OpenAI Content Provenance and Authenticity Initiatives

OpenAI is implementing a multi-layered approach to content provenance to combat deceptive AI-generated media

OpenAI has joined the C2PA Steering Committee and is developing internal detection tools to provide transparency regarding the origin of audiovisual content. These efforts combine open industry standards with proprietary technical solutions to help users verify whether digital content was created or edited using OpenAI tools.

Adoption of the C2PA Open Standard

OpenAI has joined the Steering Committee of the Coalition for Content Provenance and Authenticity (C2PA), a widely used standard for digital content certification. This standard allows creators and platforms to prove the source of digital content, whether it is a raw camera output or an AI-generated image.

Implementation in DALL-E 3 and Sora

OpenAI has integrated C2PA metadata into all images created or edited via DALL-E 3 in ChatGPT and the OpenAI API. This metadata tracks the history of an image, including the tools used (e.g., OpenAI's DALL-E) and actions taken (e.g., format conversion or edits). OpenAI has also committed to integrating C2PA metadata for the Sora video generation model upon its broad launch.

Societal Resilience Fund

To promote the adoption and understanding of provenance standards, OpenAI and Microsoft have launched a $2 million societal resilience fund. This fund supports AI education through organizations such as the Partnership on AI, International IDEA, and Older Adults Technology Services from AARP.

Proprietary Detection and Watermarking Tools

OpenAI is developing internal provenance methods designed to be more resistant to removal than standard metadata.

DALL-E 3 Image Detection Classifier

OpenAI has launched a Researcher Access Program to allow research labs and journalism nonprofits to test an image detection classifier. This tool predicts the likelihood that an image was generated by DALL-E 3. Internal testing indicates the following performance metrics:

  • Accuracy: Correctly identifies ~98% of DALL-E 3 images.
  • False Positives: Less than ~0.5% of non-AI generated images are incorrectly tagged as DALL-E 3 images.
  • Robustness: Maintains performance through common modifications like cropping, compression, and saturation changes.
  • Limitations: Performance is lower when distinguishing DALL-E 3 images from those generated by other AI models, with 5-10% of images from other models being flagged as DALL-E 3 images.

Audio Watermarking

OpenAI has incorporated tamper-resistant audio watermarking into Voice Engine, its custom voice model currently in a limited research preview. This involves embedding an invisible signal into the audio that is difficult to remove.

Research into Text Provenance

OpenAI is prioritizing audiovisual content due to the higher perceived risks, but is actively researching text provenance solutions. Current findings include:

Text Watermarking

OpenAI has developed a text watermarking method that is effective against localized tampering (such as paraphrasing) but vulnerable to globalized tampering. Methods such as using translation systems, rewording with another generative model, or inserting and deleting special characters can circumvent the watermark.

Additionally, research suggests text watermarking may disproportionately impact non-native English speakers by stigmatizing the use of AI as a writing tool.

Text Metadata

OpenAI is exploring the use of cryptographically signed metadata for text. Because metadata is signed, it eliminates false positives, which OpenAI expects will be critical as the volume of AI-generated text increases.

Sources