Jazyk AI GUIDE

Text vygenerovaný LLM

Text watermarking hides a statistical signal in the words a language model chooses, so a detector holding a secret key can later tell that a passage came from that model.

Poslechněte si tento návodAI-vyprávěný · ~8 min

Na této stránce9 min čtení
  1. Přehled
  2. Klíčové věci
  3. Hluboký ponor
  4. Computing a green list detection score by hand
  5. Real-World Implementace
  6. Budoucnost vodoznaku Text generovaný LLM
  7. Uvedení do praxe
  8. Zdroje a další čtení
  9. Pokračujte v objevování
  10. Často kladené otázky

Přehled

It matters because the EU AI Act now requires machine-readable marking of AI-generated text, and Google, OpenAI and others have deployed schemes, yet paraphrasing, translation and short passages still weaken detection.

Klíčové věci

  1. A text watermark biases a language model's word choices with a secret key so a detector can later test whether text came from that model.
  2. The 2023 Maryland green list watermark detected about 98.4% of 200-token samples at a z-score threshold of 4.
  3. Google DeepMind's SynthID-Text, published in Nature in October 2024, was tested on about 20 million Gemini responses without measurable quality loss.
  4. Article 50(2) of the EU AI Act requires machine-readable marking of AI-generated text from August 2, 2026, where technically feasible.
  5. Paraphrasing, translation, short passages and low-entropy answers all weaken detection, and a missing watermark does not prove human authorship.

Hluboký ponor

A text watermark is a hidden pattern placed in machine-generated writing at the moment it is produced, so that software can later recognize the text as coming from a particular model. Unlike a visible label, it is invisible to readers and survives copy and paste because it lives in the choice of words themselves. It differs from AI text detectors that guess authorship from style, which have a poor record: OpenAI shut down its own AI text classifier in 2023 because of its low accuracy, according to TechCrunch. A watermark instead gives a statistical test with a known false positive rate, but only for text from a model that used that watermark with that key.

The idea works because language models generate text one token, a word or word piece, at a time by sampling from a probability distribution over their vocabulary. Often several next tokens are almost equally good, and the model has freedom in which it picks. A watermark uses that freedom. A secret key and the preceding tokens feed a pseudorandom function, which nudges the choice toward a hidden subset of tokens or a hidden ranking. To a reader the text reads normally. To a detector with the key, the text shows far more of the favored choices than chance would predict.

The best-known public scheme came from John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers and Tom Goldstein at the University of Maryland in January 2023. At each step, the previous token seeds a random number generator that splits the vocabulary into a green list and a red list. In the soft version, the model adds a constant bias, delta, to the scores of green tokens before sampling, so green words become more likely without being forced. Detection counts green tokens and computes a z-score, a measure of how many standard deviations the count is above what unwatermarked text would show. With half the vocabulary green, a bias of 2 and a threshold of z = 4, about 98.4% of 200-token samples from the OPT-1.3B model were detected, and the paper reported no false positives in its runs. A z threshold of 4 corresponds to a false positive rate of about 3 in 100,000.

Google DeepMind's SynthID-Text, published in Nature on October 23, 2024 by Dathathri and colleagues, uses a different mechanism called tournament sampling. The model proposes several candidate tokens, which compete in a knockout tournament over 30 layers by default, with winners decided by pseudorandom g-values derived from the secret key and the previous four tokens. A non-distortionary setting preserves text quality, while a distortionary one trades some quality for stronger detection. The paper reports that the non-distortionary version beat Gumbel sampling, another published scheme, on detectability, especially at low entropy, and added only 0.57% latency per token on Gemma 7B. SynthID-Text had been rolled out to the Gemini app and web experience, as Google DeepMind announced in May 2024, and its code was open-sourced.

Adoption accelerated with regulation. Article 50(2) of the EU AI Act requires providers of AI systems that generate synthetic text, images, audio or video to mark outputs in a machine-readable format so they are detectable as artificially generated, with solutions that are effective, interoperable, robust and reliable as far as technically feasible; it applies from August 2, 2026. OpenAI had built a text watermark by 2024 but held back; in August 2024 it told TechCrunch it was taking a deliberate approach because of circumvention risks and possible unfair effects on non-English speakers. In October 2026 TechCrunch reported that OpenAI would watermark ChatGPT and Codex text in the EU, using a method its technical report calls textGrain, and that Anthropic had announced in August 2026 that it would watermark Claude's text worldwide.

Compared with alternatives, watermarking has clear strengths and limits. Post-hoc detectors need no cooperation from the model provider but are unreliable. Watermarks travel with the words but only cover providers who adopt them; the SynthID paper notes that watermarking requires coordination among providers and is hard to enforce on open-source models that anyone can run without it.

The failure modes are well documented. Low-entropy text, where only one answer is reasonable, such as the capital of France or a math result, leaves little room to embed a signal; Google's blog names factual prompts as a weak case, and TechCrunch's account of OpenAI's report names math answers. Short passages give too few tokens for a confident test. Paraphrasing and translation scramble the pattern: OpenAI's own test, reported by TechCrunch, found that replacing 10% of words with synonyms dropped detection from about 92% to 66%. The Maryland group's 2023 reliability study found watermarks remained detectable after strong human paraphrasing, but only at around 800 tokens. Attackers can also try to steal the pattern and spoof it onto human text, or scrub it out, threats the SynthID paper lists. And absence of a watermark proves nothing, as OpenAI states.

In practice, a provider adds watermarking at the sampling step, with no retraining, and must protect its keys; Google's reference code warns that its random number generator is not a vetted cryptographic primitive and offers no cryptographic security guarantees. Detection needs the key, so access is usually restricted, as with OpenAI's initial limit to approved researchers and expert organizations. Anyone using a detector result for decisions about people, such as students, should treat it as one piece of evidence with a stated false positive rate, never as proof.

04Zpracovaný příklad

Computing a green list detection score by hand

  1. Assume a watermark with green list fraction gamma = 0.5 and a detection threshold of z = 4.

  2. Take a 200-token passage and suppose the detector, using the key, finds 140 of the tokens on their green lists.

  3. Compute the expected green count for human text: 0.5 times 200 = 100.

  4. Compute the standard deviation: the square root of 200 times 0.5 times 0.5, which is about 7.07.

  5. Compute z = (140 - 100) / 7.07, which is about 5.66, above the threshold of 4.

  6. Repeat for a 50-token excerpt with 35 green tokens: z = (35 - 25) / 3.54, about 2.83, below the threshold.

Co ukazuje

The 200-token passage is flagged as watermarked and the 50-token excerpt is not, even though both have the same 70% green rate. This shows why longer text gives more confident detection; check by recomputing z with your own counts.

Real-World Implementace

Google DeepMind's SynthID-Text watermarks Gemini output; a 2024 Nature paper reports a live test on about 20 million Gemini responses with no significant change in user thumbs-up or thumbs-down rates.

Developers adding SynthID watermarking to open models through Google DeepMind's Apache-2.0 reference implementation, which wraps Hugging Face Gemma and GPT-2 classes and ships a frequentist detector that returns an exact p-value.

OpenAI watermarking ChatGPT and Codex text for users in the European Union from October 2026 to comply with the EU AI Act, with an opt-in watermark for select API models worldwide, per TechCrunch.

Researchers testing robustness with the University of Maryland green list watermark, whose 2023 follow-up found watermarks still detectable after strong human paraphrasing at about 800 tokens with a one-in-100,000 false positive rate.

Compliance teams mapping obligations under Article 50(2) of the EU AI Act, which requires providers of systems that generate synthetic text to mark outputs in a machine-readable, detectable format where technically feasible.

Budoucnost vodoznaku Text generovaný LLM

Several threads are visible in sources from 2023 to 2026. Robustness is the main technical problem: the Maryland group's ICLR 2024 study showed detection survives paraphrasing only with long passages, and OpenAI's October 2026 report found that replacing 10% of words with synonyms cut detection from about 92% to 66%. Security is a second: the SynthID authors list stealing, spoofing and scrubbing attacks, and Google's reference code disclaims cryptographic guarantees. Coverage is a third, since watermarks only mark text from providers that use them; the EU AI Act's Article 50, applying from August 2, 2026, pushes major providers to mark output, and TechCrunch reported that Anthropic, Google, Meta, Microsoft and OpenAI committed to the EU code of practice on AI-generated content. Interoperability is unresolved, because each company's detector currently recognizes only its own watermark. Fair use of detectors in education and hiring remains contested, given the concerns OpenAI raised in 2024 about effects on non-English speakers. Whether a shared standard for text watermarks will emerge is an open question.

Uvedení do praxe

Proč na tom záleží

  • Rychlost a měřítko. Jazykové pracovní postupy se mohou pohybovat rychleji, aniž by byla obětována konzistentnost.
  • Přístup a dosah. Rozšiřuje přístup napříč jazyky a komunikačními styly.
  • Jasnější rozhodnutí. Týmy mohou strávit více času úsudkem, zatímco automatizace zvládne opakování.

Pozor na

  • Halucinovaná fakta mohou tiše vstupovat do zpráv, podpůrných toků nebo výstupů výzkumu.
  • Citlivost na výzvy může způsobit nekonzistentní výsledky napříč podobnými požadavky.
  • Citlivá textová data mohou být vystavena, pokud je řízení přístupu slabé.

Další kroky

  1. Před zavedením definujte výstupní formát, tón a standardy kvality.
  2. Pozemní reakce s důvěryhodnými zdroji, kdykoli záleží na přesnosti.
  3. Udržujte kontrolní bod lidské kontroly pro vysoce důležité výstupy.
  4. Sledujte vzorce selhání a pravidelně opakujte výzvy nebo pracovní postupy.

Zdroje a další čtení

  1. arXivA Watermark for Large Language Models
  2. arXiv (ICLR 2024)On the Reliability of Watermarks for Large Language Models
  3. Nature (via PubMed Central)Scalable watermarking for identifying large language model outputs
  4. GitHub (Google DeepMind)google-deepmind/synthid-text
  5. Google DeepMindWatermarking AI-generated text and video with SynthID
  6. EU Artificial Intelligence Act (Future of Life Institute)Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems
  7. TechCrunchOpenAI says it's taking a 'deliberate approach' to releasing tools that can detect writing from ChatGPT
  8. TechCrunchOpenAI will start watermarking ChatGPT's text in the EU

Pokračujte v objevování

Newsletter zdarma

Získejte denní přehled o AI

Tři ověřené příběhy ze světa AI každý všední den ráno, psané srozumitelnou angličtinou. Navždy zdarma, bez reklam.

Jeden e-mail každý všední den. Odhlásit se můžete jedním kliknutím. Vaši adresu nikdy neprodáváme ani nesdílíme.

Otestujte se

Take the Watermarking LLM-Generated Text quiz

Okamžitá zpětná vazba ke každé odpovědi a sdílitelný certifikát s ověřitelným ID po absolvování kurzu.

Spustit kvíz

Podpořte bezplatnou AI vzdělávání. AI Understanding je nezisková organizace 501(c)(3) — žádné reklamy, žádná platba za přístup, nikdy. Darujte

Často kladené otázky

Co je text generovaný vodoznakem LLM?

Text watermarking hides a statistical signal in the words a language model chooses, so a detector holding a secret key can later tell that a passage came from that model. It matters because the EU AI Act now requires machine-readable marking of AI-generated text, and Google, OpenAI and others have deployed schemes, yet paraphrasing, translation and short passages still weaken detection.

How does the soft green-list watermark in the cited scheme influence token generation?

The scheme softly promotes selected tokens while retaining possible nongreen outputs.

What must match when applying a detector for a keyed green-token watermark?

The tested implementation must reproduce its partitions; original model parameters are not inherently required.

Why can a longer eligible passage provide more evidence for a statistical text watermark?

Length is one factor; constrained outputs, repeated contexts and the scheme also matter.

Which published text watermark uses tournament sampling and is documented for Gemini text?

The research and official Google report describe the named method, distinct from every other watermark.

Jaký je známý způsob oslabení nebo odstranění textového vodoznaku?

Protože mnoho vodoznaků žije ve volbách povrchových tokenů, parafrázování, překlad nebo úpravy mohou narušit vzor zeleného žetonu.