Google Opens the Door to Removable AI Watermarks

Google is giving people more control over visible marks on AI-generated content. In the coming days, users will be able to remove visible watermarks from images, videos, and songs created with several Google models.

The option comes with a clear limit: removing the visible mark will not remove the hidden information that helps identify AI-generated work. Google says the invisible SynthID watermark and C2PA metadata will remain in place, giving the company a way to balance creative control with content tracking.

Visible marks will become optional

The new toggle will work with the Nano Banana, Omni, and Lyria models. All three use the familiar sparkle icon to show that content came from AI. Google VP Josh Woodward said, “A toggle will soon be available for the Nano Banana, Omni and Lyria models, which all use the familiar sparkle icon to denote AI.”

Users will find the control under Settings > Media Watermark, where they can turn the visible watermark on or off. The setting will be available in Gemini and Google’s video editor Flow, while support for Search is coming soon.

Woodward described the decision as a compromise between user choice and safety: “We’re striking a balance here between creative control and safety.” The visible mark can disappear, but the invisible layers remain. “This move will not impact the invisible SynthID watermark and C2PA metadata, as there are are no toggles to shut those off,” he said.

Google is also open sourcing a library called Credentio. Developers can use it to embed validation mechanisms, adding another way to check the origin or handling of AI-generated content.

Watermark rules are spreading

Google’s change arrives as governments and AI companies set different rules for labeling generated media. The European Union recently announced new AI transparency rules that mandate visible watermarks on deepfake content, and South Korea made a similar move earlier this year.

The EU’s transparency rules aim to prevent misinformation, manipulation, fraud, impersonation, and deception. The guidelines say people should know when they are interacting with AI or viewing AI-generated content, making labels part of a wider effort to explain how digital media was created.

The EU AI Act requires AI providers to watermark AI-generated or manipulated outputs from models released after August 2. The law includes a grace period until December 2026. Violations can bring fines of up to 15 million euros or 3 percent of a company’s worldwide annual revenue.

That creates a tricky situation for companies building tools that produce many types of media. A watermark may help people recognize AI content, but it does not guarantee that the content is genuine, safe, or easy to trace.

Invisible systems have limits too

Other AI generators, including OpenAI and Meta, do not apply visible watermarks. Instead, they rely on systems such as SynthID, C2PA, or Content Seal standards. These methods keep information attached to the file or embedded inside it without placing a clear mark in front of viewers.

Anthropic has taken a similar path with Claude. The company announced that it was adding invisible watermarks to text and files generated by Claude to comply with EU regulations. It plans to watermark content processed by its models, including audio, image, text, and video outputs.

Anthropic says all new models will mark AI-generated content from day one. Its watermarks will be embedded in text outputs and include digitally signed provenance metadata where supported. The approach applies to all processed content, even when the EU law does not require a label.

The watermarks will travel with text when people copy and paste it, and they may survive some editing. Screenshotting or editing metadata can destroy them, though, and the marks can be bypassed with ease. Anthropic also says a detected mark shows that content was processed by Claude, but it is not conclusive.

That warning matters because Anthropic’s system may mark content that does not need a label under EU law. The company acknowledged this issue and plans to release a tool for detecting its watermarks, while continuing work on watermarking and detection methods to meet EU demands.

Google’s decision highlights the split between visible labels and hidden verification. A visible watermark gives people an immediate signal, but users can remove it in Gemini and Flow. Invisible marks and metadata preserve another layer of information, yet they can also disappear through actions such as screenshotting or metadata editing.

For users, the practical change is simple: Google-generated images, videos, and songs can soon appear without a visible mark. For developers and regulators, the larger question is whether hidden validation systems can give people enough information about AI content without making creative tools harder to use.

Stay connected via Google News

Follow us for the latest AI News