A lab watermarked pixels instead of stamping a visible logo

The process wall

On 29 August 2023, Google DeepMind announced SynthID, describing the problem it was built to solve in its own post: a visible watermark 'stamped' on an image 'can easily be edited out', for instance by cropping the corner it sits in, while some earlier imperceptible watermarks were 'lost through simple editing techniques like resizing.' The tool was released in beta to a limited set of Vertex AI customers generating images with Imagen, not as a general-purpose add-on for any generator.

What the documents show

The announcement describes two trained models working together, one embedding a watermark into an image's pixels during generation, one scanning a pixel array to detect it, optimised jointly so the mark stays imperceptible while surviving common edits. The company's own account states the watermark 'remains detectable' after 'filters, changing colours and brightness' and lossy JPEG compression, describing this as based on internal testing; it explicitly does not claim the watermark survives 'extreme image manipulations'. Retrieved on 16 September 2026, DeepMind's current SynthID technology page states the tool has since expanded to text, audio and video, and that a 'SynthID Detector' verification portal was launched, dated on that living page only as 'earlier this year', for uploading a file to check for the mark, currently being tested with journalists and media professionals rather than opened generally.

Authorship and limitation

The mechanism is scope-limited by construction: SynthID can only mark an image at the moment DeepMind's own models generate it, so it says nothing about an image made by a different generator, edited heavily after the fact, or captured by a camera. The 2023 announcement itself draws this line, positioning the pixel watermark as complementary to metadata-based approaches like C2PA rather than a replacement: metadata can carry rich information about who made an edit but is lost if stripped, while a pixel watermark survives typical edits but can only ever answer one narrow question, whether this specific lab's tool produced this image.

What survives

What a maker or a platform can rely on is bounded to that one question, and only for images the same company's models generated. Editorially, a 'no watermark detected' result cannot be read as 'not AI-generated', since it may mean the image came from an unrelated tool carrying no SynthID signal at all, only that this particular signal is absent.

• Check which company's model produced an image before treating a watermark-detection result as meaningful.
• Ask whether a stated detection confidence level, not a bare yes or no, was reported.
• Compare a watermarking claim against the announcement's own stated exception for extreme manipulation.

A pixel-level signal that survives ordinary editing is a real, documented advance over a corner logo, but the company's own account keeps it to one narrow claim: whether its models made this image, not whether any image anywhere was AI-made.