AI Text Watermarks Explained: How Detection Works
Quick answer: An AI text watermark is an invisible statistical pattern in the tokens a language model chooses. Using a secret key, the model slightly favors certain tokens as it writes; a detector with the key checks whether that pattern is unusually strong. In a simple green-list test, if 140 of 200 scored tokens fall in a random 50% green list, the z-score is about 5.7, far beyond chance. In 2026, Google’s SynthID, OpenAI’s textGrain and Anthropic’s Claude watermark all use related ideas. A missing watermark does not prove a human wrote the text.
Facts checked November 22, 2026.
Teachers, publishers and platforms all want to know: was this text written by AI? Ordinary “AI detectors” guess from writing style and are often wrong. A watermark is different: the AI company deliberately hides a mathematical signal in the text as it is generated, which a detector can later test for. The math combines hashing, probability and a statistics test you may have met in class.
Step 1: Split the Vocabulary With a Hash
Before choosing each word (token), the system takes the previous token, feeds it into a hash function with a secret key, and uses the result as a random seed. That seed splits the vocabulary into a green list (say, a fraction \(\gamma = 0.5\) of all tokens) and a red list. The split changes at every position and looks random to anyone without the key. Hash functions are the same tools used in Bitcoin.
Step 2: Nudge the Model Toward Green Words
A language model assigns a score to every possible next token (see the probability behind AI). The watermark adds a small bonus \(\delta\) to the scores of green tokens before sampling. When many words would fit equally well, the model usually picks a green one; when only one word makes sense, it still picks that word, so quality barely changes. Those low-choice (low-entropy) positions carry a weaker watermark signal, which is why factual passages, math and code are harder to mark.
Step 3: Detect With a z-Score
Human writers know nothing about the secret lists, so in \(T\) scored tokens they hit green about \(\gamma T\) times (assuming the lists behave like independent random draws; repeated phrases can break this assumption), with standard deviation \(\sqrt{T\gamma(1-\gamma)}\) (a binomial distribution). The detector counts green tokens \(G\) and computes
\[ z = \frac{G – \gamma T}{\sqrt{T \gamma (1 – \gamma)}}. \]Worked example: \(T = 200\) tokens, \(\gamma = 0.5\), and \(G = 140\) green tokens:
\[ z = \frac{140 – 100}{\sqrt{200 \times 0.5 \times 0.5}} = \frac{40}{\sqrt{50}} \approx \frac{40}{7.07} \approx 5.7. \]| Text | Green tokens out of 200 | z-score | Verdict at threshold z = 4 |
|---|---|---|---|
| Typical human text | About 100 (±7) | About 0 | Not watermarked |
| Lightly edited watermarked text | 125 | \(25/7.07 \approx 3.5\) | Below this threshold; depends on calibration |
| Watermarked text | 140 | About 5.7 | Watermarked |
Under the normal approximation, a z-score above 4 happens by chance only about 3 times in 100,000 for human text, so false positives are rare when the threshold is strict. (The exact binomial chance of 140 or more green tokens out of 200 by luck is even smaller, about 8 in a billion.) Longer texts give stronger evidence, because the signal grows like \(T\) while the noise grows only like \(\sqrt{T}\).
Which AI Chatbots Watermark Their Text in 2026?
| System | Approach | Status |
|---|---|---|
| Google Gemini: SynthID-Text | Keyed “tournament” sampling that picks among candidate tokens | Published in Nature in October 2024 and released as open source; a live test of nearly 20 million Gemini responses found no measurable drop in user-rated quality |
| OpenAI: textGrain | Statistical signal in generated text | Announced October 5, 2026: opt-in for API customers worldwide; automatic for eligible ChatGPT and Codex text in the EU, rolling out over the following weeks; detector access initially for approved research and academic groups |
| Anthropic: Claude | A version of the SynthID-Text approach; no hidden characters or extra tokens | Announced August 14, 2026 for future Claude models, applied globally; detection API in private preview for eligible organizations such as regulators, researchers and fact-checkers |
| Kirchenbauer et al. (2023) | Green-list bonus, as above | Influential research method from the University of Maryland |
The companies publish their own limits. OpenAI reported that for 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to about 17%. Anthropic says its watermark works poorly on small samples, is sparser on factual passages and code, survives light editing but not a complete rewrite, and cannot show whether text was written by a human or by a different AI. Images are handled differently: Anthropic and others attach C2PA “Content Credentials,” signed metadata in the file.
The Law: Marking AI Content in 2026
The European Union’s AI Act includes transparency rules (Article 50) that apply from August 2, 2026: providers of generative AI must mark synthetic text, images, audio and video in a machine-readable way so it can be detected. For generative AI systems already on the market before that date, the marking obligation in Article 50(2) has a transition period until December 2026. This law is the main reason OpenAI and Anthropic launched text watermarks in 2026.
Limits of Watermarking
- Paraphrasing by a person or another AI replaces tokens and weakens the signal, as in the “lightly edited” row above.
- Short texts carry too few tokens for a confident test.
- Coverage: a watermark only works if the company that made the model added it; open-source models can be run without one.
- Not the same as AI detectors: style-based detectors have no secret key and are much less reliable, which leads to false accusations.
Common Misconceptions
“A watermark is hidden text or special characters.” No. It is a statistical preference for certain ordinary words; the text looks normal.
“Watermarks prove a person cheated.” They give strong statistical evidence that a model generated text, not proof of intent, and editing weakens them.
“Any AI text can be watermark-detected.” Only text from models that embed a watermark, checked with the right key.
Watch: How AI Watermarks Work
If you learn better by watching, these videos are worth your time.
Check Your Understanding
Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.
Answer each question and press Submit to see your result right away.
Frequently Asked Questions
How do AI text watermarks work?
The model is secretly nudged to prefer a pseudo-random “green list” of words chosen by hashing the previous word with a key. A detector with the key counts green words; far more than expected by chance means the text was watermarked.
Can AI watermarks be detected without the model?
Yes. In the green-list method, detection needs only the hashing key and a statistical test, not the model itself.
What is SynthID?
Google DeepMind’s watermarking technology. SynthID-Text, published in Nature in October 2024, marks Gemini outputs by adjusting the sampling step; a live test covering nearly 20 million Gemini responses found no measurable loss of quality.
Does ChatGPT or Claude watermark text?
In 2026, yes in part. OpenAI announced textGrain on October 5, 2026, opt-in for the API and automatic for eligible ChatGPT and Codex text in the EU. Anthropic announced on August 14, 2026 that future Claude models watermark text globally.
Can AI watermarks be removed?
They can be weakened by paraphrasing or heavy editing, and short texts give weak evidence. They also only exist if the model provider added them.
Does the law require AI watermarks?
The EU AI Act’s Article 50 transparency rules, applying from August 2, 2026, require providers of generative AI to mark synthetic content in a machine-readable, detectable way, with a transition to December 2026 for systems already on the market.
Keep Exploring
- Why Does AI Hallucinate? Math, Examples and Fixes
- The Math Behind ChatGPT: Explained With Examples
- Is Using ChatGPT for Math Homework Cheating?
Sources and Further Reading
- Kirchenbauer et al., A Watermark for Large Language Models (arXiv 2301.10226)
- Dathathri et al., Scalable watermarking for identifying large language model outputs (Nature, Oct 2024)
- European Commission: Quick facts, transparency rules for AI systems
- Anthropic: How Claude’s text watermark works (Aug 14, 2026)
- The New Stack: OpenAI text watermarking (textGrain) (Oct 5, 2026)
- C2PA: Content Credentials
Related to This Article
More math articles
- How to Balance the Scales: Inequalities with Addition and Subtraction of Mixed Numbers
- Adding and Subtracting Fractions with Unlike Denominators Calculator (with Steps)
- Matrix Calculator (Add, Multiply, Determinant & Inverse)
- How to Evaluate Logarithm? (+FREE Worksheet!)
- Gödel’s Incompleteness Theorems Explained Simply
- The Ultimate PSAT Math Course (+FREE Worksheets & Tests)
- 3rd Grade SBAC Math Worksheets: FREE & Printable
- Perfect Competition: Firm and Market Graphs Explained
- Convert Rational Numbers to a Fraction
- Sleep Calculator


What people say about "AI Text Watermarks Explained: How Detection Works - Effortless Math"?
No one replied yet.