Google Hid AI Watermarks — Right After Anthropic Exposed Them

By Imran Khan (AI Tech Safar)

Three days before this announcement, Anthropic made it harder to hide that you used AI. Now Google is making it easier. On August 14, 2026, Google confirmed that users can now switch off the visible watermark on images, videos, and songs generated with Gemini — the same week Claude users were publicly furious that Anthropic's new watermarks would expose them for using AI at work and in school. Same industry, same eight-day window, two AI giants moving in exactly opposite directions.

Google's version comes with a catch that matters more than the headline: turning off the visible mark doesn't touch the invisible one.

What's Changing: Visible vs. Invisible Watermarks

Watermark Type	Can You Turn It Off?	What It Does
Visible Watermark	Yes, as of this update	The small on-image logo/mark you can see, removable via Settings > Media Watermark
SynthID (Invisible)	No, stays on regardless	Embedded signal detectable via Gemini or Search, unaffected by the new toggle
C2PA Metadata	No, stays on regardless	Digitally signed provenance data attached to the file itself

What Google Actually Announced

Josh Woodward, Google's VP for Gemini, confirmed on X that the toggle applies to content generated with Nano Banana, Omni, and Lyria — covering images, video, and music respectively. The setting is rolling out first inside Gemini and Google's video editor Flow, with support for Search coming later. Once live, it's a simple switch: Settings > Media Watermark, on or off. Alongside the toggle, Google also open-sourced a new library called Credentio, letting developers build local C2PA validation directly into their own apps — a move aimed at third-party tools, not just Google's.

Why This Is the Opposite of What Anthropic Just Did

The contrast only makes sense with the full timeline. On August 11, Anthropic confirmed it would start embedding invisible watermarks into text generated by Claude, specifically to comply with the EU AI Act's Transparency Code, which took effect August 2, 2026. Unlike a visible logo, Anthropic's text watermark travels with copied and pasted content and can survive some editing — meaning someone submitting Claude-written work for a job or a class assignment could plausibly get caught by a detector later, even after editing the text. That announcement triggered visible backlash on Reddit, with some users calling it an invasion of privacy and arguing Claude shouldn't "claim credit" for work they directed and refined themselves.

Google's move points the opposite way for a specific reason: visible watermarks are a genuine usability problem for anyone doing professional creative work — a marketer generating a product image, a filmmaker testing a video concept — where a logo stamped across the output makes it unusable without extra editing. Google's actual position isn't "hide that this is AI," though; it's "separate the annoying visible marker from the permanent identification layer." SynthID and C2PA metadata stay in place either way, meaning the content is still detectably AI-generated to anyone who checks — it just won't announce itself visually to a casual viewer anymore.

The Real Tension Both Companies Are Navigating

Strip away the surface-level contrast, and both companies are actually managing the same underlying problem from different angles: regulators and platforms increasingly require AI content to be identifiable, while regular users increasingly want AI content to look and function like anything else they create. Anthropic is under direct EU legal obligation with real financial teeth — fines up to €15 million or 3% of global annual turnover for non-compliance — so it has far less room to soften its approach, even against user backlash. Google faces no equivalent legal pressure specifically forcing visible marks to stay on, so it can afford to prioritize creative usability while quietly keeping the compliance layer (SynthID, C2PA) intact underneath. Same regulatory era, same detection technology stack, just very different amounts of legal room to maneuver.

💡 AI Tech Safar Insight
The detail worth sitting with here is that neither company actually removed AI detectability — they both just changed who sees it by default. Google hid the marker from casual viewers while keeping it fully detectable to anyone who runs a check through Gemini or Search. Anthropic did roughly the inverse: made the marker invisible to casual viewers from the start, while making it harder to strip out through editing. Functionally, both companies landed on nearly the same place — AI content stays identifiable to anyone who looks for it, and invisible to anyone who doesn't. The visible-watermark toggle is really a UX decision dressed up as a policy shift; the actual accountability layer, SynthID and C2PA, isn't going anywhere for either company.

Frequently Asked Questions (FAQs)

Q1: If I turn off Gemini's visible watermark, can people still tell an image is AI-generated?
Yes. The invisible SynthID watermark and C2PA metadata remain active regardless of the toggle, and can still be checked through Gemini or Google Search.

Q2: Does this watermark toggle apply to text generated by Gemini too?
No, this specific update covers images, video, and music generated with Nano Banana, Omni, and Lyria — not text output, which is a separate area Anthropic's recent changes specifically targeted for Claude.

Q3: Why is Anthropic's approach to watermarking so different from Google's?
Anthropic's watermarking is driven by direct EU AI Act compliance requirements with significant financial penalties for non-compliance, giving it far less flexibility than Google, which faces no equivalent legal mandate forcing visible marks to stay on.

Q4: What is Credentio, the library Google also released?
Credentio is an open-source C2PA validation library Google released alongside this update, letting third-party developers build local content-authenticity checks directly into their own apps.

What Do You Think?
Does hiding the visible watermark while keeping invisible detection intact strike the right balance, or does it just make AI content easier to pass off as human-made at a glance? Share your take in the comments below!

Related Reading:

• Gemini Omni Tutorial: Step-by-Step Guide to Google's Free AI Video Generator — Our full walkthrough of the same Omni model now getting the watermark toggle.
• How to Use Claude Code: A Complete Beginner's Guide (2026) — Background on Claude, the model now carrying Anthropic's new text watermarks.

Source: Reporting based on TechCrunch, with additional context on Anthropic's watermarking from TechCrunch's original report and Euronews.