Figma launches AI-driven tools with strong privacy focus

Figma has launched a suite of AI-driven features designed to enhance efficiency and creativity in design workflows. Known collectively as Figma AI, these tools offer inspiration, explore various design directions, and automate repetitive tasks.

"All of the generative features we've launched to date are powered by third-party, out-of-the-box AI models and were not trained on private Figma files or customer data," according to the company. The announcement underscores the use of publicly accessible Community files to fine-tune visual and asset search capabilities.

Figma has indicated plans to create new models that better integrate with the platform's specific concepts and tools. "To make these improvements, we need to train models that better understand design concepts and patterns, and Figma's internal formats and structure through Figma content," the company stated.

Data privacy and security are key components of Figma's AI strategy. The company reiterated, "Our model development process is designed to protect your privacy and confidential information." Several measures are in place, including encryption of all data in transit and at rest, tailored permissions, and strict user access controls. Additionally, third-party providers are restricted from using uploaded data for their own model training and are limited in how long they can store user data. For example, OpenAI and other third-party providers only store data temporarily or not at all, as necessary to process requests and enable AI features.

To further ensure privacy, data used for model training is de-identified and aggregated. Figma noted, "We do not use any data from Figma for Education or Figma for Government accounts for model training." The company also offers users the option to share their content for AI model training, requiring admin control to initiate such sharing by August 2024. Examples of customer content include text, images, comments, annotations, and layer properties.

Figma distinguishes between customer content and usage data. The latter involves metrics relating to how the platform is accessed and utilised but does not include user content itself. This data type is employed in an aggregated, de-identified format to ensure user privacy. Examples of usage data include information on the frequency of content access and various technical logs.

Admins have two new settings to manage AI use and content training. These settings can be adjusted anytime to control AI features and content training access. AI features are active for all plans by default, while content training settings vary. For instance, content training is switched on by default for Starter and Professional plans, whereas for Organisation and Enterprise plans, it is off by default.

"These are team-level settings for Starter and Professional plans and organisation-level settings for Organisation and Enterprise plans," Figma elaborated. However, the content training setting will only take effect from August 2024. Until then, no content training will occur. Post-implementation, new content and edits will fall under the chosen settings unless admins opt to turn off content training, thereby excluding new inputs from being used for model training.

Regarding the Figma Community, the company stressed its appreciation for Community creators. It confirmed that free files within the Community are governed by the CC BY 4.0 licenses, enabling transformation but necessitating attribution. Presently, Figma does not train generative models that generate designs based on Community files, though it uses public, free files for specific needs like semantic and visual search improvements. Paid files remain excluded from AI model training efforts.