AB2013 Statement

Last updated: July 2, 2026

This document is published pursuant to California Civil Code Section 3111 (AB2013). Recraft will update this document before each substantial modification to a covered system, as required by Section 3111.

This documentation applies to Recraft's proprietary generative AI image models offered to users in California. Third-party models accessible through the Recraft platform are developed by their respective providers and are not covered by this notice.

1. Purpose of Training Data

Recraft offers publicly available generative AI image models in the state of California, designed to produce high-quality raster and vector images in response to user prompts. We select our training data to enable our models to learn visual patterns, design styles, typography, and the relationship between visual characteristics and textual descriptions across a wide range of subjects and categories. Our training datasets support our models in generating diverse, design-oriented visual content from text inputs.

2. Dataset Sources and Ownership, Synthetic Data

Recraft models are trained on a dataset that consists of a mixture of the following source categories:

Publicly available data: Content from publicly accessible repositories.

Data from third parties and Recraft users: Non-public data from third-party providers, and data Recraft users provide through use of the Recraft service. By default, content from Teams plan accounts and content generated through the Recraft API is not used for training. For individual accounts, user content is used for training by default; users may change this in account settings so that their content is not used for training.

Internally generated data: Data generated and annotated internally by Recraft, including synthetic data to help support various training objectives and supplement real-world data.

3. Dataset Characteristics

The training data used to train Recraft models encompasses a diverse range of content types, including raster images and vector images, across a variety of subjects, styles, and categories, the textual metadata associated with these images, and human-provided annotations, ratings, and preferences to help reinforce model learning.

4. Dataset Scale

Recraft models are trained on datasets comprised of billions of images and associated text. The exact number of data points varies depending on the model version and phase of training.

5. Intellectual Property Status

Given the varying sources of our training data, Recraft's training dataset includes a mixture of data, including licensed data, data that may be protected by copyright and is used under fair use, data in the public domain, and data that is not eligible for copyright protection.

6. Personal Information

A portion of Recraft's training dataset includes publicly available content that may relate to individuals, such as images depicting people and associated textual metadata (e.g., captions or tags). Recraft does not intentionally seek out or use personal information to identify specific individuals in training data. However, because some data is sourced from publicly available content, personal information and aggregate consumer information, as defined in California Civil Code Section 1798.140, may be incidentally included.

7. Dataset Processing

Recraft training data undergoes several processing steps during model development, including deduplication; removal of low-quality images; safety filtering to remove certain data with known risk of containing child sexual abuse material (CSAM) and other categories of sensitive or disallowed content; and categorization based on relevance, quality, or image format.

8. Training Timeline

Recraft began collecting data to develop Recraft models in 2022, and continues to collect data today. These datasets were first incorporated into model development in 2022.