Lab

AI Text Watermarking Lab

Generate, detect, and edit a SynthID-style statistical watermark for AI text, entirely in your browser.

Research model, not a Claude detector

This lab uses a visitor-selected key and a small, transparent tournament-style generator. It teaches the mechanics of statistical text watermarking, but it does not know Anthropic's production key or detect Claude's watermark.

What this is

A keyed generation and detection experiment. The generator chooses among plausible words using the preceding context and your key. The detector recomputes those choices and asks whether the observed pattern is unlikely under a 50/50 null model.

What it demonstrates

  • The matching key reveals a distributed statistical signal.
  • A different key makes the same text look like ordinary noise.
  • Editing and shortening reduce the amount of usable evidence.

Step 1

Choose the experiment key

The key stays in this browser tab and is not saved or transmitted.

Key fingerprint
No key selected

Tournament depth: each layer doubles the candidate draws and adds another keyed knockout round. Three layers means 8 draws → 4 winners → 2 finalists → 1 selected word.

Step 2

Generate or edit a sample

No sample

Select a key, then generate a keyed or control sample.

Step 3

Detection result

Waiting for a sample

Generate or paste text and analyze it with the selected lab key.

z = 0 strong threshold · z = 3 z = 8+
Z-score
One-sided p-value
Positive g-values
Usable evidence

Token evidence map

positive mixed negative ineligible

What the key does

A color is not permanently attached to a word. For every eligible word, the detector combines four inputs and computes one binary g-value per tournament layer:

key + previous words + current word + layer → 0 or 1

Changing the key changes those bits, so the same passage gets a different color pattern. The preceding words matter too: the same word can be green in one sentence and red in another.

How to read the colors

Green
At least two-thirds of its keyed bits are 1.
Amber
Its keyed bits are mixed between 0 and 1.
Red
At most one-third of its keyed bits are 1.
Gray
Not scored: it is outside the lab vocabulary or repeats a context.

Red words are expected. The generator biases the whole passage toward candidates with more positive bits; it does not force every word to be green. Detection comes from the cumulative excess of positive bits across many eligible words, compared with the 50% null expectation.

Token-level evidence appears here after analysis.

Word inspector

hover · focus · tap
Normalized preceding context → selected word Select a colored word

Analyze a passage, then inspect a colored word to compare its alternatives.

Future integration

Claude production watermark

The adapter is intentionally unavailable until Anthropic publishes an official detector API and response contract.

How this generator works

At each flexible position, the generator has several plausible next words. It hashes your key, the preceding context, the candidate word, and a layer number to produce binary g-values. Candidates compete in pairs: a candidate with g = 1 defeats one with g = 0, while ties are resolved randomly. Each new round uses the next layer's g-value until one candidate remains. A control sample chooses from the same alternatives without consulting the key.

The detector repeats that calculation and tests the observed positive-g rate against an expected rate of 50%. Repeated contexts are counted only once. This is intentionally readable teaching code, not Google's or Anthropic's production implementation.

Depth Candidate draws Knockout rounds Bits scored per word
1 layer 2 1 1
2 layers 4 2 2
3 layers 8 3 3
4 layers 16 4 4

More layers create stronger selection pressure and give the detector more keyed observations per eligible word. They do not mean extra writing passes or visible layers in the final text.

Try the key-swap experiment

  1. Generate a keyed sample and note its colors and z-score.
  2. Change one character in the key without changing the text.
  3. Analyze again: the colors reshuffle and the aggregate signal falls.
  4. Restore the original key and analyze once more to recover the signal.
Can I watermark text I already wrote?

Yes, but the watermark is added through word choice—not by attaching a hidden character or file marker. You need to rewrite enough flexible parts of the text using alternatives favored by your key.

  1. Choose a word or phrase that can change without changing the meaning.
  2. List several natural alternatives that fit the sentence.
  3. Score each alternative using the key, preceding context, candidate, and layer.
  4. Prefer a higher-scoring alternative that still sounds natural.
  5. Continue from left to right and periodically analyze the complete passage.

The team inspected the system.

inspected · 1/3 positive reviewed · 3/3 positive examined · 2/3 positive studied · 0/3 positive

In this hypothetical choice, the key favors “reviewed.” With another key or preceding context, the ranking would probably change.

Human judgment remains important: the highest-scoring replacement may be grammatically valid but wrong for the intended tone or meaning. This lab's generator already has curated alternatives; it does not currently create alternatives for arbitrary pasted prose.

Browse the lab's choice vocabulary

These are the generator's actual context-specific choice groups—not a general thesaurus. At each flexible position, the generator samples from one row and uses the key to favor candidates with stronger g-values.

54 choice groups · 189 unique words

Choice 1
review assessment analysis evaluation
Choice 2
examined studied considered inspected
Choice 3
service system platform application
Choice 4
several multiple various some
Choice 5
implicit unstated informal unverified
Choice 6
mapped traced documented charted
Choice 7
changed revised adjusted updated
Choice 8
design plan model approach
Choice 9
initial first early opening
Choice 10
small limited compact focused
Choice 11
inspect review examine study
Choice 12
later subsequent follow-up second
Choice 13
additional extra new further
Choice 14
responded behaved performed reacted
Choice 15
proof certainty confirmation fact
Choice 16
chosen selected configured declared
Choice 17
stronger clearer better greater
Choice 18
many numerous repeated additional
Choice 19
fewer limited scarcer reduced
Choice 20
weaker harder uncertain fragile
Choice 21
separate distinct isolated apart
Choice 22
settings parameters configuration options
Choice 23
returned fell moved settled
Choice 24
expected ordinary baseline normal
Choice 25
reduced weakened diluted lowered
Choice 26
difficult hard unlikely impractical
Choice 27
little minimal limited negligible
Choice 28
characters spacing markup metadata
Choice 29
usable eligible available meaningful
Choice 30
interpret understand evaluate judge
Choice 31
locally privately entirely directly
Choice 32
retain store upload transmit
Choice 33
simple small readable compact
Choice 34
production provider commercial deployed
Choice 35
supported documented stable verified
Choice 36
calibrated interpreted reported validated
Choice 37
disabled unavailable separate inactive
Choice 38
simulation demonstration model prototype
Choice 39
different new changed separate
Choice 40
scores measurements values results
Choice 41
visible concrete clear tangible
Choice 42
sounds appears reads feels
Choice 43
authorship style origin provenance
Choice 44
keyed statistical configured watermark
Choice 45
practical hands-on useful focused
Choice 46
private secret production internal
Choice 47
unusual strong notable elevated
Choice 48
particular selected given specific
Choice 49
observable interactive measurable accessible
Choice 50
explicit visible documented clear
Choice 51
multiple several different varied
Choice 52
broad general strong confident
Choice 53
reproduce repeat verify audit
Choice 54
inputs parameters settings configuration
Common questions about the signal

Is one green word evidence?

No. Any individual word can be green by chance. The evidence is the excess of positive bits accumulated over many eligible words.

Is there a secret list of green words?

No. Color depends on the key, preceding context, word, and layer. A word that is green here can be red somewhere else.

Where do the vocabulary and alternatives come from?

They are hand-curated in this lab's source code. Each flexible position has a small choice(...) group of plausible alternatives, such as “review,” “assessment,” “analysis,” and “evaluation.” The eligible vocabulary is the union of those groups. The lab does not call a thesaurus or language model, and the alternatives are context-specific rather than guaranteed synonyms everywhere.

Why does changing one word affect later words?

The preceding words are part of each calculation. An edit can change the context—and therefore the colors—for several following tokens.

How much text is enough?

This lab requires at least 80 scored observations. Observations are keyed bits, not total words, so the required passage length depends on how many words are eligible and how many layers you select.

What happens if I edit the text?

Formatting alone usually changes little here. Replacing or removing scored words can weaken or reshape the signal, while context-changing edits can affect later scores too. Reanalyze after each edit to measure the result.

Does the key need to stay secret?

Not for this teaching exercise. In a real provenance system, the embedding key would normally be controlled by the provider: anyone who has the key and algorithm could generate text that matches its signal.

Can the lab detect AI writing without the key?

No. It verifies agreement with this lab's key and vocabulary model. It does not estimate whether arbitrary prose was written by AI, Claude, or a person.

Reading the statistics

The z-score measures how far the observed positive-g rate sits above the null expectation in standard-error units. The one-sided p-value asks how often an unwatermarked sample would produce a score at least this large under this lab's simplified assumptions. The lab calls z ≥ 3 strong evidence and z ≥ 2 weak evidence; fewer than 80 observations is reported as insufficient.

Limits and privacy
  • The eligible vocabulary is specific to this built-in research generator.
  • A score on arbitrary pasted prose is not an AI-authorship judgment.
  • The educational hash is deterministic but is not a cryptographic PRF.
  • Your text, key, and results remain in the browser and are not persisted.
  • The future Claude adapter currently sends no network request.
Primary references