Watermark Forensics: How Sightova Detects Invisible AI Watermarks Hidden in Your Images

Google and OpenAI hide invisible watermarks in the pixels of the images their models generate — not in metadata, in the image itself. Every Sightova scan checks for them.

Sightova Research6 min read
Contents
  1. 01The Problem with Metadata-Based Provenance
  2. 02How Invisible Watermarks Work
  3. 03What We Built
  4. 04What You See in Results
  5. 05Why This Matters for Detection
  6. 06The Limitations We're Transparent About
  7. 07The Bigger Picture

When Google generates an image through Gemini, it doesn't just create pixels. It embeds an invisible watermark, called SynthID, directly in the image. OpenAI does something similar for images from its GPT-Image models. These watermarks are designed to survive many of the everyday changes that strip traditional metadata, such as re-saving, compression and sharing, although heavy edits can weaken them.

Today, we're adding watermark forensics to every Sightova scan.

The Problem with Metadata-Based Provenance

The digital provenance ecosystem has long relied on metadata. Standards like C2PA embed signed credentials into file headers. EXIF data records camera information. XMP stores processing history.

The fundamental weakness is obvious: metadata lives in the file container, not in the image itself. Save the image to a different format and the metadata is gone. Send it through Telegram, WhatsApp, or Instagram — stripped. Take a screenshot — completely lost.

This means metadata-based provenance only works in controlled chains of custody. The moment an image leaves that chain — which happens almost immediately on the open internet — the provenance data disappears.

How Invisible Watermarks Work

Invisible watermarks take a fundamentally different approach. Instead of storing information alongside the image, the generator changes the pixel values themselves in a pattern that people can't see but a matching detector can.

Because the watermark is part of the image, changing the file format doesn't remove it, and clearing the metadata doesn't touch it. It is designed to persist through compression and format conversion. Strong resizing, tight cropping or deliberate attacks can still weaken it below the point where it can be detected, which is why a watermark check should never be the only check.

What We Built

Every Sightova scan runs two watermark detectors on the image's pixels:

Google SynthID — a decoder trained to recognise the watermark Google adds to images made with its Gemini image models, including Nano Banana.

OpenAI — a detector for the watermark OpenAI adds to images made with its GPT-Image models.

Both look for the signal in the pixels only, so they work on files whose metadata has been stripped. Each returns a confidence score, and a watermark counts as found only when that score clears the detector's threshold. You can also run the same check on its own, for free, with our SynthID decoder.

What You See in Results

The watermark check is the first step in every scan result:

  • The verification timeline shows the watermark result first: found, with the decoder's confidence, or clear.
  • The verdict names the source when a watermark is found, for example that the image was generated by Google Gemini or by OpenAI's ChatGPT.
  • The AI Evidence section lists the generator and notes that it was identified by its invisible watermark.

This sits alongside the rest of the result: the AI probability, the content credentials, and the manipulation heatmap. It's another independent signal in the forensic picture.

Why This Matters for Detection

Watermark forensics and AI detection are complementary, not redundant. They answer different questions through different methods:

AI detection (our detection model) identifies synthetic images by analyzing visual patterns and statistical artifacts left by the generation process. It works on any AI-generated image regardless of whether it was watermarked.

Watermark forensics identifies images that a specific provider has marked. It confirms origin — this image came out of Google's or OpenAI's generation pipeline — with an independent method.

When both signals agree — our detection model says AI-generated, and the watermark check finds a watermark — you have two independent confirmations from fundamentally different analytical approaches. That's a level of certainty neither method provides alone.

The Limitations We're Transparent About

Watermarks only exist where a provider adds them. Google and OpenAI do; most open-source image models don't. A missing watermark therefore proves nothing on its own, which is why every scan also runs the detection model.

A watermark can also be weakened by heavy resizing, cropping or deliberate attacks, and the decoder reports a confidence score rather than a certainty. Our testing methodology describes how we evaluate the watermark check. As more providers add watermarks to their images, we'll add detectors for them.

The Bigger Picture

The AI provenance landscape is bifurcating into two approaches: metadata standards (C2PA, IPTC) that require cooperation from every link in the content chain, and invisible watermarks that travel with the pixels.

Both have roles to play. But for detection purposes — for answering the question "was this image generated by AI?" — invisible watermarks offer something metadata cannot: they are far harder to lose by accident than an EXIF tag.

By running watermark forensics alongside our detection model, metadata analysis and the manipulation heatmap, Sightova gives you a fuller forensic picture of any image you need to verify.

Try it today — watermark forensics is included in every scan at sightova.com.