AI Watermark
AI Watermark
An AI watermark is an invisible, machine-readable signal embedded in content produced by an AI model, so the content can later be identified as machine-generated.
An AI watermark is an invisible, machine-readable signal embedded in content produced by an AI model, so the content can later be identified as machine-generated.
Anthropic began watermarking Claude’s text output in August 2026, making this a live issue for anyone who writes with AI.
Anthropic began watermarking Claude’s text output in August 2026, making this a live issue for anyone who writes with AI.
What an AI watermark is
What an AI watermark is
A signal carried inside the content itself, not a visible label.
A signal carried inside the content itself, not a visible label.
Invisible to readers, and does not change how text reads
Survives copying and pasting into other applications
Applied by the AI provider at the model level
For files, providers use the C2PA provenance standard
Invisible to readers, and does not change how text reads
Survives copying and pasting into other applications
Applied by the AI provider at the model level
For files, providers use the C2PA provenance standard
You cannot see it, and you cannot strip it by pasting into a plain text editor.
You cannot see it, and you cannot strip it by pasting into a plain text editor.
Why AI watermarking started
Why AI watermarking started
The EU AI Act. Article 50(2) requires providers of generative AI to mark their output so it can be identified as machine-made. The accompanying Code of Practice on Transparency took effect on 2 August 2026.
The EU AI Act. Article 50(2) requires providers of generative AI to mark their output so it can be identified as machine-made. The accompanying Code of Practice on Transparency took effect on 2 August 2026.
Anthropic signed the Code and applied watermarking globally rather than only for EU users. The broader pressure is the volume of unlabelled machine-written content now circulating, what the press calls AI slop.
Anthropic signed the Code and applied watermarking globally rather than only for EU users. The broader pressure is the volume of unlabelled machine-written content now circulating, what the press calls AI slop.
Which AI models use watermarks
Which AI models use watermarks
As of August 2026:
As of August 2026:
Claude (Anthropic). Text watermarking on models released after 2 August 2026, with older models to follow. Applies across the Claude API, Claude, Claude Code, Claude Cowork, and Claude Tag.
Other major providers. No equivalent text watermarking announced at the time of writing.
Claude (Anthropic). Text watermarking on models released after 2 August 2026, with older models to follow. Applies across the Claude API, Claude, Claude Code, Claude Cowork, and Claude Tag.
Other major providers. No equivalent text watermarking announced at the time of writing.
This will change. We update this page as providers announce. More on Claude’s watermark.
This will change. We update this page as providers announce. More on Claude’s watermark.
Watermarks versus AI detectors
Watermarks versus AI detectors
They are different mechanisms and are constantly confused.
They are different mechanisms and are constantly confused.
A detector analyses writing patterns and returns a probability. A watermark reads an embedded signal and is simply present or absent.
A detector works on any text. A watermark only exists in text from participating models.
A detector can flag human writing, which is a documented problem. A watermark cannot, but its absence proves nothing.
A detector analyses writing patterns and returns a probability. A watermark reads an embedded signal and is simply present or absent.
A detector works on any text. A watermark only exists in text from participating models.
A detector can flag human writing, which is a documented problem. A watermark cannot, but its absence proves nothing.
A detector guesses whether writing looks machine-generated. A watermark records that text passed through a specific model. Watermarks do not replace detectors, because they only cover providers who add one.
A detector guesses whether writing looks machine-generated. A watermark records that text passed through a specific model. Watermarks do not replace detectors, because they only cover providers who add one.
What an AI watermark proves
What an AI watermark proves
It indicates: the text passed through a particular AI model.
It indicates: the text passed through a particular AI model.
It does not establish: who wrote the underlying ideas, whether any rule or policy was broken, whether a person edited or fact-checked the output, or whether the use was legitimate.
It does not establish: who wrote the underlying ideas, whether any rule or policy was broken, whether a person edited or fact-checked the output, or whether the use was legitimate.
Translation, accessibility tools, grammar checking, and reformatting your own writing all put text through a model. A watermark cannot tell those apart from wholesale generation.
Translation, accessibility tools, grammar checking, and reformatting your own writing all put text through a model. A watermark cannot tell those apart from wholesale generation.
What is still unknown
What is still unknown
Being straight about the gaps, because there are several.
Being straight about the gaps, because there are several.
How much editing removes or degrades a watermark is not publicly documented
No public tool exists for checking whether text carries a watermark
Whether schools and publishers will get access to checking, and how, is undecided
How institutions will treat watermarks in policy is unresolved
How much editing removes or degrades a watermark is not publicly documented
No public tool exists for checking whether text carries a watermark
Whether schools and publishers will get access to checking, and how, is undecided
How institutions will treat watermarks in policy is unresolved
Last checked 12 August 2026.
Last checked 12 August 2026.
Where Lunchbreak fits
Where Lunchbreak fits
Lunchbreak does not read or remove watermarks, and no public tool does. What it does is the step you actually control: seeing how your writing reads, and making it sound like you.
Lunchbreak does not read or remove watermarks, and no public tool does. What it does is the step you actually control: seeing how your writing reads, and making it sound like you.
AI checker. Run your writing against major AI detectors at once and see each result separately. Detectors disagree with each other more than people expect, and seeing the spread is more useful than one number. Free with an account.
AI checker. Run your writing against major AI detectors at once and see each result separately. Detectors disagree with each other more than people expect, and seeing the spread is more useful than one number. Free with an account.
AI humanizer. Rewrites for tone, rhythm, and flow so writing reads naturally instead of stiffly. You compare the original against the rewrite and keep what you want.
AI humanizer. Rewrites for tone, rhythm, and flow so writing reads naturally instead of stiffly. You compare the original against the rewrite and keep what you want.
Useful whether or not watermarking affects you: knowing what a detector says about your work before someone else runs it, fixing prose that reads robotically, making translated or non-native English writing flow, and understanding false positives if human writing gets flagged.
Useful whether or not watermarking affects you: knowing what a detector says about your work before someone else runs it, fixing prose that reads robotically, making translated or non-native English writing flow, and understanding false positives if human writing gets flagged.
Read next
Read next
Frequently asked questions
Frequently asked questions
What is an AI watermark?
What is an AI watermark?
An invisible, machine-readable signal embedded in AI-generated content so it can be identified as machine-made. It does not change how the content looks or reads.
An invisible, machine-readable signal embedded in AI-generated content so it can be identified as machine-made. It does not change how the content looks or reads.
Can you see an AI watermark?
Can you see an AI watermark?
No. It is imperceptible by design.
No. It is imperceptible by design.
Which AI models watermark their text?
Which AI models watermark their text?
Anthropic’s Claude models, for releases after 2 August 2026. No other major provider had announced equivalent text watermarking at the time of writing.
Anthropic’s Claude models, for releases after 2 August 2026. No other major provider had announced equivalent text watermarking at the time of writing.
Can you remove an AI watermark?
Can you remove an AI watermark?
No public documentation establishes what removes one. Coverage says the mark may persist through some editing, which is not a specific answer. Treat confident claims in either direction with scepticism.
No public documentation establishes what removes one. Coverage says the mark may persist through some editing, which is not a specific answer. Treat confident claims in either direction with scepticism.
Can Turnitin detect an AI watermark?
Can Turnitin detect an AI watermark?
No announcement has been made. Turnitin’s AI detection analyses writing patterns, a different mechanism from reading an embedded signal.
No announcement has been made. Turnitin’s AI detection analyses writing patterns, a different mechanism from reading an embedded signal.
Does a watermark mean the text is AI-generated?
Does a watermark mean the text is AI-generated?
It means the text passed through an AI model. That is not the same as saying a human contributed nothing, or that any rule was broken.
It means the text passed through an AI model. That is not the same as saying a human contributed nothing, or that any rule was broken.
Does no watermark mean a human wrote it?
Does no watermark mean a human wrote it?
No. It could be human writing, output from a model that does not watermark, or a mark that degraded. Absence proves nothing.
No. It could be human writing, output from a model that does not watermark, or a mark that degraded. Absence proves nothing.
Is AI watermarking required by law?
Is AI watermarking required by law?
The EU AI Act requires providers of generative AI to mark their output. The obligation sits with the AI companies, not with people using them.
The EU AI Act requires providers of generative AI to mark their output. The obligation sits with the AI companies, not with people using them.