Can I tell when an image or video has a SynthID watermark just by looking at it?

No. SynthID watermarks are designed to be imperceptible to humans. The signal is embedded in pixel values or audio waveforms in a way that does not visibly or audibly alter the content. To detect it, you need Google's detection tooling, such as uploading the file to Gemini and asking whether it was created by Google AI, or using the SynthID Detector portal available to journalists and researchers.

Does Claude's watermark affect the quality or meaning of the text it writes?

According to Anthropic, the watermark is designed to have no practical impact on the quality, creativity, or readability of Claude's output. The signal is embedded by subtly influencing word-choice probabilities during generation, and the result is text that reads identically to unwatermarked output for any human reader.

What is C2PA, and how is it different from an embedded watermark?

C2PA, which stands for Coalition for Content Provenance and Authenticity, is an open standard that attaches cryptographically signed provenance metadata to a media file, recording who created it, when, and what tools were used. Unlike an embedded watermark, which is woven into the content's pixels or words, C2PA metadata lives in the file container. This makes it publicly readable and tamper-evident, but also fragile: if the metadata is stripped or the file is converted to a format that discards it, the provenance record is lost even though the content looks the same.

Why is the EU AI Act driving AI companies to watermark their content globally, not just in Europe?

The EU AI Act's Article 50 transparency obligations took effect on August 2, 2026, and they apply to any AI system used in the EU regardless of where the company is based. Non-compliance carries fines of up to 15 million euros or 3 percent of global annual turnover. Because building separate regional versions of models is technically complex, companies like Anthropic have found it simpler to apply the marking worldwide. This is an example of what is sometimes called the Brussels effect, where one jurisdiction's regulations become a de facto global standard.

If a watermark can be removed by paraphrasing or image processing, what is the point?

Watermarks are not designed to be tamper-proof against every possible adversary. They are designed to survive accidental or routine modification, such as compression, resizing, copying, and light editing, so that the origin of content can be verified in ordinary circumstances. They also create a legal record: under the EU AI Act, deliberately removing or altering an AI watermark is prohibited, so the act of stripping one carries legal risk even if it is technically possible. Researchers and regulators treat watermarks as one layer of a multi-signal transparency system rather than a complete solution on their own.

Does the absence of a watermark mean content was made by a human?

No. A missing watermark only means the content was not detectably marked by the specific system doing the checking. Content could be unwatermarked because it was made by an AI system that does not yet watermark its output, because the watermark was removed, because it was created before watermarking was deployed, or because it genuinely was made by a human. Watermarks can confirm AI origin when present; they cannot confirm human origin when absent.

AI Policy & Safety

How AI Watermarks Work: Gemini, Claude, and Can You Remove Them?

By The Agent5 founder·August 18, 2026·How we fact-check

Invisible watermarks are now baked into the AI content you create every day. Here is how the technology actually works inside Gemini and Claude, what regulators are demanding, and what the honest limits of the system are.

ShareX LinkedIn WhatsApp

Key takeaways

Google's SynthID and Anthropic's Claude both now embed invisible watermarks in AI-generated content, using statistical signals in text and pixel-level changes in images and video that survive routine editing but are undetectable to human readers or viewers.
Claude's file outputs also carry C2PA provenance metadata, a cryptographically signed record that shows the file was created by Claude and whether it has been tampered with, though this metadata can be lost if a file is reformatted or stripped.
The EU AI Act's Article 50 transparency obligations took effect on August 2, 2026, requiring machine-readable marking of AI-generated synthetic media and threatening fines of up to 15 million euros or 3 percent of global turnover for non-compliance.
Users cannot opt out of watermarking on supported platforms; the marks are applied at the model level. However, motivated adversaries can degrade or remove watermarks through paraphrasing, image regeneration, or other techniques, though deliberately doing so is illegal under EU rules.
Watermarks are a meaningful transparency tool but not a silver bullet: they cannot prove content is human-made when absent, they are not yet interoperable across AI providers, and short or heavily edited content may not carry a reliable signal.

How AI Watermarks Work: Gemini, Claude, and Can You Remove Them?

What an AI Watermark Actually Is

How Google's SynthID Works

How Claude's Watermarking Works

The C2PA Standard: A Shared Language for Provenance

The EU AI Act: Why Regulation Is Driving Deployment

Can You Turn AI Watermarks Off?

What Watermarks Cannot Do

The Agent5 Angle: Thinking in Probabilities About What Comes Next

Sources