en.Wedoany.com Reported - OpenAI announced on May 19 a comprehensive upgrade to its provenance system for AI-generated content. Building upon the existing Content Provenance and Authenticity (C2PA) metadata standard, the company has officially integrated Google DeepMind's SynthID invisible digital watermarking technology, forming a "metadata + watermark" dual-layer verification architecture. This is initially applied to images generated via ChatGPT, Codex, and the OpenAI API. A simultaneously launched public verification tool allows users to upload images to detect whether they were generated by OpenAI models.

In its official announcement, OpenAI defined this update as a multi-layered provenance model. C2PA content credentials, based on metadata and cryptographic signatures, record the creation source, edit history, and signature information within the file, allowing platforms and users to trace the complete content lineage. However, C2PA metadata is highly susceptible to being stripped or accidentally lost during screenshots, format conversions, or platform uploads; a single layer of metadata alone cannot cover real-world online distribution scenarios. SynthID takes a completely different technical path—it embeds an imperceptible digital signal directly into the pixel layer of an image. This signal remains detectable by algorithms even after common operations like screenshots, compression, rotation, or cropping, technically compensating for the durability shortcomings of metadata solutions. OpenAI summarized this by stating: "The two systems reinforce each other—C2PA helps content carry detailed context, while SynthID preserves a signal when metadata cannot survive. Watermarks are more durable across transformations like screenshots, while metadata can provide richer information than a watermark. Together, they make provenance more resilient than either single-layer solution."

SynthID was first launched by Google DeepMind in 2023, with its core technology based on steganography principles. Its watermark embedding and detection are performed by two deep learning models respectively. The embedding model adds imperceptible signals directly into pixels, audio waveforms, or text probability distributions during AI content generation, while the detection model extracts this signal from the content and assesses the likelihood of it being AI-generated. Google DeepMind scientist Pushmeet Kohli explained in public interviews that the team invested significant research to ensure SynthID's sufficient robustness against various transformations. Prior to this integration by OpenAI, SynthID had already been used to label over 100 billion images and videos, as well as audio content equivalent to 60,000 years in duration, within Google's internal products. NVIDIA also announced the integration of SynthID into its Cosmos world foundation model, with companies like Kakao and ElevenLabs joining the adoption ranks simultaneously. SynthID is evolving from an internal Google tool into a cross-platform watermarking infrastructure.

The public verification tool launched alongside SynthID is another key focus of OpenAI's release. After users visit the verification portal and upload an image, the system automatically scans for C2PA content credentials and SynthID watermarks to confirm whether the image was generated via ChatGPT, the OpenAI API, or Codex. Currently, this tool only covers images produced by OpenAI's own models, and the company stated it plans to support more verification systems and expand to more content types in the coming months. The tool's limitations are also clearly noted: metadata or watermarks may be intentionally or unintentionally stripped under certain circumstances. If no provenance signal is detected, the tool will not make a definitive conclusion on whether the image was generated by OpenAI. OpenAI wrote directly in the announcement: "No detection method is foolproof, so we take a cautious stance when detection fails."

OpenAI's choice to strengthen its provenance mechanism at this time aligns with its broader AI content safety strategy. The company began adding C2PA content credentials to images generated by DALL·E 3 as early as 2024, and has since extended this practice to the ImageGen and Sora video models. Sora 2 deployed dual visible and invisible provenance identifiers this March; all Sora videos embed C2PA metadata, and most outputs also include a visible moving watermark bearing the creator's name. OpenAI simultaneously uses internal reverse image and audio search tools to track content recirculation. The company has officially joined the C2PA Steering Committee and become a C2PA-compliant generator product member, meaning platforms have a trusted way to read, retain, and pass on the provenance information attached by OpenAI. The C2PA compliance program, which OpenAI announced it is joining concurrently, further requires products to correctly produce and verify C2PA data.

Global regulation on AI content labeling is shifting from encouraging guidance to mandatory requirements, a trend that forms the policy backdrop for OpenAI's update. China's "Artificial Intelligence Generated Synthetic Content Labeling Measures" officially took effect on September 1, 2025, requiring explicit labels on AI-generated content and the embedding of implicit identifiers in file metadata, while also encouraging the addition of digital watermarks. Article 50 of the EU AI Act similarly requires the labeling of AI-generated content. OpenAI's simultaneous deployment of C2PA metadata and SynthID watermarks directly responds to the dual requirements for "explicit identification" and "implicit identification" under different regulatory frameworks. By upgrading from a single-layer metadata approach to a multi-layered provenance architecture, OpenAI is elevating signal durability from a technical capability to a compliance baseline.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com