OpenAI adds SynthID watermarks and a public checker to spot AI images

What is worse than a fake image going viral? Having no reliable way to prove where it actually came from.

OpenAI is tackling this trust gap by rolling out a three-part content provenance push. The company is now an official C2PA conformant generator, has partnered with Google DeepMind to bake invisible SynthID watermarks directly into images made with ChatGPT, Codex, and its API, and launched an open verification portal at openai.com/verify.

Double-bagging digital origins

For a long time, content credentials relied heavily on C2PA metadata. Think of it like a digital birth certificate attached to a file, carrying cryptographic proof of how and when an image was created.

The catch? Metadata is fragile. Download an image, crop it, take a screenshot, or upload it to a social media feed, and that data frequently gets wiped clean.

That is where Google DeepMind's SynthID comes into play. Instead of attaching a label to the outside of the file, SynthID weaves an invisible digital watermark directly into the image pixels. Even if the metadata gets stripped by transformations or screenshots, the embedded signal survives. C2PA provides the detailed paper trail, while SynthID acts as the persistent fingerprint.

OpenAI has since expanded this layered setup beyond static pictures. The team brought SynthID watermarking to supported audio generated across ChatGPT and the API, opened a dedicated verification API for enterprise workflows, and started extending text provenance to European users to align with the EU AI Act.

The reality of detection

The public verification tool lets anyone drop in an image to scan for OpenAI-originating credentials and SynthID markers. But OpenAI is notably upfront about its boundaries: if the tool spots a signal, it can confirm the image came from its platform; if it finds nothing, it refuses to issue a clean bill of health. In the wild, aggressive manipulation can still scramble watermarks.

Meanwhile, Google is pushing the same standard hard across its ecosystem. Google shared that SynthID has already watermarked over 100 billion images and videos alongside 60,000 years of audio, with its Gemini app verification feature clocking 50 million global uses. Partners like Kakao, ElevenLabs, and Nvidia are adopting it as well.

Why it matters

Seeing OpenAI adopt Google's watermarking tech is the clearest sign yet that the industry cannot afford a format war over AI safety. Front-line frontier models will keep competing fiercely, but when it comes to provenance, having five different incompatible watermarks is practically useless for social networks and newsrooms.

For everyday users, the verification site is worth bookmarking, but do not treat it as an omniscient truth machine. It only knows what OpenAI built. It will not flag rogue open-source deepfakes, nor can it stop determined bad actors from washing out signals using advanced re-rendering tricks.

Still, we think this multi-layered approach makes practical sense. Pairing external metadata with imperceptible pixel signals is far more resilient than either method alone. As platforms like Instagram prepare to display camera-shot credentials and Google embeds detection across Chrome and Search, provenance is finally moving from a PR talking point into everyday web plumbing.