On May 20, OpenAI quietly announced that it has become a C2PA Conforming Generator Product and, crucially, integrated Google DeepMind's SynthID watermark into its entire image product line, including ChatGPT, Codex, API, DALL·E 3, ImageGen, and Sora.
Note: that's Google DeepMind. OpenAI and Google have been locked in fierce competition over large models for three years, with search and Workspace already direct battlegrounds. Yet on watermarking, they now stand together.
Three-layer provenance
OpenAI's approach is called 'multi-layered provenance,' stacking three layers:
Layer 1: C2PA metadata
The Coalition for Content Provenance and Authenticity is an open standard developed since 2021 by Adobe, Microsoft, BBC, and others. It adds cryptographically signed metadata to images, indicating who created them, when, with what tool, and whether they were modified. The problem: metadata is easily lost—screenshots or format conversions strip it away.
Layer 2: SynthID invisible watermark
Google DeepMind's technology, introduced in 2023, embeds an imperceptible signal directly into image pixels that survives screenshots, compression, and format changes. SynthID watermarks will now be applied to all images generated by OpenAI.
Layer 3: openai.com/verify public verification tool
Anyone can upload an image to check both C2PA metadata and SynthID watermark, receiving a verdict on whether the image was generated by OpenAI systems.
'If neither signal is detected, the tool will not draw a conclusion—because these signals may have been deliberately removed.'
This is important: OpenAI does not claim the system is foolproof but clearly states its limits.
OpenAI acknowledges limitations
The announcement includes candid statements:
- 'No single provenance technique can independently solve the problem'
- The tool currently only verifies content generated by OpenAI's own products
- Content created before this standard was established, or deliberately bypassing the system, cannot be identified
This 'we did it but it's not a silver bullet' stance is more honest than typical PR hype.
Why OpenAI and Google need to work together
Watermarking has an inherent problem: it only works if everyone uses the same system. If only OpenAI images carry SynthID watermarks, images from other models can still be passed off as 'un-generated'—forgers just switch models.
Thus, this must be an industry standard, not a product differentiator. Adobe already uses Content Credentials in Firefly, and Google deploys SynthID across its products. OpenAI's move effectively aligns it with Google-led watermarking, rather than building its own.
The next question is Anthropic, Meta, and xAI. Anthropic's Claude does not generate images, so it is not immediately affected; Meta's Imagine and xAI's Aurora have not yet stated their positions.
Not charity, but regulatory pressure
The real driver is government pressure. Regulators in the US, EU, and Asia have been demanding traceability for AI-generated fake images, political deepfakes, and child sexual abuse material. OpenAI itself faces over 8,000 reports of AI-generated CSAM every six months. The industry's best strategy before regulation lands is to 'build a solution ourselves,' preventing lawmakers from defining technical standards.
Watermarking is not a technology issue; it is governance. OpenAI and Google stand together because their legal teams have run the same numbers.
The remaining questions—when everyone will adopt it, whether it can be forged, and whether regulators will accept it—will be answered in the coming year.
Sources: Advancing content provenance for a safer, more transparent AI ecosystem (OpenAI official blog); OpenAI builds a new system to identify AI-generated images (The Media Copilot); CocoLoop; OpenAI joins C2PA and adds Google SynthID watermarks to provenance stack (Resultsense)