OpenAI, a leading artificial intelligence (AI) research organization, announced plans to adopt the C2PA (Coalition for Content Provenance) standard. This will be used to add metadata to images rendered by DALL-E 3, a part of their generative AI model family. The implementation is set for February 12th and will be available for both desktop and mobile users.
This decision aims at combating the use of AI products for propagating disinformation. It’s part of a broader industry-wide initiative where companies like Meta, previously known as Facebook, have shown interest in identifying and labeling AI-created content.
C2PA is an open technical convention that imprints metadata on images. This information represents elements such as the origin, editing history, and other related specifics. OpenAI stated that images produced directly through DALL-E 3, or indirectly via ChatGPT or its API, will now incorporate C2PA data.
The incorporation of C2PA data into an image allows it to be submitted to platforms like Content Credentials Verify for provenance verification purposes. The data is also useful for social media platforms intending to mark and label AI-produced images.
However, integrating C2PA data does slightly increase the file sizes. Estimates from OpenAI suggest an approximate elevation of 3% for PNG files, 5% for WebP files via API, and a 32% increase for WebP via ChatGPT.
The move to incorporate C2PA isn’t without its limitations. OpenAI admits that it’s not a concrete solution for verifying provenance since the metadata can be easily or accidentally removed. Standard actions like taking screenshots or converting PNG files to JPG can lead to unintentional loss of C2PA data. Additionally, most social media platforms remove metadata when uploading images.
Despite these shortcomings, OpenAI believes the addition of C2PA data into DALL-E generated images will advance the credibility of digital data by enabling users to spot key signals. However, critics point out that its ease of removal renders it less effective against deliberate disinformation campaigns. Alternative solutions like Google’s SynthID show promise in providing watermark-like security but are yet to be considered by OpenAI due to competition dynamics.
