Analysis
Google will now let users remove the visible watermark from images generated by its AI tools, according to TechCrunch -- a change that puts Google's approach to AI provenance on a notably different track than Anthropic's, which published details this same week on a text watermark specifically engineered to survive copy-paste and resist removal.
The distinction Google is drawing is between the visible watermark -- the small logo or mark stamped onto an image that any viewer can see -- and the invisible provenance signal built on Google's SynthID technology, which remains embedded in the image regardless of whether a user opts to remove the visible mark. In theory, that preserves forensic detectability: anyone running SynthID's detection tooling can still confirm an image was AI-generated even after the visible watermark is gone. In practice, removing the visible mark makes AI-generated images functionally indistinguishable from unmarked ones to the overwhelming majority of people who will ever see them, since almost no one outside a research lab or a platform's trust-and-safety team actually runs watermark-detection tooling on images they encounter.
“## Two labs, two philosophies, same week The contrast with Anthropic is instructive.”
## Two labs, two philosophies, same week The contrast with Anthropic is instructive. Anthropic's text watermark, driven by EU AI Act compliance requirements, is designed to be invisible to readers by default and specifically built to resist removal even through copy-paste -- transparency-by-default, opt-out not available. Google's image watermark takes the opposite default: visible by default, but removable at user discretion, with only the invisible layer persisting regardless. Neither approach is strictly more or less compliant with current regulation -- the EU's Transparency Code requires machine-detectable marking, which Google's SynthID layer still satisfies even with the visible mark stripped -- but they represent genuinely different bets about what 'meaningful transparency' looks like in practice for typical users who will never check for an invisible signal.
The practical effect is that Google is making it easier for users to share AI-generated images without an obvious visual tell, right as concern about AI-generated misinformation and deepfakes remains a live regulatory and platform-trust issue. Google's defense is that the underlying detectability hasn't actually changed -- only the visible surface layer has -- but that distinction will matter little to platforms, journalists or ordinary users trying to make a fast visual judgment about whether an image is real, since almost none of them will run a SynthID check before sharing.
Watch whether social platforms and news organizations start requiring SynthID or equivalent invisible-watermark verification as a standard step in content moderation, now that a visible mark can no longer be relied on as a fast, low-effort transparency signal.