AI Watermarking

An AI watermark is a hidden signal that the system generating content builds into it as it is created, so the company behind that system can later confirm where the content came from. That makes it a different kind of evidence from an AI detector, which reads finished text and guesses from style, and whose output is unreliable for any individual page. No platform has said a watermark affects ranking or AI citation; the reason to understand watermarking is that regulatory pressure is now bringing it to text, images and video.

What is AI watermarking?

AI watermarking embeds a machine-readable signal into generated output so that its origin can later be established. The signal is imperceptible to a reader and, in the designs currently shipping, does not change what the content says.

This matters because the alternative is so weak. The two most-quoted figures about how much of the web is AI-written both come from detectors: Ahrefs sampled 900,000 newly published pages in April 2025 and classified 74.2% as containing at least some AI-generated content,¹ and Graphite estimated 42.7% of the references ChatGPT cited in June 2026 were themselves AI-generated.² Both studies are careful, and both rest on an instrument that misclassifies human writing as machine-written often enough that no individual result from it should be trusted.

Watermarking answers a narrower question more reliably. It cannot tell you whether a page is any good, and as the next sections show, for text it cannot yet tell you much at all unless you work at the company that applied the mark.

How does SynthID work?

SynthID is Google DeepMind’s watermarking system, and it covers images, video, audio and text.

For text, it works by adjusting the probability scores the model uses to choose each word. Every candidate word is assigned a score based on how likely it is to come next, and SynthID modulates those scores so a pattern is embedded across the output without degrading quality.³ Nothing is added to the text and nothing is removed, which is why the mark survives copying and pasting in a way a visible label never would.

Scale is no longer the constraint. Google reports watermarking over 180 billion images and videos and 240,000 years of audio.⁴ OpenAI, Kakao, ElevenLabs and NVIDIA bring SynthID to their own generated media,⁵ and Google names Apple as a partner “soon”.⁴

Verification was the constraint until October 2026, and for text it still is. The SynthID Detector, previously a waitlist for journalists and media professionals, opened to everyone worldwide in English on 7 October 2026: upload an image, video or audio file at synthid.com and it reports whether the file carries a SynthID mark.⁴ It does not check text, and it only finds marks from Google’s own tools and its partners’, so a clean result does not show that a person made the file. The other route is to ask “Is this made with AI?” through Lens, AI Mode, Circle to Search or Gemini in Chrome. Google expanded that verification to Search in May 2026, with Chrome following,⁵ and says its built-in checks now handle over 1 million requests a day.⁴

What are C2PA Content Credentials?

C2PA Content Credentials take the other approach. Rather than hiding a signal inside the content, they attach cryptographically signed metadata to the file recording who created it and how it has been edited since.

The two are complementary rather than competing. A watermark travels with the pixels or the words and survives stripping the metadata; Content Credentials carry a fuller history but can be removed with the metadata. Google’s C2PA verification reached the Gemini app in May 2026 and was scheduled for Search and Chrome in the months after.⁵ Anthropic attaches C2PA provenance metadata to .svg, .png and .jpg files Claude generates,⁶ and its free Claude Content Checker reports whether a file carries one.⁷

For a publisher, Content Credentials are the more actionable of the two, because they are something you can apply to your own images rather than something a model applies to its output.

Can you detect AI-generated text?

Not reliably, and not yet through watermarks either.

Anthropic marks all text from supported Claude models with an imperceptible watermark, worldwide rather than only in the EU.⁷ OpenAI announced a narrower approach on 5 October 2026, framed as its response to the EU AI Act: its text watermark, textGrain, is being added to ChatGPT and Codex output in the EU only, and is opt-in and off by default for API customers everywhere.⁸ Two limits, stated by both companies, set the ceiling on what any of this proves.

A mark does not prove the content was written by AI. Anthropic states a detected mark means content “may have been processed by Claude”, which covers proofreading, translating or summarising something a person wrote.⁷ OpenAI says a mark “does not measure human contribution”.⁸

The absence of a mark proves nothing at all. Anthropic lists text that has been heavily edited, paraphrased, translated or mixed into other writing, very short passages, and output from a model generated before marking was supported for it; for files, format conversion or re-saving strips the metadata.⁷ OpenAI adds other companies’ tools to the list.⁸

Light editing weakens OpenAI’s text mark quickly. OpenAI published its own detection rates, measured with the detector set to wrongly flag about 1 in 100 unmarked passages. It found the mark in about 95% of 400-token passages (roughly 300 words) and about 80% of 200-token ones, for psychology answers, and substantially fewer for maths, where word choice is constrained. Swapping 10% of the words in a 400-token passage for synonyms cut detection from about 92% to 66%; swapping 25% cut it to 17%.⁸ Anthropic publishes no detection rates; it says light editing “probably won’t remove the watermark completely”, while a complete rewrite where every word is replaced will.⁶

None of the three offers the public a way to check text. Anthropic’s detection API is in private preview for organisations eligible under EU law, such as regulators, law enforcement, media, fact-checkers and independent researchers, and for enterprises with their own compliance obligations;⁶ OpenAI’s text detector is open only to approved researchers and expert organisations, by application;⁸ and Google’s public SynthID Detector covers images, video and audio only.⁴ So provenance marking will eventually make the origin of text more knowable than a detector ever could, and today it changes nothing you can check.

It will also never produce a clean split between human and machine writing, because most content is neither. Ahrefs’ own breakdown makes the point: of those 900,000 pages, only 2.5% were classified as pure AI and 25.8% as pure human, with the large majority a mixture.¹

Do you have to label AI-generated content?

If you publish in or into the EU, there is a rule, and there is an exemption that most editorial workflows already satisfy.

Article 50 of the EU AI Act, which applies from 2 August 2026, places an obligation on deployers, not only on the companies building the models. A deployer publishing AI-generated text to inform the public on matters of public interest must disclose that the content was artificially generated. The exemption is the part worth reading closely, because it is written in the language of ordinary publishing:⁹

where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content

A human review step with someone accountable for what goes out removes the disclosure requirement. That is the same workflow AI content risks and E-E-A-T already argue for on quality grounds, now with a second reason attached.

Whether it reaches a UK publisher is genuinely unsettled. Article 2 applies the Regulation to deployers “that have their place of establishment or are located within the Union”, which excludes a UK business, but extends it to third-country deployers “where the output produced by the AI system is used in the Union”.¹⁰ Whether a UK website read by EU visitors counts as output used in the Union is an interpretation question that has not been tested. Treat it as a reason to keep the review layer documented rather than as a settled obligation, and take legal advice if the answer would change what you publish.

Product images in Google Shopping have their own rule. Google Merchant Center requires AI-generated product images to carry the IPTC DigitalSourceType value trainedAlgorithmicMedia, the standard metadata label for AI-generated media embedded in the image file, and tells merchants not to remove it.¹¹ The rule covers the main, additional and lifestyle image fields. If your image pipeline compresses or resizes files, check the tag survives it.

Does AI watermarking affect search rankings?

No platform has said so.

Google’s announcement of SynthID and C2PA verification in Search describes a capability for users checking an image, not a signal in ranking, and says nothing about search results being affected.⁵ Nothing in Google’s guidance on AI-generated content ties provenance marks to ranking either. The position across generative engine optimisation applies unchanged: content is assessed on quality, usefulness and evidence, not on how it was produced.

The plausible near-term consequence is a reader-facing one rather than an algorithmic one. Verification is now open to anyone holding the file, inside Search and Chrome or through the public detector, so an image on your site that returns “made with AI” is answering a question in front of the reader, at the moment they are deciding whether to trust you.

What to do now

For most publishers, three things, none of them expensive.

1. Keep the human review layer, and record who owns it. A named author, a documented editorial process and someone holding editorial responsibility is what the Article 50 exemption describes. It is also what earns citations.
2. Disclose AI-generated imagery in the caption. A one-line credit costs nothing, matches what verification tools will report anyway, and reads better coming from you than from a right-click.
3. Do not buy a detector to police your own writers. Detectors misclassify human writing, and a false positive against a real author is a worse outcome than the problem it was bought to solve.

What not to do: nothing here justifies engineering work. Beyond keeping the metadata tag on AI-generated product images, there is no markup to add, no header to set, and no confirmed benefit to chase.

Frequently asked questions

Can Google tell if my content was written by AI?

Not from a watermark in most cases. Text marks are applied by the model provider, Google’s public SynthID Detector checks images, video and audio but not text,⁴ OpenAI’s text detector is limited to approved researchers,⁸ and Anthropic’s is a private preview for regulators, media, researchers and similar organisations, a list that does not name search engines.⁶ Google’s stated position is that it rewards helpful, reliable content regardless of how it is produced, and penalises content made primarily to manipulate rankings, also regardless of how it is produced.

Does a SynthID watermark survive editing?

For images, video and audio Google says the mark survives cropping, filters, frame-rate changes, lossy compression, added noise and speed changes.³ For SynthID text Google does not document it. OpenAI does for its own text watermark, and the figures show how fragile a text mark is: replacing a quarter of the words with synonyms cut detection to 17%.⁸ Anthropic says light editing probably won’t remove its text mark completely, and a complete rewrite will.⁶

Should I add Content Credentials to my images?

It is reasonable for original photography and for AI-generated illustrations where you want the provenance to travel with the file. It has no confirmed search benefit. Treat it as a transparency measure rather than an SEO one.

Is a watermark the same as an AI content detector?

No, and the difference is the point. A detector inspects finished text and infers from style, which is unreliable per page. A watermark is inserted deliberately by the system that generated the content. The detector is a guess; the watermark is a record, and for images, video and audio from Google and its partners, anyone can now read it.

Footnotes

1. What percentage of new content is AI-generated? — Ahrefs ↩ ↩²

2. AI search collapse: AI responses collapse when AI retrieves its own generations — Graphite, Gregory Druck and Ethan Smith, June 2026; academic version RAG Collapse, arXiv 2608.22118, 22 August 2026. Collapse figures from Graphite’s own simulations (three models, 1,019 questions, 1,528 runs); the AI-generated-reference shares are live ChatGPT references from January and June 2026, classified with the GPTZero detector. ↩

3. SynthID — Google DeepMind ↩ ↩²

4. We’re making it easier to identify AI-generated content globally — Google, Pushmeet Kohli, VP, Science and Strategic Initiatives, Google DeepMind. Published 7 October 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

5. Making it easier to understand how content was created and edited — Google. Published 19 May 2026. ↩ ↩² ↩³ ↩⁴

6. How Claude’s text watermarking works — Anthropic. Published 14 August 2026, updated 1 September 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

7. How Claude marks AI-generated content — Anthropic ↩ ↩² ↩³ ↩⁴

8. Our approach to EU text provenance rules — OpenAI. Published 5 October 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

9. Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems — EU Artificial Intelligence Act. Chapter IV applies from 2 August 2026. ↩

10. Article 2: Scope — EU Artificial Intelligence Act ↩

11. Preserving metadata tags for AI-generated images in Merchant Center — Google ↩