AI watermarking in content is not a new phenomenon although it has taken the industry by storm since Anthropic made their latest announcement.
In a statement released in August, they said future Claude models will contain a watermark in the text it generates. They also added that older models launched before August 2nd 2026 will be retrofitted in line with EU law. We’ll be covering more detail on the EU legislation and why this has come into force further in this guide.
The announcement has inevitably thrown up questions among business owners and marketers who are wondering:
- How can text be watermarked?
- Are there going to be special characters or hidden symbols?
- Will this be visible to the naked eye?
- Will this affect the quality of my content?
- How are Google going to respond?
- How can I remove the AI watermark from content?
Many of these questions are circling the industry, but we’re going to break down what it all means – and simply.
The main message around AI content and how this works in SEO has always remained unchanged:
Google rewards original, high-quality, people-first content.
That has never been disputed and is cited in Google’s own documentation including their Helpful Content Update. This emphasises the importance of helpful, reliable content created to benefit real people rather than to manipulate search engines.
Let’s start by exploring what the AI watermark in content is.
What Is the AI Watermark In Content?
The AI watermark in content is not what you might initially think. It differs from watermarks that you’re traditionally used to such as those often found on stock imagery and design mock ups.
The watermark is also not additional letters, characters, or symbols as all of these things would be too easy to remove. These things are visible to the human eye so there would be nothing stopping people from removing them to rid content of its output method.
Instead, the system used by Anthropic for Claude is inspired by SynthID which has existed in Google’s own AI platform, Gemini, since 2024.
Instead of using symbols or characters, this uses a clever pattern-based method that changes the randomness of which words precede other words in response to an AI prompt.
For example, if you ask AI to complete the sentence ‘The weather today was…’
It might complete this with any of the following:
- Sunny
- Hot
- Humid
- Close
- Warm
- Bright
All of which make perfect sense in this context. Any one of these words would not affect the quality or meaning of the sentence, which is the most important thing. AI simply chooses the most suitable word based on the preceding text and it is completely randomised.
Imagine rolling a die. In this scenario:
- 1 = Sunny
- 2 = Hot
- 3 = Humid
- 4 = Close
- 5 = Warm
- 6 = Bright
AI is rolling through each of these options when deciding its next word.
Watermarking uses choices like this across many times over a piece of text to create a pattern. So instead of choosing the next word completely randomly, it favours particular words over others. For instance, using the example above, it might favour ‘sunny’ over ‘humid’.
If this happens a lot over a piece of text it suggests Claude has been involved in some way.
One choice is not enough though, so watermark detection is easier over a large piece of content as a detection tool can amalgamate all of these choices.
It’s like flipping a coin – if you only get heads once, you wouldn’t think anything of it. However, if you flipped the coin 100 times and got heads 80 times it would suggest something unusual was happening.
How Does A Detection Tool Spot the Pattern?
The watermarking system has a key which is unknown to the general user. That key influences which words Claude is more likely to pick out of the suitable collection (i.e. humid or hot).
The detection tool has access to the same key, so it can analyse the finished text and see whether the word choices consistently match the pattern that the key would produce.
The same concept applies to Gemini, created by Google, as it also uses SynthID. For a detection tool to identify whether content was written by Gemini, it would need access to the key used by Gemini. This is because Claude and Gemini do not use the same key, even though their watermarking systems are based on the same underlying concept.
Essentially, each key creates its own pattern so a detection tool looking for Claude’s watermark would need Claude’s key, while one looking for Gemini’s watermark would need the key for Gemini’s pattern.
Does AI Watermarking In Content Have Limitations?
There are limitations, even to this sophisticated pattern-based approach. Whilst it’s more clever than using symbols or characters which are easily removed, marketers can still get around the feature if they edit the copy heavily enough.
The more human input that a piece of AI-generated content has, the less likely it is to be watermarked. That’s why only using AI for one or two sentences or a short email is a lot less likely to be flagged than using it for long-form.
This includes things like essays, blog posts, and website copy, as there are numerous opportunities for AI to make word choices that contribute to the watermark. This makes it easier for detection tools to use the pattern-based system by spotting these choices across a larger piece of text.
However, if a human has edited or changed the AI content enough, the question begs has it really still been written by AI at this point? If the tool is being used as an assistance, but the writer still edits and influences the majority of the output, you could argue that AI has no longer written it. It has been used in the right way – to aid the original creator.
Furthermore factual prompts with a clear answer are also less likely to be watermarked as there are less available, correct answers.
For example, if you asked AI ‘what is the capital of France?’
Paris is the correct answer.
There are no other options for the platform to scroll though (unlike the example of the hot weather) so everyone gets the same answer.
Another scenario where AI watermarketing falls short is when text is translated into another language. This is because the wording and structure of the sentence can completely change, so the patterns can become undetectable to an analysis tool.
Is Using AI Content Bad for SEO?
Google does not penalise AI content by default.
However it will penalise content that is unoriginal, repetitive, and generic. This leads to a bad user experience (high bounce rates, low session duration, keyword dilution), all of which are negative signals to Google.
It also affects E-E-A-T (experience, expertise, authority, trust) which is a credibility score used by Google to determine which content ranks highly.
However by its very nature, AI content that has had little-to-no human input lacks expertise, experience, authority, and trust as it has been written by the platform and not by a person.
Furthermore there is a real issue with everyone posting the same content that has been written heavily using AI. The way AI models work is by learning from information published online and then gathering the best responses for user prompts.
However:
If every piece of content that lives in Google has been written by AI, what fresh material does the model have to learn from? This leads to a cycle where AI is increasingly learning from content that has itself been generated by AI rather than a human.
That’s why Google puts so much emphasis on originality, otherwise everyone ends up with the same work and there is no real differentiation between the quality of content.
What Are the New EU Guidelines for AI Watermarking?
The new EU AI Guidelines aim to provide more transparency as companies need to declare when content has been heavily generated using AI.
This relates to major platforms such as OpenAI, Anthropic, and Google as users rely on these tools every day. Whilst there is no obvious sign to the human eye, it will allow detection tools to know when content has been influenced in some way by AI.
This is the important point to highlight – even if content has not been written by AI it shows that the content has ‘passed through’ AI at some point in its creation.
The act comes after growing concern around deepfakes and misinformation as users are taking what they see online and running with it as being the truth. However they have a right to know if a piece of text or image was created or altered by AI as this gives them the opportunity to make their own decisions.
The guidance will affect all AI labs including any company that offers the functionality in some way whether that’s an AI voice tool or AI image generator.
Will Users Be Affected By AI Watermarking?
As the EU AI act affects companies (such as OpenAI and Anthropic) individuals should not be affected. Furthermore the watermark will not change who owns the content or who is legally responsible for it.
According to Anthropic’s website, the watermark will help test whether Claude (or another form of AI) was involved in the process of its production. This is one of the limitations as it does not mean the content was entirely written by AI.
For users, the good news is that watermarking shouldn’t change their experience, as the watermark itself is invisible. As Claude’s system uses SynthID, there are no extra tokens being added to the response.
Tokens are the small pieces of text that an AI model reads and generates (such as words and punctuation). As the watermark is created through patterns in the words rather than anything additional, nothing extra needs to be added to the text. This means the length and quality of its responses stays the same without a lag.
