a shipped product is a property of that material, not a random accident of the model. Editorially, a reasonable check before publishing a Make Designs draft, or its successor folded into Figma's newer agent tooling, is to ask whether the output echoes a specific, identifiable app rather than a generic pattern.

• Read the retrospective's account of which components were removed and why.
• Compare a generated layout against the commercial app it most resembles before reusing it.
• Ask a design-AI vendor whether its reference material is commissioned, licensed, or scraped.

Figma's own record turns a single embarrassing prompt into a legible process failure: a reference library outran its own review, and the company's fix was curatorial rather than technical.