OpenAI is strengthening its approach to identifying AI-generated content, announcing a multi-layered provenance system that combines cryptographic metadata, invisible watermarking, and a public verification tool, though the company acknowledges no single method is foolproof.
What do 1,000 journalists and PR pros know about AI that you don't? They took AI Quick Start, a 1-hour live class from The Media Copilot. 94% satisfaction. Find out how to work smarter with AI in just 60 minutes. Get 20% off with the code AIPRO: https://mediacopilot.ai/
The centerpiece of the update is a new partnership with Google DeepMind to embed SynthID watermarking into images generated through ChatGPT, Codex, and the OpenAI API. SynthID, Google’s invisible watermarking technology, is designed to survive transformations like screenshots and file format changes that can strip standard metadata from content.
The two approaches are meant to work together. C2PA — the open technical standard backed by the Coalition for Content Provenance and Authenticity, a cross-industry group — uses metadata and cryptographic signatures to carry information about where content came from, who created it, and how it was edited. But that metadata can be lost. SynthID provides a backup signal that persists through more transformations.
“C2PA helps content carry detailed context; SynthID helps preserve a signal when metadata does not survive,” OpenAI said in a blog post. “Together, they make provenance more resilient than either layer would be on its own.”
OpenAI also became a C2PA Conforming Generator Product — meaning platforms can now reliably read, preserve, and pass along the provenance information attached to OpenAI-generated content. The company has been adding Content Credentials to images since 2024, when it began with DALL·E 3, later extending to ImageGen and Sora.
A public verification tool
OpenAI is also previewing a public verification tool at openai.com/verify that lets people check whether an uploaded image was generated on ChatGPT, the OpenAI API, or Codex. The tool checks for both Content Credentials and SynthID watermarks.
The approach is deliberately cautious. If no metadata or watermark is detected, the tool will not conclude the image was not generated with OpenAI tools — since provenance signals can be stripped.
“If no metadata or watermark is detected, for example, the tool will not make a definitive conclusion about whether the image was generated with OpenAI tools since provenance signals can in some cases be stripped,” the company said.
At launch, the verification tool is limited to content generated by OpenAI. The company said it aims to support cross-industry verification across platforms in the coming months.
The limits of provenance
The announcement arrives as the question of AI content authentication has become acutely relevant. A Florida Tribune investigation published this week identified a network of AI-generated fake local news sites — complete with fabricated reporters and AI-recycled content — built specifically to manipulate search results. Provenance tools like SynthID and C2PA would not have prevented that scheme, which used content scraped from real outlets and processed through AI. But they could make it harder for the operators of such sites to pass their output as genuinely human-produced.
“No single provenance technique is enough on its own,” OpenAI acknowledged. The company’s answer is the layered approach — shared standards, durable watermarking, and public verification — in the hope of building “a more interoperable provenance ecosystem” over time.
The broader industry has been moving in similar directions. Adobe has embedded Content Credentials in its Firefly-generated images, and Google has been rolling out SynthID across its own products. But adoption remains voluntary, and the tools do nothing to address content that was AI-generated before provenance standards existed — or content deliberately created outside these systems.
