OpenAI unveils a tool to track AI-generated images

OpenAI Introduces a Tool for Verifying AI-Generated Images

OpenAI recently announced the launch of a new tool in research preview aimed at identifying whether an image has been generated by its technologies, such as ChatGPT. This initiative is part of a broader effort by the California-based company to comply with the standards of the C2PA, an open standard that allows for the addition of provenance metadata to digital files. This aims to indicate whether content has been created or modified by artificial intelligence.

Two Types of Signals for Effective Detection

The tool, available since May 19, 2026, uses two complementary types of signals to verify the origin of images. The first, Content Credentials (C2PA), associates provenance metadata with digital content, allowing users to determine if an image was generated via the API or ChatGPT. The second, SynthID, is an invisible watermark developed by Google DeepMind. This watermark is designed to withstand modifications such as cropping, adding filters, or lossy compression.

When a signal is detected on an uploaded image, it indicates that it likely originates from OpenAI's tools. Although the company admits that the tool is not infallible, it claims that the detected signals are reliable and that false positives are rare.

OpenAI emphasizes the importance for users to understand and verify the origin of media in order to interpret it confidently. Provenance signals provide context about the origin of content, how it was created or modified, and whether it is authentic.

Towards Broader Verification

Currently, the tool focuses on detecting visuals generated via the API, ChatGPT, or Codex. However, OpenAI plans to expand verification to all platforms in the coming months. The company hopes to contribute to a more interoperable provenance ecosystem by leveraging Content Credentials, adopting SynthID, and offering public verification tools.

Using the Detection Tool

OpenAI's tool is accessible in research preview and can be used by uploading a suspicious file to the dedicated platform. The analysis, which takes a few seconds, is compatible with several image formats, including PNG, JPEG, and WEBP.

To optimize results, OpenAI recommends cropping screenshots as closely as possible to the image and avoiding the import of files containing multiple images. It is also noted that uploaded images are not used for training purposes.