That last part changes everything about AI detection, and it is the part most explainers skip. Here is how it actually works, who is doing it, and what it means when you paste a text into a detector — including this one.

The intuition: a thumb on the scale of word choice

When a language model writes, it does not choose each next word deterministically. At every position it holds a probability distribution — many plausible next words, each with a weight — and samples one. “The sky was overcast” and “The sky was grey” might both be perfectly good continuations, each with a reasonable probability.

A watermark exploits that freedom. Instead of sampling purely at random, the model uses a keyed source of randomness that consistently nudges it toward certain words in certain contexts — say, “overcast” rather than “grey” whenever the preceding words hash a particular way. One nudged choice proves nothing. Hundreds of them, across a full document, form a statistical fingerprint that is overwhelming evidence — if you know which words were the favored ones. Anthropic describes its scheme in exactly these terms: the choices are still random, but “the source of the randomness is different” [Anthropic, Aug 2026].

How it works under the hood (the SynthID-Text scheme)

The scheme now used in production by Google, and adopted by Anthropic in August 2026, is SynthID-Text, published by Google DeepMind researchers in Nature in October 2024 [Dathathri et al., Nature, Oct 2024]. It works in three steps at each word position:

  1. Seed from context + key. The model hashes a sliding window of the last few generated words together with a secret watermarking key. That hash seeds a pseudorandom function.
  2. Assign g-values. The pseudorandom function assigns a hidden score — a g-value — to every candidate next word (in fact, several layers of g-values per position).
  3. Tournament sampling. Several candidate words are drawn from the model's normal probability distribution and compete in a small bracket: at each layer, the candidate with the higher g-value advances. The winner is what the model writes.

The result is text that still reads naturally and still follows the model's own distribution — Google reports running the scheme across roughly 20 million live Gemini responses with no measurable quality loss [Dathathri et al., Nature, Oct 2024] — but that systematically leans toward high-g-value words.

Detection simply replays the math. Given a text and the same secret key, the detector re-computes the g-values for every position and asks: are the g-values of the actual words suspiciously high, position after position? Random human writing scores like a coin flip. Watermarked text scores like a coin that came up heads 800 times out of 1,000.

A tiny worked example

Suppose the text so far is “By afternoon the sky had turned” and the model's top candidates are grey (35%), overcast (30%), and dark (20%). The key + context hash assigns g-values: overcast 0.91, dark 0.44, grey 0.12. Candidates are drawn by their normal probabilities, then the tournament favors the higher g-value — so overcast wins more often than its 30% would suggest. A human reader sees an ordinary sentence. The key-holder sees one more data point where the “hot” word won. Multiply by every position in a 1,000-word document and the signal becomes unmistakable — to the key-holder, and only to the key-holder.

Why can't anyone check for the watermark?

Because if anyone could check it, anyone could remove it.

The g-values are the output of a keyed pseudorandom function. Without the key, they are indistinguishable from noise — by design; that is the security property [Dathathri et al., Nature, Oct 2024; Google AI dev docs, accessed Aug 2026]. A public detector would hand attackers an oracle: keep paraphrasing until the detector says “clean,” and the watermark is gone. Keeping the key secret keeps the watermark robust — at the price that only the provider can ever verify its own watermark. Anthropic puts it plainly: the pattern “is detectable to anyone who has a key that encodes it” — which in practice means Anthropic [TechCrunch, Aug 11, 2026].

Google has open-sourced the SynthID-Text algorithm (it ships in Hugging Face Transformers), but that only lets you watermark and detect your own model's output with your own key. It does not let anyone detect Google's or Anthropic's production watermarks, whose keys and detector configurations are secret [Google AI dev docs, accessed Aug 2026].

What breaks a watermark — and what doesn't

The watermark lives in word choices, so it survives anything that preserves them and degrades with anything that rewrites them [Anthropic, Aug 2026; Dathathri et al., Nature, Oct 2024]:

  • Survives: copy-paste, formatting changes, light edits, trimming a few sentences.
  • Weakens: moderate editing, mixing AI text with your own writing, short excerpts (fewer positions = weaker statistics).
  • Destroys: heavy paraphrasing or a comprehensive rewrite (Anthropic's own words: “comprehensive rewrites will eliminate” it), and translation to another language — the watermarked word choices simply no longer exist.
  • Barely there to begin with: low-entropy output. When there is only one right answer — exact code syntax, arithmetic, a factual list — the model has no free choice to bias, so there is little watermark to find. Anthropic notes Claude's code output is minimally watermarked, mostly in comments [Anthropic, Aug 2026].

Who watermarks text today (as of August 16, 2026)

  • Google — watermarking Gemini app and web text with SynthID-Text since 2024 [Google DeepMind, accessed Aug 2026].
  • Anthropic — announced August 11–14, 2026: Claude models launched after August 2, 2026 watermark text worldwide across all surfaces (API, Claude Code, cloud platforms); older models to follow “over the coming months.” A detection API is “coming soon” but not yet available [Anthropic, Aug 2026].
  • OpenAI, Meta, Mistral, Microsoft — no text watermark shipped, but all signed the EU's Transparency Code of Practice in July 2026, and EU law requires machine-readable marking of AI text from existing systems by December 2, 2026 [European Commission, Jul 2026; ReedSmith, accessed Aug 2026]. Expect movement in this window.
  • xAI (Grok) — did not sign the Code and has made no watermark commitment, though it remains bound by the EU AI Act itself [European Commission signatory list, Aug 2026; CityAM, Aug 12, 2026].

We track the per-provider picture — live, from the same configuration our detection backend uses — on Who watermarks AI text today?

The danger scenario: watermarked text that looks human

Here is the trap the current moment creates. Suppose someone submits a long, unedited essay written by a new Claude model. That text carries a keyed watermark — across a document like that, statistical evidence strong enough for Anthropic to demonstrate AI generation with overwhelming confidence (remember: the signal weakens with editing and shortness, as covered above). But no tool other than Anthropic's can see that evidence, and Anthropic's detection API is not yet public.

Meanwhile, a statistical detector — ours included — analyzes the same text and might score it “likely human,” because the watermark does not change the text's statistical texture in ways classifiers reliably catch, and modern models are increasingly hard to distinguish from human prose. So the strongest evidence in existence (a keyed watermark) and the only evidence available to you (a statistical estimate) can point in opposite directions on the same document.

This is why any detector that returns a confident “human-written” verdict is overpromising. Absence of statistical evidence is not evidence of human authorship.

What this site can and cannot tell you

Two checks are live on Cobalynx today. The text scan: paste text and get one of three statistical verdicts — Likely AI, Inconclusive, or Likely human — as a calibrated probability with published error rates (our methodology page defines each state). And the file check: upload a PNG or JPEG and we verify its C2PA Content Credentials cryptographically; upload a DOCX or PDF and we read its metadata for tool fingerprints — honestly framed as the weak signal metadata is (what a credentials check can and cannot prove is walked through in the Content Credentials guide). Pasted text itself carries no file-level provenance (pasting strips it), so the text scan only looks for invisible Unicode artifacts, another clearly-labeled weak hint.

Where provenance signals can exist, we grade the evidence into three states, honestly:

  1. Confirmed AI — reported only when a deterministic signal exists: a valid provenance manifest (C2PA Content Credentials) in an uploaded image whose credentials record AI generation — this is live today on the file check — or, once providers open their detection APIs, a provider-verified watermark hit. No text-watermark API is open yet (the keyed asymmetry above), so no text scan can currently reach this state.
  2. Indicators present — statistical analysis finds patterns typical of AI generation, reported as a calibrated probability with published error rates, never as a verdict. This is what the live scan does today: its “likely AI” and “inconclusive” verdicts are exactly this kind of evidence, never a deterministic confirmation.
  3. Unknown — no deterministic signal and statistics inconclusive. We will never label this “human-verified,” because nothing available today can prove human authorship.

What we cannot do — and neither can anyone else without the keys — is detect Google's or Anthropic's text watermarks. The EU's Code of Practice requires providers to stand up interoperable detection access by February 2, 2027 [TechPolicy.Press; Stibbe, accessed Aug 2026]. The day those doors open, we will connect to them and say so on this page.

Sources

  1. Anthropic, “A text watermark for Claude,” anthropic.com/news/claude-text-watermark (announced Aug 11–14, 2026; accessed Aug 16, 2026)
  2. Dathathri et al., “Scalable watermarking for identifying large language model outputs,” Nature (Oct 2024), nature.com/articles/s41586-024-08025-4
  3. Google DeepMind, SynthID overview, deepmind.google/models/synthid/ (accessed Aug 16, 2026)
  4. Google AI developer docs, SynthID safeguards, ai.google.dev/responsible/docs/safeguards/synthid (accessed Aug 16, 2026)
  5. TechCrunch, “Anthropic says it will watermark text generated by its AI models” (Aug 11, 2026) and follow-up detail piece (Aug 15, 2026)
  6. Kirchenbauer et al., “A Watermark for Large Language Models,” ICML 2023, arxiv.org/abs/2301.10226
  7. European Commission, Code of Practice on Transparency of AI-generated Content — policy page and signatory announcement (Jul 2026; accessed Aug 16, 2026)
  8. CityAM, “ChatGPT might follow Claude's watermark pledge — but Grok to swerve it” (Aug 12, 2026)
  9. ReedSmith, analysis of the Art. 50 Commission Guidelines and Code adequacy (accessed Aug 16, 2026)
  10. TechPolicy.Press, “The EU's AI Transparency Code of Practice, Explained” (accessed Aug 16, 2026)
  11. Stibbe, “(Water-)marking the machine” (accessed Aug 16, 2026)