If your team has generated a single image or video inside Google’s ad tools this year, you have shipped watermarked creative. You just could not see it.
That watermark is SynthID, Google DeepMind’s family of systems that hides a machine-readable signal inside AI-generated images, video, audio and text. PPC Land has published a detailed explainer on how it works, and there is plenty in it that performance marketers should not skip.
Provenance, not detection
DeepMind draws a hard line between two things marketers often confuse. An AI “detector” guesses whether content looks synthetic, gets it wrong often, and cannot tell you which system produced it. A watermark applied at generation answers the origin question directly, but only for content from participating systems.
The reason invisible marks exist at all: metadata dies. Screenshot a file or upload it to a platform that strips EXIF, and your provenance record is gone. SynthID is designed to survive cropping, compression, filters and screenshots.
How the image mark actually works
SynthID-Image is post-hoc and model-independent. A trained encoder adds a learned perturbation to a finished image; a matching decoder reads it back. One encoder can mark output from every current and future Google model, and the decoder can be retrained without touching the generative model.
The decoder returns two things: a detection score, and a multi-bit payload. The external variant, SynthID-O, encodes 136-bit payloads in 512×512 images. Detection and payload recovery are deliberately separated, so a badly damaged file can still register as watermarked even if the payload is unreadable.
Robustness is benchmarked against 30 transformations across six categories. Calibrated to a 0.1% false positive rate, SynthID-O recorded a 99.72% true positive rate across worst-case transformations, versus 83.37% for the next best method tested. Under combined transformations at maximum strength it held 98.06%, while the strongest rival dropped to 55.96%.
Text uses a completely different mechanism called tournament sampling, published in Nature on 23 October 2024. Across close to 20 million live Gemini responses, watermarked and unwatermarked outputs showed no significant difference in thumbs-up and thumbs-down rates.
Why this is now a compliance line item
Article 50(2) of the EU AI Act obliges providers of generative systems to mark synthetic output in machine-readable form. The Commission’s Code of Practice, finalised on 20 July 2026, requires at least two marking layers for audio, images and video, plus watermarking of free-form text longer than 200 tokens. Obligations became applicable on 2 August 2026, with penalties up to 15 million euros or 3% of worldwide turnover. Google signed the Code on 24 July 2026.
Practically, that means:
- Google embeds SynthID and C2PA markup automatically in any image or video generated in its ad tools, regardless of your visible AI label setting.
- Documentation for the July 2026 label rollout put the disclosure duty on advertisers, not AdSense publishers.
- YouTube treats SynthID and C2PA signals as automatic AI-label triggers, and uploaders cannot adjust those labels.
The performance cost you should plan for
Disclosure is not free. An NYU Stern figure cited in the IAB’s August 2026 framework found AI labelling cut click-through rates by 31.5%. Separate IAB research recorded a 37-point gap between what advertisers assume and what consumers actually feel about AI-generated advertising.
So the smart move is not to hide the label. It is to build creative that survives it: stronger hooks, clearer value props, and testing labelled versus unlabelled variants where you legally can, so you know your real delta rather than guessing.
The limits worth knowing
A negative result proves nothing. Unmarked content may be a photograph, an open-weights generation, or a scrubbed file. Text is the weakest modality, degrading under paraphrasing, copy-paste edits and back-translation, per research published in August 2025.
Images have been attacked too. An independent March 2026 project by Alosh Denny used Fourier analysis on roughly 123,000 Gemini images to build a spectral bypass, reporting a 91% drop in phase coherence at 43 dB PSNR, with the same structure permitting forgery. Google has not published a response, and the claims rest on the author’s own measurements.
Access is the quieter issue: no public decoder, no released weights, no local detection. Verification runs on Google’s servers.
What to do this quarter
Audit which of your creative assets were generated in Google’s tools. Assume they carry the mark. Map disclosure duty across your agency, client and publisher contracts. And build a scale of 100 billion marked images and videos, plus a February 2027 interoperability deadline, into your 2027 creative ops planning.
Source: PPC Land
