Watermarking and Detecting AI-Generated Text: How Green-List Watermarks and SynthID Work, Their Limits, and Why AI Detectors Mislead

How AI text watermarking works (Kirchenbauer et al.'s green-list scheme and Google DeepMind's SynthID Text), what research shows about paraphrasing and robustness, why style-based AI detectors are unreliable, EU AI Act Article 50 marking duties, and practical guidance for schools, publishers and platforms.

PUBLISHED // UPDATED
DURATION // 06 MIN READ
Watermarking and Detecting AI-Generated Text: How Green-List Watermarks and SynthID Work, Their Limits, and Why AI Detectors Mislead
FIG.01 // VISUAL DOCUMENTATION Watermarking and Detecting AI-Generated Text: How Green-List Watermarks and SynthID Work, Their Limits, and Why AI Detectors Mislead

AI-generated text can be watermarked, and watermarks are far more reliable than style-based “AI detectors”, but neither offers proof on its own. Watermarking works by subtly biasing a model’s word choices in a pattern a detector with the right key can recognise; research shows it can survive moderate paraphrasing given enough text, but short, heavily rewritten or low-entropy text is hard to call, and it only works if the generating model applied the watermark. Classifiers that guess from writing style have high error rates and are biased against non-native English writers; OpenAI withdrew its own classifier in 2023 for low accuracy. The realistic approach combines watermarking, provenance standards, disclosure rules and process-based policies.

Two ways to identify AI text

ApproachHow it worksStrengthsWeaknesses
WatermarkingThe model embeds a statistical signal during generation; a detector with the key tests for itLow false-positive rates with enough text; quantifiable confidenceRequires the model provider’s cooperation; weakened by heavy editing; needs enough tokens
Post-hoc classifiersA model guesses whether text “looks” AI-writtenWorks on any textHigh error rates; biased against some writers; easy to evade
Provenance metadataSigned records of how content was createdStrong for images and videoMetadata is easily stripped from plain text

How green-list watermarking works

The best-known scheme comes from Kirchenbauer et al., “A Watermark for Large Language Models” (arXiv:2301.10226). At each generation step, a hash of the previous token pseudo-randomly splits the vocabulary into a “green” list and a “red” list, and sampling is nudged toward green tokens. Individual word choices still look natural, but over many tokens the text contains more green tokens than chance would predict. A detector that knows the hashing key counts them and computes how unlikely that count would be in human writing.

Because the bias is statistical, confidence grows with length. A short reply may not contain enough tokens for a reliable call, while a long essay can be classified with very low false-positive rates.

How robust are watermarks?

A follow-up study by the same group, “On the Reliability of Watermarks for Large Language Models” (arXiv:2306.04634), tested watermarked text after rewriting by humans, paraphrasing by a non-watermarked model and mixing into longer human-written documents. The authors found watermarks remained detectable after both human and machine paraphrasing, because paraphrases tend to leak n-grams from the original. Detection needed more text: after strong human paraphrasing, the watermark was detectable after about 800 tokens on average at a one-in-100,000 false-positive rate.

Other researchers are more pessimistic. Sadasivan et al., “Can AI-Generated Text be Reliably Detected?” (arXiv:2303.11156), introduced recursive paraphrasing attacks that degraded a range of detectors, including watermark-based ones. The practical conclusion is that watermarks catch casual, lightly edited AI text well, while a determined adversary willing to rewrite heavily can often weaken them.

Google DeepMind’s SynthID Text

SynthID Text uses a related approach called tournament sampling. Its design and a large-scale live evaluation in Gemini were published in Nature in October 2024, and the implementation was open-sourced and made available through Hugging Face Transformers (Nature paper; Hugging Face; Google developer docs; SynthID overview). Open-sourcing lets other model developers apply compatible watermarks to their own outputs. For more on SynthID across media, see our guide to SynthID detection limits.

Known limitations

  • Coverage: a watermark only helps if the generating model applied it. Open-weight models run locally and many commercial systems produce unmarked text.
  • Short texts do not contain enough tokens for reliable detection.
  • Low-entropy text, such as code, formulas or factual lists with few valid word choices, leaves little room to embed a signal.
  • Heavy rewriting and translation can weaken or remove the signal.
  • Key access: detection requires the provider’s key or detector, so third parties cannot check arbitrary watermarks without cooperation.
  • Quality trade-offs: stronger watermarks can slightly constrain word choice, so providers tune the strength.

Detectors without watermarks are unreliable

Classifier-based “AI detectors” that guess from writing style have well-documented problems. When OpenAI launched its AI text classifier in January 2023, it correctly flagged only 26% of AI-written text in its challenge set and mislabelled human writing 9% of the time; it withdrew the tool on July 20, 2023, citing its low rate of accuracy (Ars Technica). A Stanford study found that several popular detectors frequently misclassified essays by non-native English speakers as AI-generated (Liang et al., 2023, arXiv:2304.02819). Schools and employers should never treat a detector score as proof.

Regulatory direction

Article 50 of the EU AI Act requires providers of generative AI systems to mark synthetic audio, image, video and text outputs in a machine-readable format and make them detectable as AI-generated, with obligations applying from August 2026 (EU AI Act Article 50). Deployers must also disclose deepfakes and AI-generated text published to inform the public on matters of public interest, unless it has undergone human review and editorial responsibility. For images and video, the C2PA Content Credentials standard provides signed provenance metadata (c2pa.org). See also our guide to EU AI labelling rules.

Practical guidance

For educators: focus on process rather than detection. Ask for drafts, notes and version history, use in-class writing and oral follow-up, and set clear rules on acceptable AI use. Do not discipline students based on a detector score alone.

For publishers and platforms: require disclosure of AI involvement, label AI-generated content where rules require it and use watermark detection where providers offer it, as one signal among several.

For businesses deploying AI: check whether your model provider applies watermarks and how they can be detected, and plan for EU Article 50 obligations if you serve EU users.

For writers: keep drafts and notes. If your work is ever wrongly flagged, a documented process is the best defence.

Our analysis

Perfect AI text detection is probably impossible, because high-quality AI writing and human writing overlap and determined adversaries can rewrite. But perfect detection is not required for progress. Watermarking makes casual, unedited AI text reliably identifiable when providers participate, which covers much real-world misuse such as spam and low-effort content farms. Style-based detectors should not be used to accuse anyone. The most durable solutions combine watermarking, provenance standards, disclosure rules and institutional policies that reward transparency, so that being dishonest about AI use becomes risky and costly rather than undetectable.