Forensics · Detection

AI watermark detection: what can actually be verified?

Search results and product marketing often collapse three unrelated technologies into one “AI detector.” Understanding the distinction is the foundation of any serious detection workflow.

Updated 14 Aug 202612 min readPrimary-source based

Short answer: people usually mean one of three things: a generic AI-text classifier that estimates whether prose resembles model output; a keyed watermark detector that tests for a deliberately embedded statistical signal; or a C2PA provenance verifier that checks cryptographically signed claims bound to an asset. Each returns a different kind of evidence with different failure modes.

AI detector vs watermark detector vs provenance verifier

The same input text can produce contradictory labels depending on which system is used. Treating them as interchangeable leads to false confidence and bad policy.

SystemWhat it checksNeeds secret/provider data?Result typeMain limitation
Generic AI classifierWriting/content patterns, perplexity, stylistic featuresNoProbability or classificationNot provenance; unrelated error modes
Statistical watermark detectorDeliberately embedded token-selection signalOften yes (key, rule, threshold)Statistical evidence (score, p-value)Key-, length- and threshold-dependent
C2PA verifierSigned provenance claims and manifest bindingPublic trust lists and verification dataCryptographic pass/fail on signaturesBinding can be lost; claims ≠ truth

Do not equate them. A generic “AI detector” that flags Claude-like prose is not automatically detecting Anthropic's secret watermark. A valid C2PA signature on a file does not prove that copied plain text carries the same provenance chain.

Detecting statistical LLM watermarks

Generation-time text watermarks work by introducing a secret, keyed relationship between preceding context and token selection. At each position the decoder subtly prefers tokens consistent with a hidden rule. A detector that knows the rule recomputes it over observed text and aggregates weak evidence across many positions.

How detection accumulates evidence

In the canonical Kirchenbauer green-list construction, a fraction γ of the vocabulary is designated “green” for each context. The detector counts green tokens and compares against the expected base rate. The statistic grows with usable token count — longer, higher-entropy passages generally provide more observations.

z = (NG − γT) / √(Tγ(1−γ))

Production systems such as SynthID Text use different sampling rules (Tournament Sampling) but the principle is similar: aggregate correlation with keyed scores rather than visible markers.

False positives and false negatives

Statistical detectors operate at chosen thresholds. A low false-positive rate is essential at scale — even 0.1% specificity produces enormous numbers of incorrect flags across billions of documents. Short text, low-entropy passages, mixed human/AI documents and heavily edited content all increase false-negative risk.

A serious detector should expose calibrated evidence — a score, confidence band or p-value — rather than a binary “AI / human” verdict without context.

LLM watermarking primer → · Kirchenbauer / KGW guide → · SynthID Text →

Can Claude's watermark be detected?

Confirmed Anthropic says supported Claude models embed an imperceptible watermark directly into generated text at model level, and that it travels with copy/paste.

Confirmed Anthropic says it is working to enable users and third parties to detect supported marks and will publish technical documentation describing the detection mechanisms.

Unknown As of August 2026, Anthropic has not publicly released the generation algorithm, detector statistic, verification key or a general-purpose third-party detection API for Claude text watermarks.

Important: unless a tool uses Anthropic's published detector, approved API or verification key, it is most likely performing generic AI-text classification — not Claude watermark verification. Do not treat unofficial “Claude detectors” as equivalent to an authorized watermark check.

Anthropic also notes that heavy editing, paraphrasing, translation, mixing and very short passages can prevent reliable detection — consistent with statistical aggregation over many token positions.

Read the full Claude watermark analysis →

SynthID detection

SynthID is a family of watermarking systems from Google DeepMind. The detection approach depends on modality:

SynthID Text

Google describes text watermarking via keyed, context-dependent token scores and Tournament Sampling during generation. Detection recomputes expected scores and measures aggregate correlation. Google reports production use in the Gemini app and web experience. Detection works best on longer, diverse outputs.

SynthID Text guide →

SynthID for images, audio and video

Google's image, audio and video SynthID systems use modality-specific pixel- or sample-domain embedding mechanisms — not the same construction as SynthID Text. Metadata stripping on a container file does not remove a pixel-domain watermark embedded in image data.

Inference Unofficial reverse-engineered scoring tools may exist for research purposes. They should not be presented as equivalent to Google's official detector or verification infrastructure.

C2PA verification

C2PA (Coalition for Content Provenance and Authenticity) defines Content Credentials: cryptographically signed manifests that record provenance assertions about a digital asset. Verification is fundamentally different from probabilistic watermark detection.

  1. Signature verification — check that claims were signed by a credential in the trust model.
  2. Provenance manifest — read assertions about creation, editing and AI involvement.
  3. Trust chain — validate issuer certificates against configured trust lists.
  4. Asset binding — confirm the manifest is bound to the content (hard binding) or reconnect via soft-binding fingerprints where supported.

C2PA answers “does this asset carry valid signed provenance claims?” — not “does this plain text statistically resemble AI output?” Copy/paste of text usually severs the file-level provenance binding even when an embedded text watermark survives.

C2PA vs watermarking comparison →

How to interpret a detection result

ResultReasonable interpretationWhat it does NOT prove
Positive watermark signalText is statistically consistent with a particular watermarking system at the chosen threshold.Full authorship, intent, or that every word came from that model.
Negative resultNo supported watermark signal was recovered at the chosen threshold.That content is human-written or never touched an AI system.
Weak / borderline resultInsufficient evidence — passage may be short, edited, mixed or low-entropy.Reliable attribution without additional evidence.
Valid C2PA credentialSigned provenance claims authenticate under the trust model; binding intact.That semantic content is truthful or that copied text preserves the chain.
Missing C2PA credentialNo verifiable signed provenance was found on this asset.That content is fake, unmodified, or free of any in-content watermark.

AI watermark detection FAQ

Can I detect a Claude watermark?

Anthropic says third-party detection is planned and technical documentation will follow. As of August 2026 there is no public Claude watermark detector or verification API. Unofficial tools are unlikely to detect Anthropic's actual mark.

Is an AI detector the same as a watermark detector?

No. Generic AI detectors classify writing patterns. Watermark detectors test for a signal deliberately embedded during generation by a cooperating model and key holder.

Can a watermark prove which model wrote a document?

Only if the scheme supports attribution keys and the detector is calibrated for that purpose. Even then, a positive result is statistical evidence about a marking process — not automatic proof of authorship.

Can paraphrasing break detection?

Strong paraphrasing, translation and regeneration can substantially weaken token-level watermarks by replacing many original token choices. Robustness is scheme- and threshold-dependent.

Does C2PA detect AI-generated text?

C2PA verifies signed provenance claims, which may include assertions that content was AI-generated. It does not perform statistical watermark detection on plain text detached from its signed manifest.

Related ZK-F guides