Since August 2026, text written by Claude carries an invisible mark. It is not a hidden character you could find and delete. Anthropic states explicitly that nothing is added to the text and that there are no hidden characters. The watermark is a statistical pattern in which words the model picks. Invisible characters are a completely separate phenomenon, they mostly concern ChatGPT, and they have nothing to do with watermarking. For your thesis, the important part is that your university cannot verify either of them today.
Two things that get confused in every discussion
When someone says "the AI hid something in my text", they usually mean one of two entirely different things.
Invisible Unicode characters. Real characters that are physically present in the text but display nothing, or look like an ordinary space. They can be found, counted and deleted. They are not watermarks and nobody is tracking anything with them.
A statistical watermark. Nothing is added. The model simply chooses differently between words that would fit the sentence equally well. The resulting pattern can only be read by whoever holds the key. Find and replace is powerless against it, because there is nothing to replace.
The first is an unintended by-product of training. The second is a deliberate feature that European legislation asked for. The difference matters, because each one means something different for your thesis.
What Claude actually does as of August 2026
Anthropic explained how the watermark works on 14 August 2026 in How Claude's text watermarking works. The key sentence:
Nothing is added to the text and there are no hidden characters.
Anthropic, How Claude's text watermarking works
So if you have read that Claude inserts invisible markers into its answers, that is a misunderstanding the vendor itself refutes.
Why Anthropic is doing this
The reason is Article 50 of the EU AI Act, which applies from 2 August 2026. Paragraph 2 requires providers of AI systems generating synthetic audio, image, video or text to ensure that the outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated".
How that duty is met in practice is spelled out by the Code of Practice on Transparency of AI-Generated Content. The European Commission reports that around 190 organisations had signed it by the end of July 2026, among them Anthropic, Google, OpenAI, Meta, Microsoft and Mistral.
Anthropic switched watermarking on worldwide rather than only in Europe, explaining that it does not yet have a durable way to scope it by region. According to its documentation on marking AI-generated content, the watermark covers the Claude app, the API, Claude Code, Claude Cowork and Claude Tag. The documentation mentions no opt-out.
How it works technically
As the model builds a sentence, it repeatedly reaches a point where several words would fit equally well. After "the weather today was cold and", it could continue with overcast, grey, damp or dull. Normally a random number generator settles it.
The watermark replaces that randomness. A secret key, together with a few preceding words, decides the pick. The result reads just as naturally, but anyone holding the key can verify afterwards that the choices were not random. Anthropic compares it to a board game where you use the digits of pi instead of dice rolls: the moves look random, yet they can be reproduced exactly.
The approach is not new. It builds on SynthID-Text, which Google DeepMind published in 2024 in the paper Scalable watermarking for identifying large language model outputs in Nature and later released as open source.
What the watermark survives and what it does not
Anthropic is fairly open about the limits:
- A translation produced by Claude carries the watermark, because every word in it is chosen by the model.
- Light editing probably will not remove it. A complete rewrite in which every word is replaced will.
- On factual passages the watermark is sparser, because there the model has fewer choices it can make without hurting accuracy.
- In code it barely applies. Where an exact output is required, it is not applied at all.
- Proofreading your own text leaves almost nothing for the watermark to attach to.
- On short excerpts detection is unreliable. Confidence grows with the length of the passage.
So if you use Claude to proofread a chapter you wrote yourself, there is practically no watermark in the result. If you have a whole subsection generated and paste it in untouched, the whole thing carries one.
Who can actually read it
Almost nobody, for now. The watermark can only be verified with the key, and Anthropic says the detection API is still being prepared and that it is working out the details of how to make it available. Until then, Claude's text carries the mark but nobody outside Anthropic can read it.
Where the other models stand
Gemini. Google deployed SynthID-Text in Gemini back in 2024 and published the method openly. It is the same principle Anthropic now uses.
ChatGPT. OpenAI signed the code of practice as well, but it has not confirmed blanket text watermarking in the public version of ChatGPT.
Watch December 2026. Systems that were on the market before 2 August 2026 get a grace period until 2 December 2026 to meet the machine-readable marking duty, according to the practical guide to Article 50. In other words, the number of tools that watermark text is very likely to grow before the year is out.
Invisible characters: a separate story
What people commonly find in ChatGPT output is not a watermark. These are real Unicode characters that reached the output from the training data. The usual suspects:
- the narrow no-break space (U+202F), which looks like a space, only thinner,
- the zero-width space (U+200B), which takes up no room at all,
- the em space (U+2003) in place of an ordinary space,
- the em dash (U+2014), which is visible but is set differently in most academic style guides.
This is not covert marking. The models were trained on typographically careful material, meaning academic papers and professional publishing, where these characters are used entirely legitimately, and they learned to write them too.
Why they are a nuisance in a thesis
Not because they give you away. They are a nuisance in practical terms:
- Document search will not find them, so a bulk find-and-replace of a term will silently skip some occurrences.
- Line breaking behaves unexpectedly around a no-break space, which shows up mainly when you export to PDF.
- Automatic bibliography generation and citation checking can trip over such a character.
- When you check your length, characters you cannot see still count. If you are running just under the limit, your word count will not be what you think it is.
How to find and remove them
You do not need a third-party tool for this, and you should not upload your unpublished thesis to one either.
- Find one such character in your text (typically between a number and its unit, for example before a percent sign) and copy it.
- In Word, open Find and Replace, paste the character into the Find field and type an ordinary space into the Replace field. Then Replace All.
- Repeat for every type of character you found.
- Finally check that line breaking in tables and in the reference list did not fall apart.
If you write in Google Docs, the same logic works through Edit and Find and replace.
It matters to be clear about what you have and have not solved. You cleaned up the typography. You have said nothing about whether the text was written with AI, because these characters were never evidence, and removing them is not covering your tracks.
What this means for your thesis
A watermark is not an AI detector, and a detector does not read watermarks. They are two independent technologies. Turnitin's documentation on its AI writing detection model describes its own model, which estimates the probability that text is machine-generated. It does not hold Anthropic's watermark key, so it cannot verify it. How detection works and what its score means is covered in our article on how AI-generated text gets detected.
Your university cannot verify the watermark today. The detection interface is not public. That can change, and in the meantime the mark stays in the text, because it survives copy and paste.
A watermark does not say you cheated. It only says that a given passage came out of a model. Whether that was acceptable is decided by your faculty's policy, not by technology. The reverse holds too: an absent watermark does not prove a human wrote the text.
None of this changes the basic rule. Declare and cite what you actually used, following your university's rules. We wrote a separate piece on whether students can use ChatGPT.
A practical minimum before you submit
- Write in a document with version history. It remains the strongest evidence of how the text came about, and no watermark changes that.
- Do not paste generated paragraphs untouched. Not only because of the watermark. Text you did not rewrite in your own words is text you cannot defend at your thesis defense.
- Clean up typography at the end, not as you go. One pass before submission is enough.
- Read your faculty's AI policy. University rules are being updated through 2026 and what applied last year may not apply now.
- Keep your notes, extracts and raw data. Your own data collection is the one thing a model cannot manufacture.
What not to do
- Do not treat deleting invisible characters as a solution. It has nothing to do with the statistical watermark and says nothing about where the text came from.
- Do not upload your thesis to online cleaners and detectors. You hand your unpublished text to a third party and gain no usable evidence.
- Do not run your text through "humanizer" tools. Since 27 August 2025, Turnitin also detects the use of detection-evasion tools and folds it into the score.
- Do not make the watermark a bigger problem than it is. Nobody outside the vendor can read it today. The real risk in a thesis is still the same one: text you cannot explain.
If you need expert support with the literature review, the structure or the factual groundwork, take a look at our master's thesis services or bachelor's thesis services. We prepare the materials, you write the final text. That way no watermark can arise in the first place.
Frequently asked questions
Does Claude add invisible characters to text?
No. In its watermarking documentation, Anthropic states explicitly that nothing is added to the text and that there are no hidden characters. The watermark is a statistical pattern in word choice, not an inserted character. The invisible Unicode characters people find in AI output are a separate phenomenon and relate mainly to ChatGPT.
Can my university tell that Claude wrote my text?
Not today. The watermark can only be read with the key, and Anthropic is still preparing the detection interface. Common tools such as Turnitin do not check for watermarks, they use their own statistical model. That may change, and the mark stays in the text in the meantime.
Does the watermark disappear if I rewrite the text in my own words?
According to Anthropic, light editing probably will not remove it, while a complete rewrite in which every word is replaced will. In practice, genuinely rewriting the text in your own style weakens the watermark substantially. That is not the reason to rewrite it, though. The reason is that rewritten text is text you understand and can defend.
Does watermarking affect the quality of the text?
Anthropic says internal testing showed no impact on content, creativity or readability, and that raters in a controlled study saw no difference between watermarked and unwatermarked answers. Google has used the same principle in Gemini since 2024.
How do I find out whether my thesis contains invisible characters?
Copy a suspicious spot, typically the space before a unit or a percent sign, paste it into Word's Find field and look at how many occurrences it reports. If it finds dozens, you are dealing with a narrow no-break space or another special character. They can be replaced in bulk with an ordinary space.
Does a watermark mean I broke my university's rules?
No. The watermark only says the text came out of a model. Whether that use was permitted is decided by your faculty's policy. Some universities allow AI for language editing, while almost all of them prohibit generating entire chapters.
Does ChatGPT watermark text too?
OpenAI has not confirmed blanket text watermarking in the public version of ChatGPT, even though it signed the EU transparency code. The invisible characters that show up in its output are a training artefact, not a watermark. That situation may change by 2 December 2026, when the grace period ends for systems placed on the market before August 2026.