Google Gemini gør det nu muligt at slå synlige AI-vandmærker fra – med 3 undtagelser

On Friday, August 14, 2026, Google announced a new "Media Watermark" setting in Gemini and Flow that lets users toggle the visible AI watermark on or off — a reversal of the previous policy that always forced it on . Here is a full breakdown of the announcement and the surrounding landscape.

What Google changed

• The visible watermark — a small sparkle (spark) icon in the bottom-right corner of AI-generated content — is now optional .
• The setting is found at Settings > Media Watermark inside the Gemini app and on gemini.google.com .
• When toggled off, the visible mark disappears from new generations, but the file still carries invisible SynthID watermarks and C2PA provenance metadata .

Affected tools, models, and content types

Content type	Tool / Model	Visible watermark toggle applies?
Images	Nano Banana (Gemini)	Yes
Videos	Omni (Gemini) / Flow (Google's AI video generator)	Yes
Music / songs	Lyria (Gemini's music generator)	Yes
Search results	Google Search (Gemini-integrated)	Expected next, not yet available

• The toggle covers content generated in both the Gemini app and Flow .
• Text output is not part of this visible-toggle change; Google's SynthID for text remains invisible and is not user-controlled .

Where the toggle is mandatory or restricted

• In countries where visible watermarks are required by law, the toggle is unavailable — users cannot turn off the visible mark .
• Google did not specify which countries, but the restriction is designed to comply with local regulations (likely including EU jurisdictions with transparency mandates).
• Outside those jurisdictions, the toggle is fully user-controlled.

Invisible markers that remain

Two forensic layers stay embedded in every file regardless of the toggle:

1. SynthID (by Google DeepMind) — an imperceptible digital watermark embedded in images, audio, video, and text at generation time. It is machine-detectable but invisible to humans and survives compression and format changes . By May 2026, more than 10 billion pieces of content had been SynthID-watermarked .
2. C2PA metadata — open-standard provenance metadata (Content Credentials) attached to the file. Google Search and Chrome can surface this to let users verify AI origin .

Notably, reverse-SynthID tools have emerged — a proof-of-concept (reverse-SynthID) claims 91% removal of the SynthID watermark on Nano Banana images via FFT-based spectral analysis , and diffusion-based regeneration has also been shown to evade detection . Google has not publicly patched these.

Comparison with competitors

Company	Visible watermark	Invisible watermark	Text watermarking	User opt-out
Google (Gemini)	Toggle on/off (optional)	SynthID (always on)	SynthID for text (always on)	Yes, for visible only
Anthropic (Claude)	None	Invisible text watermark + C2PA for files (always on since Aug 2)	Yes, embedded in token sampling	No opt-out at all
OpenAI (DALL-E / GPT-4o)	No visible overlay by default	SynthID + C2PA since May 2026	Not in text output (planned)	No toggle; invisible always on

Key competitive differences:

• Anthropic went the opposite direction from Google — it activated invisible text watermarking on all Claude models released on or after August 2, 2026, globally and with no opt-out . Older models (Opus 5, Sonnet 5, Haiku 4.5, Fable 5) are not yet watermarked under an EU transition period . The move is driven by the EU AI Act (Article 50), which mandates transparency labels on AI outputs . Anthropic's watermark is a machine-readable signal woven into token selection during text generation .
• OpenAI uses both C2PA metadata (since Feb 2024) and SynthID (since May 2026) on DALL-E 3 images, plus C2PA on Sora video . OpenAI's images carry no visible overlay by default — unlike Google's previous always-on visible mark. OpenAI started watermarking supported audio via SynthID in July 2026 .
• The EU AI Act is the common regulatory driver: all "most advanced" generative AI providers must label outputs as AI-generated, with compliance deadlines starting August 2, 2026, and a transition period until December 2, 2026 for existing models already on the market .

Related industry developments

AI music labeling: OpenAI announced on July 31, 2026, that it expanded SynthID watermarking to supported audio generated through ChatGPT and the OpenAI API, and its public verification tool now handles audio files alongside images . Google's Lyria (within Gemini) now also falls under the visible-toggle change, with SynthID remaining as the invisible layer for audio . The EU AI Act Article 50 equally applies to AI-generated music, requiring labeling .

Google's Gemini assistant strategy: Google has been integrating SynthID and C2PA deeper into the broader Gemini ecosystem. At Google I/O 2026 (May), the company announced that C2PA Content Credentials verification would be accessible in the Gemini app and that Google Search and Chrome would surface AI provenance information to users . The visible-watermark toggle is a direct user-experience concession following months of complaints that the always-on sparkle icon made Gemini-generated content less usable for professional and commercial purposes .