The Race to
Watermark
Reality

That's how many images were generated through ChatGPT in a single week — the number that landed May 19, 2026 and put OpenAI on notice fast enough to ship a major announcement.

The problem: almost none of them can be reliably verified as AI-generated once they leave ChatGPT and start travelling across the internet.

They get screenshotted. Reposted. Compressed. Cropped. Saved as different formats. By the time a fake headshot of a CEO or a fabricated protest photo lands on your timeline, the trail is cold.

OpenAI's response: adopt Google DeepMind's SynthID watermarking (the exact same system Google already built into Gemini) and pair it with an open standard called C2PA. Then launch openai.com/verify so anyone can check.

But these two technologies work completely differently, fail in completely different ways — and one is orders of magnitude harder to break than the other. Let's go deep.

The dual-layer system OpenAI (and Google) uses combines two fundamentally different approaches. Understanding why they're different is the whole point.

Step through each stage to see exactly what happens to both signals as an AI-generated image travels from ChatGPT through the real internet.

Most explainers skip the actual mechanism. Here's what SynthID's encoder does at a technical level — and why it survives what kills metadata-based approaches.

After an image is generated by the diffusion model, SynthID's encoder network takes the finished image and applies learned perturbations to pixel values. These aren't random noise — they're structured according to a cryptographic key pattern, distributed across the entire image in the frequency domain.

Human vision is most sensitive to low-frequency changes — broad colours, overall composition. SynthID targets the high-frequency zones — fine textures, sharp edges — where modifications of ±1–2 RGB units are statistically detectable but perceptually invisible.

Pixel-level demo

The three panels show: original image, SynthID-watermarked (±1.5 RGB), and amplified 25× (what the neural detector reads). Three modes of the same ten-by-ten patch, laid out side by side so you can compare them at once: normal view, apply watermark, amplify 25× (detector view).

The left panel is the baseline: no watermark. The middle panel carries the watermark and looks identical to human eyes at normal zoom — that is the entire design goal. The right panel is the same watermarked data amplified 25× — bright spots = high correlation with the key. The perturbations were never random, so amplifying them separates the structure out cleanly — a structure the detector already knows how to look for.

SynthID confidence across real-world transformations

Each transformation costs the signal something. Here is what survives what.

The 40% dashed threshold — signals below it return “no watermark found.”

This is what openai.com/verify returns for different scenarios — including images from Gemini Nano and Gemini Omni verified through Google's own tool.

Open any row below for the full verification note the tool returns.

Here's what barely got covered in the OpenAI announcement. Google didn't just invent SynthID — they already deployed it.

Gemini Nano, Gemini Flash, Gemini Pro, and Gemini Omni all have SynthID embedded at generation time. Google built this, shipped it quietly, and had been running it in production while OpenAI was still figuring out their C2PA compliance.

More importantly — Google built the verification side. You can upload any image to Gemini and ask: “Was this made by Google AI?” Gemini checks both the SynthID watermark in the pixels and the C2PA metadata, and gives you a confidence-weighted answer.

Google also expanded SynthID detection into Chrome and Google Search — meaning verification is moving to infrastructure level. Every image you see in your browser could soon carry a small provenance indicator, without you needing to visit a verification tool at all.

OpenAI's choice to adopt SynthID — a competitor's technology — is genuinely unusual. It signals that SynthID is robust enough that building a competing watermarking standard would fragment the ecosystem in a way that's bad for everyone. So they collaborated instead.

The announcements won't always say this clearly. Here's the honest breakdown.

SynthID is technically impressive. C2PA is architecturally sound. But both require buy-in from every link in the chain — generators, editors, platforms, browsers, end users. The image that goes ChatGPT → downloaded → Photoshop → Telegram → Twitter? It arrives clean. No manifest. Watermark degraded. No definitive claim.

If you're building anything in the AI space — or thinking about where the internet is heading — here's what actually matters.

1. C2PA is becoming table stakes. The EU AI Act mandates disclosure of AI-generated content. If you're building any image generation tool, C2PA compliance is the path. OpenAI just did it. Adobe's been there. Expect every major tool to follow.
2. SynthID-equivalent watermarking will become standard. Expect Anthropic, Midjourney, and others to adopt pixel-level watermarking. Open-source models will lag — and that gap is where the most dangerous content will continue to come from.
3. Verification moves into infrastructure. Google putting SynthID detection in Chrome and Search is the signal. This will become a browser-level feature. Every image, a small provenance indicator — without users needing to know what C2PA or SynthID even mean.
4. “Made by a human” becomes a differentiator. As AI images flood the internet, cryptographic proof from a C2PA-signed camera — Leica, Nikon, Sony — becomes a commercial differentiator. Photojournalism, commercial photography, stock images: “human-made” will carry a premium.

Deep technical breakdowns, build walkthroughs, and real AI product tutorials — for founders, operators, and developers who want to stay on the right side of the gap.

Haroon K M · @ closefuture

Visit Haroon K M →