Paragraph matches

First paragraph53 words

Start with ordinary model probabilities

For the unfinished sentence “The sky looked …”, imagine the model assigns vast 40%, endless 30%, blue 20% and grey 10%. These numbers are invented for teaching; Anthropic has not published Claude’s probabilities or tournament parameters. A nonsensical continuation would have negligible probability and would be correspondingly unlikely to enter the candidate draw.

37.5%

Second paragraph93 words

Start with ordinary model probabilities

For the unfinished sentence “The sky looked …”, imagine the model assigns vast 40%, endless 30%, blue 20% and grey 9%. An unrelated token such as toaster has less than 1% probability, with any remaining probability distributed across other tokens. These are teaching numbers—not published Gemini probabilities. They show the crucial boundary: high-probability tokens enter the candidate draw more often, so semantic fit remains controlled by the language model. Simplified from Google DeepMind’s published SynthID-Text reference design. The probabilities, candidate draw and g-values are illustrative; production configurations can differ.

First paragraph79 words

Start with ordinary model probabilities

The key and recent context create repeatable scores Conceptually, the provider combines the secret watermark key with “The sky looked” to seed pseudorandom scores for the sampled tokens. For this toy example, imagine layer-one scores of 0 for vast #1, 1 for endless, 0 for vast #2 and 0 for blue. The seed stays with the provider; it is not inserted into the text and does not encode a user. The key and recent context create repeatable scores

37.2%

Second paragraph60 words

Start with ordinary model probabilities

Turn the key and recent context into scores The watermark key and recent token IDs seed repeatable pseudorandom g-values for each candidate and tournament layer. The key stays with the watermarking system; it is not inserted into the prose. The same word can receive a different value in a different context. Turn the key and recent context into scores

First paragraph40 words

How Claude watermark detection works

Detection is separate from generation. The checker receives the finished passage and the matching secret configuration. It does not recover a hidden key from the prose and does not need to know which losing candidates appeared in the original tournament.

35.5%

Second paragraph59 words

How SynthID-Text detection works

Detection is a separate process. The checker receives a finished passage and the matching watermark configuration. It tokenizes the text, reconstructs the recent context at each eligible position and recomputes the keyed g-values for the token that actually appears. Nothing is decoded from a hidden string, and the underlying language model is not required for the published detector.

First paragraph147 words

A worked detection example

Take the toy continuation “vast and endless, stretching beyond the distant hills.” At each eligible token position, the detector combines the same key with the preceding tokens, recomputes the expected keyed value for the token that actually appears, and records that observation. Imagine 12 of 16 layer-level checks align in this eight-position teaching example. That 12/16 result is illustrative—not a Claude threshold or confidence score. The detector follows four conceptual stages: Tokenize the submitted passage and identify eligible positions using the expected tokenizer and masking rules. At each position, combine the matching secret key with the preceding context to reproduce the pseudorandom keyed value. Score the token that actually appears, then combine observations across many positions. A single high-scoring word proves nothing. Return calibrated statistical evidence or likelihood. Anthropic has not published Claude’s scoring function, threshold or false-positive operating point.

28.1%

Second paragraph154 words

A worked detection example

Imagine a 600-word creative explanation with hundreds of eligible positions. Some observed tokens align with the keyed values more often than ordinary text would be expected to; others are neutral. The detector combines all of those small observations. Eight highlighted positions or a meter in the diagram are purely illustrative—not a Google score threshold, confidence percentage or minimum passage length. A compatible detector follows five conceptual stages: Tokenize the passage with the tokenizer expected by the watermark configuration. Mask positions that should not contribute, including repeated contexts in the published non-distortionary design. For each remaining position, combine the matching key with recent tokens and recompute the g-values for the observed token. Aggregate evidence with a calibrated scoring method. Google’s reference work describes mean, weighted-mean, frequentist and trained Bayesian options. Interpret the result at a threshold chosen for the detector, text length, expected data and acceptable false-positive rate.

First paragraph53 words

Start with ordinary model probabilities

Candidates are drawn from the model distribution Suppose four ordinary draws produce vast, endless, vast and blue. The most probable token can appear more than once, while grey does not enter this particular draw. Quality boundary: every competitor came from the language model’s probability distribution. Candidates are drawn from the model distribution

25.8%

Second paragraph53 words

Start with ordinary model probabilities

Draw candidate copies from that distribution In a toy two-layer tournament, four draws might be vast, blue, endless and grey. Sampling is with replacement, so duplicates are possible even though this illustrative draw contains four different candidates. Every competitor was supplied by the model distribution first. Draw candidate copies from that distribution

First paragraph36 words

Text watermarks and file credentials are different

Anthropic uses “watermark” as an umbrella term for two mechanisms. Text receives a statistical signal created while tokens are generated. Supported PNG, JPG and SVG files receive a small cryptographically signed C2PA content credential in metadata.

24.5%

Second paragraph64 words

What Google SynthID is

The name covers several modality-specific systems. Text carries a statistical pattern in token choices. Images and video carry signals in visual content, while audio carries an inaudible signal in a time-frequency representation. The common idea is embed then detect; the technical carrier is different for each medium. SynthID therefore cannot identify arbitrary AI content that was never watermarked with a compatible system.

First paragraph84 words

Start with ordinary model probabilities

The final winner becomes the next token In the toy second layer, imagine endless scores 0 and vast scores 1. Vast wins, so the sentence becomes “The sky looked vast …”. A different seed or candidate draw could produce a different sensible continuation. The final winner becomes the next token In the toy second layer, imagine endless scores 0 and vast scores 1. Vast wins, so the sentence becomes “The sky looked vast …”. A different seed or candidate draw could produce a different sensible continuation.

21.5%

Second paragraph37 words

Start with ordinary model probabilities

Suppose vast wins, completing “The sky looked vast …”. New preceding context creates new keyed values for the next position. Many such choices form the passage-level signal. One selected token contains almost no useful evidence by itself.

First paragraph42 words

How the signal builds across a passage

One selected word contains almost no useful evidence. Consider a 180-word product explanation drafted entirely by a supported Claude model: dozens of eligible token choices may contribute small key-dependent signals. Those observations can accumulate into evidence a detector can score.

16.6%

Second paragraph37 words

Start with ordinary model probabilities

Suppose vast wins, completing “The sky looked vast …”. New preceding context creates new keyed values for the next position. Many such choices form the passage-level signal. One selected token contains almost no useful evidence by itself.

First paragraph128 words

What a result can and cannot prove

Watermark evidence is asymmetric. A positive result can support a limited provenance conclusion, while a missing mark is often inconclusive. How to interpret Claude watermark and C2PA results Observation Reasonable conclusion What it does not prove Claude text mark detected Claude was likely involved in generating or substantially processing some of the passage. Who authored the ideas, who used Claude, ownership, intent or misconduct. No Claude text mark detected No reliable positive Claude signal was found. Human authorship or the absence of AI involvement. Valid Claude C2PA credential A signed Claude provenance claim validates for this supported asset. The user’s identity, ownership of the concept or truthfulness of the content. No C2PA credential No readable supported credential is present. That Claude or another AI was never involved.

8.3%

Second paragraph198 words

What a SynthID result can and cannot prove

Watermark evidence is asymmetric. A detected compatible signal supports a limited inference about a supported generator or edit. A missing signal often leaves several explanations open: no compatible watermark was applied, the evidence was too weak, the asset was transformed, the wrong detector was used or another AI system created it. How to interpret SynthID results Observation Reasonable conclusion What it does not prove Media SynthID mark detected All or part of the asset was likely generated or edited by a supported Google AI system recognized by the checker. Who prompted it, who owns it, whether it is true or whether every part is synthetic. Text score exceeds a configured threshold The passage shows statistical evidence consistent with that compatible SynthID-Text configuration. That Gemini wrote every idea, that a named person used it or that misconduct occurred. No mark found The checker found no reliable positive signal under its supported configuration. Human origin, absence of AI involvement or absence of a different provider’s watermark. Only a region, segment or track is detected The marked evidence is localized to that part of the asset. That the rest of the image, video or audio is synthetic—or authentic.

First paragraph36 words

Common questions about Claude watermarking

No. Anthropic says Claude uses a version of SynthID-Text. The mark comes from a key-dependent statistical pattern across ordinary token choices; it does not append zero-width characters, a hidden string or extra tokens.

2.8%

Second paragraph29 words

Common questions about Google SynthID

No. SynthID-Text changes the statistical pattern of token sampling during generation. The watermark is not a zero-width character string, readable code, user ID or ordinary file metadata.

First paragraph46 words

Common questions about Claude watermarking

No. Anthropic says supported future models will mark eligible output, models launched in the EU on or after 2 August 2026 support marking from launch, and older models are being updated over the coming months. Very short or constrained output may also contain too little signal.

2.2%

Second paragraph34 words

Common questions about Google SynthID

No. A generative watermark must be applied by a compatible model or service during generation. SynthID-Text detection also requires the matching configuration. It is not a universal classifier for all AI-written prose.

First paragraph96 words

Common questions about Claude watermarking

No. The Claude Code attribution is a visible, configurable Git commit trailer or pull-request footer. The text watermark is an invisible statistical pattern in supported generated text. They can coexist, but they are different mechanisms. How Claude’s text watermark works - Anthropic Code of Practice on Transparency of AI-generated Content: FAQ - European Commission C2PA Technical Specification 2.4 - C2PA How Claude marks AI-generated content - Anthropic Scalable watermarking for identifying large language model outputs - Nature Regulation (EU) 2024/1689, Article 50 - European Union C2PA Security Considerations 2.4 - C2PA Claude Code attribution settings - Anthropic

1.0%

Second paragraph29 words

Common questions about Google SynthID

No. SynthID-Text changes the statistical pattern of token sampling during generation. The watermark is not a zero-width character string, readable code, user ID or ordinary file metadata.