Google Lets Users Remove AI Watermarks; Claude Keeps Them

Google is giving users control over something that's long been a fixed feature of AI-generated content: the visible watermark. In the coming days, Gemini and Flow (Google's video editor) will roll out a setting that lets users turn off the visible watermark on images, videos, and songs generated with Nano Banana, Omni, and Lyria models. Support for the same toggle in Google Search is planned to follow.

Josh Woodward, VP for Gemini at Google, confirmed the change. Users will find the option under Settings > Media Watermark, where they can switch the visible mark on or off. Notably, this only affects the visible marker — Google says the invisible SynthID watermark and C2PA-related metadata will remain intact regardless of the toggle, preserving a machine-readable trail of AI origin even when the human-visible cue disappears.

Alongside the toggle, Google is open-sourcing a new library called Credentio, designed to let developers embed a local validation mechanism directly into their own apps.

A different approach at Anthropic

The move puts Google at odds with Anthropic, which includes a mandatory watermark in text and files generated by Claude specifically to comply with EU regulations. Sources differ on whether other providers will follow Google's opt-out model or Anthropic's opt-in-by-default, compliance-driven approach — and it's unclear whether Anthropic plans to offer a similar removal option for Claude.

This divergence means users and developers moving between AI tools may encounter inconsistent watermarking expectations: visible-by-default and removable in Google's ecosystem, mandatory in Anthropic's — at least in the EU.

What's unresolved

Several open questions remain. The report doesn't specify whether EU regulations actually require visible watermarks to stay visible, which raises the question of how Google's optional approach squares with Anthropic's compliance-driven one. It's also unclear how Credentio's local validation mechanism works or how it interacts with SynthID or C2PA metadata. And while Google insists invisible watermarking remains active, there's no detail on whether downstream platforms or everyday users are actually equipped to detect or verify SynthID or C2PA data — meaning the practical transparency benefit of invisible marking is, for now, unproven at scale.

Why founders should care

Founders building consumer-facing AI products should likely evaluate whether visible watermarking needs to be optional, mandatory, or configurable depending on which regulatory markets they serve — especially the EU, where Claude's example suggests compliance obligations may apply regardless of the underlying model provider.
Teams building on Gemini, Flow, or related models may soon be able to offer less visibly-branded outputs to customers, which could be a product differentiator — but also a potential liability if it becomes harder for end users to distinguish AI-generated content from real content, as flagged in the report's risk assessment.
There's a plausible opportunity for startups to position themselves around transparent, verifiably-watermarked content, particularly for EU-facing products where visible compliance markers may carry more trust value.
The open-sourcing of Credentio suggests an emerging ecosystem of content-provenance tooling; founders may find it more efficient to adopt such libraries rather than build watermarking or validation infrastructure in-house.
Because major providers are now diverging on watermark policy, founders selecting a model provider for content-generation features should probably track these policy differences closely, as they could affect compliance posture and user trust down the line.

For now, the practical effect is straightforward: Google users get more control over how visibly their AI-generated content is marked, while Anthropic keeps the marker fixed in place for EU users. What's less settled is whether this split reflects a genuine regulatory requirement, a business choice, or something in between — and how that ambiguity will play out as more jurisdictions weigh in on AI content labeling.