Deep dives / The Word-Choice Watermark
The Word-Choice Watermark
Claude and Gemini now mark the text they write by tilting which words they pick.
We built the same kind of mark with our own key, put 150 web pages through 18 kinds of handling, and measured what survives. This guide shows how the mark works, what removed it in our tests, who can check for it, and what it means for search.
One of these is watermarked. Which one?
Text A Likely watermarked · z 3.8 No signal · z 0.8
Next, consider the specific type of cushioning you prefer. Some runners thrive on soft, maximalist cushioning
Once you understand your foot mechanics, you can start narrowing down the shoe types that suit you best. Cushioning
Text B Likely watermarked · z 3.8 No signal · z 0.8
Next, consider the specific type of cushioning you prefer. Some runners thrive on soft, maximalist cushioning
Once you understand your foot mechanics, you can start narrowing down the shoe types that suit you best. Cushioning
Both came from the same model and the same prompt. Only one was written with our secret key.
You picked Text A. Honestly, nobody can tell by reading. Apply the key to find out.
You picked Text B. Honestly, nobody can tell by reading. Apply the key to find out.
You guessed right, and it was a coin flip. Not that one, and that's the point. Text A carries the watermark. On these 150 tokens z = 3.8; on the full 446-token article it is 6.8. The other text, in full, scores 0.55.
You guessed right, and it was a coin flip. Not that one, and that's the point. Text B carries the watermark. On these 150 tokens z = 3.8; on the full 446-token article it is 6.8. The other text, in full, scores 0.55.
Our pilot · 26 Sep 2026 Gemma 4 E2B, one 400-word blog prompt, written twice; teal marks a token whose keyed score is above 0.5, faded text one below.
Teal marks what we measured ourselves. Amber marks what others report. Toy marks a simplified mechanism. Outside facts are as of 26 Sep 2026 unless dated otherwise.
Chapter 01 · The mechanism
Every word is a choice
Any text watermark hides in one step: the weighted dice roll a model makes for every word. Some text leaves far more room for it than other text.
How a model picks a word
Weighted dice
-
What Gemma 4 expects next
Our measurement · 29 Sep 2026Search traffic fell after the
the word the model rolled forced: top choice over 90% (5 of 17) Tap a word to see its oddsGemma 4 E2B, our measurement, 29 Sep 2026. The sentence was sampled with stopping turned off for 16 tokens, so at "Users" the model's top choice is to stop writing.
01 / 04
What the model expects next
We give Google's open Gemma 4 model the words “Search traffic fell after the” and read what it expects next. It gives every one of its 262,144 tokens a probability. The eight shown hold 66% of the odds.4
-
What Gemma 4 expects next
Our measurement · 29 Sep 2026Search traffic fell after the
the word the model rolled forced: top choice over 90% (5 of 17) Tap a word to see its oddsGemma 4 E2B, our measurement, 29 Sep 2026. The sentence was sampled with stopping turned off for 16 tokens, so at "Users" the model's top choice is to stop writing.
02 / 04
A roll, not the top word
The model does not take the top word. It rolls weighted dice, and here it picks “platform”, a 2.9% outcome outside the top eight, over “update” at 17.9%. A different roll gives a different, equally good sentence.4
-
What Gemma 4 expects next
Our measurement · 29 Sep 2026Search traffic fell after the
the word the model rolled forced: top choice over 90% (5 of 17) Tap a word to see its oddsGemma 4 E2B, our measurement, 29 Sep 2026. The sentence was sampled with stopping turned off for 16 tokens, so at "Users" the model's top choice is to stop writing.
-
What Gemma 4 expects next
Our measurement · 29 Sep 2026Search traffic fell after the
the word the model rolled forced: top choice over 90% (5 of 17) Tap a word to see its oddsGemma 4 E2B, our measurement, 29 Sep 2026. The sentence was sampled with stopping turned off for 16 tokens, so at "Users" the model's top choice is to stop writing.
04 / 04
Some rolls are not choices
Some rolls are not real choices. After “switched”, the word “to” holds 100% of the odds. Five of the 17 words were over 90% certain, and a watermark can only hide in the other twelve.4
Where a mark has space
Room to hide a signal
-
Room per token, in bits
Our measurement · 27 Sep 2026Juneteenth serves as a profound milestone in the American narrative, marking the moment when the promise of freedom finally reached the shores of Texas, long …1.07 bits per token on average (tokens 2 to 40; the first has no context)
In an era defined by rapid shifts in consumer behavior, the ability to transform vast oceans of data into precise, strategic direction is the ultimate competitive …1.15 bits per token on average (tokens 2 to 40; the first has no context)
<div class="wp-block-group alignfull has-background"><div class="wp-block-group__inner-container is …0.59 bits per token on average (tokens 2 to 40; the first has no context)
{"@context":"https://schema.org","@graph":[{"@type":"WebPage","@id":"https://www.example. …0.52 bits per token on average (tokens 2 to 40; the first has no context)
Guide 53 1.33Case study 40 1.24Promo / landing page 36 1.22Blog post 56 1.20Service page 59 1.20HTML markup 120 0.65Schema markup (JSON-LD) 120 0.290.00.40.81.21.6Range chart: one real company's website, a market research firm, not named; page counts after each label. Scored by Gemma 4 E2B, 27 Sep 2026. Samples: the first 28 tokens of two pages our model wrote in the main study and two hand-written markup samples; grey columns are forced tokens (top choice over 90%).
01 / 04
Measuring the room
Researchers measure that room as entropy, in bits per token: the more good options at each word, the more bits. We scored 244 prose pages from one real company's website, plus the markup of 120 of them.3
-
Room per token, in bits
Our measurement · 27 Sep 2026Juneteenth serves as a profound milestone in the American narrative, marking the moment when the promise of freedom finally reached the shores of Texas, long …1.07 bits per token on average (tokens 2 to 40; the first has no context)
In an era defined by rapid shifts in consumer behavior, the ability to transform vast oceans of data into precise, strategic direction is the ultimate competitive …1.15 bits per token on average (tokens 2 to 40; the first has no context)
<div class="wp-block-group alignfull has-background"><div class="wp-block-group__inner-container is …0.59 bits per token on average (tokens 2 to 40; the first has no context)
{"@context":"https://schema.org","@graph":[{"@type":"WebPage","@id":"https://www.example. …0.52 bits per token on average (tokens 2 to 40; the first has no context)
Guide 53 1.33Case study 40 1.24Promo / landing page 36 1.22Blog post 56 1.20Service page 59 1.20HTML markup 120 0.65Schema markup (JSON-LD) 120 0.290.00.40.81.21.6Range chart: one real company's website, a market research firm, not named; page counts after each label. Scored by Gemma 4 E2B, 27 Sep 2026. Samples: the first 28 tokens of two pages our model wrote in the main study and two hand-written markup samples; grey columns are forced tokens (top choice over 90%).
02 / 04
Prose lands close together
All five kinds of prose land close together. Medians run from 1.20 bits for blog posts and service pages to 1.33 for guides, and pages vary more than page types do.3
-
Room per token, in bits
Our measurement · 27 Sep 2026Juneteenth serves as a profound milestone in the American narrative, marking the moment when the promise of freedom finally reached the shores of Texas, long …1.07 bits per token on average (tokens 2 to 40; the first has no context)
In an era defined by rapid shifts in consumer behavior, the ability to transform vast oceans of data into precise, strategic direction is the ultimate competitive …1.15 bits per token on average (tokens 2 to 40; the first has no context)
<div class="wp-block-group alignfull has-background"><div class="wp-block-group__inner-container is …0.59 bits per token on average (tokens 2 to 40; the first has no context)
{"@context":"https://schema.org","@graph":[{"@type":"WebPage","@id":"https://www.example. …0.52 bits per token on average (tokens 2 to 40; the first has no context)
Guide 53 1.33Case study 40 1.24Promo / landing page 36 1.22Blog post 56 1.20Service page 59 1.20HTML markup 120 0.65Schema markup (JSON-LD) 120 0.290.00.40.81.21.6Range chart: one real company's website, a market research firm, not named; page counts after each label. Scored by Gemma 4 E2B, 27 Sep 2026. Samples: the first 28 tokens of two pages our model wrote in the main study and two hand-written markup samples; grey columns are forced tokens (top choice over 90%).
03 / 04
HTML leaves about half
The HTML around that prose leaves about half the room: 0.65 bits. Tags and class names decide most of the characters.3
-
Room per token, in bits
Our measurement · 27 Sep 2026Juneteenth serves as a profound milestone in the American narrative, marking the moment when the promise of freedom finally reached the shores of Texas, long …1.07 bits per token on average (tokens 2 to 40; the first has no context)
In an era defined by rapid shifts in consumer behavior, the ability to transform vast oceans of data into precise, strategic direction is the ultimate competitive …1.15 bits per token on average (tokens 2 to 40; the first has no context)
<div class="wp-block-group alignfull has-background"><div class="wp-block-group__inner-container is …0.59 bits per token on average (tokens 2 to 40; the first has no context)
{"@context":"https://schema.org","@graph":[{"@type":"WebPage","@id":"https://www.example. …0.52 bits per token on average (tokens 2 to 40; the first has no context)
Guide 53 1.33Case study 40 1.24Promo / landing page 36 1.22Blog post 56 1.20Service page 59 1.20HTML markup 120 0.65Schema markup (JSON-LD) 120 0.290.00.40.81.21.6Range chart: one real company's website, a market research firm, not named; page counts after each label. Scored by Gemma 4 E2B, 27 Sep 2026. Samples: the first 28 tokens of two pages our model wrote in the main study and two hand-written markup samples; grey columns are forced tokens (top choice over 90%).
Chapter 02 · The method
Written in, read out
The method deployed in Gemini and Claude is a keyed tournament, and only the key-holder can read it back. Here it is, with a real article from our pilot.
Every text watermark uses a secret key to bias the dice in a pattern only the key-holder can check. Scott Aaronson described a keyed scheme for OpenAI in 2022 that leaves the odds unchanged on average.10 The first widely cited paper, in 2023, favours a keyed “green list” of words at each position.9 Research schemes such as SemStamp mark sentence meaning instead of words, and no major product used one as of 26 Sep 2026.11 The method in Gemini, and in a version in Claude, is a tournament.7,5
How the mark is written
The tournament (toy)
-
Tournament for the word after “the platform”
Toy: 8 candidates, 3 roundsOdds measured · 29 Sep 2026secret key + the previous four tokens → a score of 0 or 1 for each candidate, fresh every roundMeasured odds update 27%experienced 17%announced 13%was 11%introduced 4%underwent 4%launched 3%faced 3%
Equal scores: a coin flip decides.
Search traffic fell after the platform experienced …
After “switched” totototototototo → to 100% of the odds: no contest
Our real key 30 rounds · 5-token context
Toy mechanism with measured odds (Gemma 4 E2B, 29 Sep 2026). The scores here come from a demo key; the real scheme is Dathathri et al., Nature, 23 Oct 2024.
01 / 04
Draw candidates
For each word, the model draws several candidates from its normal odds. After “the platform”, “update” is the most likely at 27%, so it can fill more than one slot.4
-
Tournament for the word after “the platform”
Toy: 8 candidates, 3 roundsOdds measured · 29 Sep 2026secret key + the previous four tokens → a score of 0 or 1 for each candidate, fresh every roundMeasured odds update 27%experienced 17%announced 13%was 11%introduced 4%underwent 4%launched 3%faced 3%
Equal scores: a coin flip decides.
Search traffic fell after the platform experienced …
After “switched” totototototototo → to 100% of the odds: no contest
Our real key 30 rounds · 5-token context
Toy mechanism with measured odds (Gemma 4 E2B, 29 Sep 2026). The scores here come from a demo key; the real scheme is Dathathri et al., Nature, 23 Oct 2024.
02 / 04
The key scores each candidate
The secret key, combined with the previous four tokens, gives each candidate a score of 0 or 1 for this round. Pairs face off; the higher score advances, and a tie is a coin flip.7
-
Tournament for the word after “the platform”
Toy: 8 candidates, 3 roundsOdds measured · 29 Sep 2026secret key + the previous four tokens → a score of 0 or 1 for each candidate, fresh every roundMeasured odds update 27%experienced 17%announced 13%was 11%introduced 4%underwent 4%launched 3%faced 3%
Equal scores: a coin flip decides.
Search traffic fell after the platform experienced …
After “switched” totototototototo → to 100% of the odds: no contest
Our real key 30 rounds · 5-token context
Toy mechanism with measured odds (Gemma 4 E2B, 29 Sep 2026). The scores here come from a demo key; the real scheme is Dathathri et al., Nature, 23 Oct 2024.
03 / 04
Rounds to a champion
Each round uses fresh scores. The winner is still a word the model found plausible, which is why Google reports no significant quality difference across about 20 million Gemini responses.7
-
Tournament for the word after “the platform”
Toy: 8 candidates, 3 roundsOdds measured · 29 Sep 2026secret key + the previous four tokens → a score of 0 or 1 for each candidate, fresh every roundMeasured odds update 27%experienced 17%announced 13%was 11%introduced 4%underwent 4%launched 3%faced 3%
Equal scores: a coin flip decides.
Search traffic fell after the platform experienced …
After “switched” totototototototo → to 100% of the odds: no contest
Our real key 30 rounds · 5-token context
Toy mechanism with measured odds (Gemma 4 E2B, 29 Sep 2026). The scores here come from a demo key; the real scheme is Dathathri et al., Nature, 23 Oct 2024.
04 / 04
Thirty rounds, and forced words
Our key runs 30 rounds; this toy shows 3. A word with no real alternative skips the contest, because every candidate is the same word.1
How the mark is read
Reading the mark
-
Running average score, token by token
Our pilot · 26 Sep 202699% noise band detection line, z 2.33 Watermarked, our keyWritten without the keyWatermarked, read with the wrong keyPilot article blog-02, Gemma 4 E2B, our key, 26 Sep 2026. The score is the running mean of each token's keyed score; 0.5 means no signal.
01 / 05
Replay the scoring
The key-holder replays the scoring on a finished text. Ordinary words average 0.5, like coin flips. This is one article from our pilot, 442 scored tokens, written with our key.1
-
Running average score, token by token
Our pilot · 26 Sep 202699% noise band detection line, z 2.33 Watermarked, our keyWritten without the keyWatermarked, read with the wrong keyPilot article blog-02, Gemma 4 E2B, our key, 26 Sep 2026. The score is the running mean of each token's keyed score; 0.5 means no signal.
02 / 05
The watermarked article
The watermarked article settles at 0.534 and clears the detection line with a z-score of 6.8.1
-
Running average score, token by token
Our pilot · 26 Sep 202699% noise band detection line, z 2.33 Watermarked, our keyWritten without the keyWatermarked, read with the wrong keyPilot article blog-02, Gemma 4 E2B, our key, 26 Sep 2026. The score is the running mean of each token's keyed score; 0.5 means no signal.
03 / 05
Written without the key
The same prompt written without the key stays inside the band, at z 0.55.1
-
Running average score, token by token
Our pilot · 26 Sep 202699% noise band detection line, z 2.33 Watermarked, our keyWritten without the keyWatermarked, read with the wrong keyPilot article blog-02, Gemma 4 E2B, our key, 26 Sep 2026. The score is the running mean of each token's keyed score; 0.5 means no signal.
04 / 05
Read with the wrong key
Scored with the wrong key, the watermarked article looks unmarked too, at z −0.42. The key is the only thing that makes the pattern readable.1
-
Running average score, token by token
Our pilot · 26 Sep 202699% noise band detection line, z 2.33 Watermarked, our keyWritten without the keyWatermarked, read with the wrong keyPilot article blog-02, Gemma 4 E2B, our key, 26 Sep 2026. The score is the running mean of each token's keyed score; 0.5 means no signal.
Chapter 03 · What we measured
Our experiments
Claude's and Gemini's keys are private, so nobody outside those companies can test the deployed marks. We built our own mark with the same open method and a key we made up, then measured what everyday handling does to it. We were not writing a removal guide; we wanted to know where things stand.
Two runs, one key
Setup
-
Two runs, one key
Our runs · 26 Sep 2026 and 28 Sep 2026Pilot · 26 Sep 2026 12 short articles→ Gemma 4 E2B, laptop→ 17 kinds of handling→ detectorMain study · 28 Sep 2026Blog post 29/30Case study 28/30Guide 28/30Promo page 25/30Service page 22/30→ brief, names removed → Gemma 4 26Brented A100 with our keywithout18 kinds of handling Copy and pasteCut to 100 tokensCut to 200 tokensLight edit, asked 5%Light edit, asked 10%Light edit, asked 15%Light edit, asked 20%Light edit, asked 30%Light edit, asked 40%Half mixed with other textAI rewrite, Mistral NemoAI rewrite, GPT-5.5Japanese and backOdd spacing added afterOdd spacing, then cleanedWritten with a symbol after each wordWritten one word per lineWritten in capitalsDetector 1% false alarms by design150pages, each written twice 5,017scored texts 9 of 2,250unwatermarked texts flagged (0.4%) 132 of 150detected untouched (88%)Main study: Gemma 4 26B on a rented A100, 28 Sep 2026; one dot per page, filled when the untouched page is detected. Pilot: Gemma 4 E2B on a laptop, 26 Sep 2026. Key: 30 rounds, 5-token context, temperature 0.8, top-k 40.
-
Two runs, one key
Our runs · 26 Sep 2026 and 28 Sep 2026Pilot · 26 Sep 2026 12 short articles→ Gemma 4 E2B, laptop→ 17 kinds of handling→ detectorMain study · 28 Sep 2026Blog post 29/30Case study 28/30Guide 28/30Promo page 25/30Service page 22/30→ brief, names removed → Gemma 4 26Brented A100 with our keywithout18 kinds of handling Copy and pasteCut to 100 tokensCut to 200 tokensLight edit, asked 5%Light edit, asked 10%Light edit, asked 15%Light edit, asked 20%Light edit, asked 30%Light edit, asked 40%Half mixed with other textAI rewrite, Mistral NemoAI rewrite, GPT-5.5Japanese and backOdd spacing added afterOdd spacing, then cleanedWritten with a symbol after each wordWritten one word per lineWritten in capitalsDetector 1% false alarms by design150pages, each written twice 5,017scored texts 9 of 2,250unwatermarked texts flagged (0.4%) 132 of 150detected untouched (88%)Main study: Gemma 4 26B on a rented A100, 28 Sep 2026; one dot per page, filled when the untouched page is detected. Pilot: Gemma 4 E2B on a laptop, 26 Sep 2026. Key: 30 rounds, 5-token context, temperature 0.8, top-k 40.
02 / 05
Pages modelled on a real site
The 150 pages follow five page types from one real company's website, 30 of each. The model turned each real page into a brief with names removed, then wrote a fresh page twice: once with our key and once without.2
-
Two runs, one key
Our runs · 26 Sep 2026 and 28 Sep 2026Pilot · 26 Sep 2026 12 short articles→ Gemma 4 E2B, laptop→ 17 kinds of handling→ detectorMain study · 28 Sep 2026Blog post 29/30Case study 28/30Guide 28/30Promo page 25/30Service page 22/30→ brief, names removed → Gemma 4 26Brented A100 with our keywithout18 kinds of handling Copy and pasteCut to 100 tokensCut to 200 tokensLight edit, asked 5%Light edit, asked 10%Light edit, asked 15%Light edit, asked 20%Light edit, asked 30%Light edit, asked 40%Half mixed with other textAI rewrite, Mistral NemoAI rewrite, GPT-5.5Japanese and backOdd spacing added afterOdd spacing, then cleanedWritten with a symbol after each wordWritten one word per lineWritten in capitalsDetector 1% false alarms by design150pages, each written twice 5,017scored texts 9 of 2,250unwatermarked texts flagged (0.4%) 132 of 150detected untouched (88%)Main study: Gemma 4 26B on a rented A100, 28 Sep 2026; one dot per page, filled when the untouched page is detected. Pilot: Gemma 4 E2B on a laptop, 26 Sep 2026. Key: 30 rounds, 5-token context, temperature 0.8, top-k 40.
03 / 05
Eighteen kinds of handling
Every page then went through 18 kinds of handling: copy and paste, cutting it short, AI light edits from 5% to 40%, mixing with other text, AI rewrites, a round trip through Japanese, and formatting changes.2
-
Two runs, one key
Our runs · 26 Sep 2026 and 28 Sep 2026Pilot · 26 Sep 2026 12 short articles→ Gemma 4 E2B, laptop→ 17 kinds of handling→ detectorMain study · 28 Sep 2026Blog post 29/30Case study 28/30Guide 28/30Promo page 25/30Service page 22/30→ brief, names removed → Gemma 4 26Brented A100 with our keywithout18 kinds of handling Copy and pasteCut to 100 tokensCut to 200 tokensLight edit, asked 5%Light edit, asked 10%Light edit, asked 15%Light edit, asked 20%Light edit, asked 30%Light edit, asked 40%Half mixed with other textAI rewrite, Mistral NemoAI rewrite, GPT-5.5Japanese and backOdd spacing added afterOdd spacing, then cleanedWritten with a symbol after each wordWritten one word per lineWritten in capitalsDetector 1% false alarms by design150pages, each written twice 5,017scored texts 9 of 2,250unwatermarked texts flagged (0.4%) 132 of 150detected untouched (88%)Main study: Gemma 4 26B on a rented A100, 28 Sep 2026; one dot per page, filled when the untouched page is detected. Pilot: Gemma 4 E2B on a laptop, 26 Sep 2026. Key: 30 rounds, 5-token context, temperature 0.8, top-k 40.
04 / 05
What counts as detected
A text counts as detected when its score clears z 2.33, a 1% false-alarm rate by design. Of 2,250 unwatermarked texts, 9 were flagged: 0.4%.2
-
Two runs, one key
Our runs · 26 Sep 2026 and 28 Sep 2026Pilot · 26 Sep 2026 12 short articles→ Gemma 4 E2B, laptop→ 17 kinds of handling→ detectorMain study · 28 Sep 2026Blog post 29/30Case study 28/30Guide 28/30Promo page 25/30Service page 22/30→ brief, names removed → Gemma 4 26Brented A100 with our keywithout18 kinds of handling Copy and pasteCut to 100 tokensCut to 200 tokensLight edit, asked 5%Light edit, asked 10%Light edit, asked 15%Light edit, asked 20%Light edit, asked 30%Light edit, asked 40%Half mixed with other textAI rewrite, Mistral NemoAI rewrite, GPT-5.5Japanese and backOdd spacing added afterOdd spacing, then cleanedWritten with a symbol after each wordWritten one word per lineWritten in capitalsDetector 1% false alarms by design150pages, each written twice 5,017scored texts 9 of 2,250unwatermarked texts flagged (0.4%) 132 of 150detected untouched (88%)Main study: Gemma 4 26B on a rented A100, 28 Sep 2026; one dot per page, filled when the untouched page is detected. Pilot: Gemma 4 E2B on a laptop, 26 Sep 2026. Key: 30 rounds, 5-token context, temperature 0.8, top-k 40.
05 / 05
Untouched: 88%
Untouched, the detector finds our mark in 132 of 150 pages, 88%. The 18 misses are concentrated in service pages, 8 of 30.2
Detection after each kind of handling
What removed it
-
Share of watermarked pages still detected
Our study · 28 Sep 20260%50%100%AI judge, 0 to 10Control
Untouched 88%Copy and paste 88%Length
Cut to 200 tokens 67%Cut to 100 tokens 29%AI light edits
Asked 5%6% changed 65% 9.9Asked 10%10% changed 45% 9.9Asked 15%13% changed 36% 9.9Asked 20%17% changed 17% 9.8Asked 30%26% changed 3.3% 9.7Asked 40%31% changed 3.3% 9.5Mixing
Half mixed with other text 38%AI rewrite
AI rewrite, Mistral Nemo 5.3% 9.5AI rewrite, GPT-5.5 (67 pages) 0% 9.7Translation
Japanese and back 2.7% 9.6Formatting
5 kinds, from added spacing to writing in capitals: see “Removal without rewording”
150 watermarked pages (GPT-5.5 rewrite: 67), Gemma 4 26B, our key, 28 Sep 2026. Whiskers: 95% intervals. Quality: one AI judge (DeepSeek V4 Flash) scoring the changed text against the original.
01 / 05
Copy and paste
Copy and paste changes nothing: 88% detected, the same as untouched. The mark is the word choice itself, so there is nothing hidden to lose.2
-
Share of watermarked pages still detected
Our study · 28 Sep 20260%50%100%AI judge, 0 to 10Control
Untouched 88%Copy and paste 88%Length
Cut to 200 tokens 67%Cut to 100 tokens 29%AI light edits
Asked 5%6% changed 65% 9.9Asked 10%10% changed 45% 9.9Asked 15%13% changed 36% 9.9Asked 20%17% changed 17% 9.8Asked 30%26% changed 3.3% 9.7Asked 40%31% changed 3.3% 9.5Mixing
Half mixed with other text 38%AI rewrite
AI rewrite, Mistral Nemo 5.3% 9.5AI rewrite, GPT-5.5 (67 pages) 0% 9.7Translation
Japanese and back 2.7% 9.6Formatting
5 kinds, from added spacing to writing in capitals: see “Removal without rewording”
150 watermarked pages (GPT-5.5 rewrite: 67), Gemma 4 26B, our key, 28 Sep 2026. Whiskers: 95% intervals. Quality: one AI judge (DeepSeek V4 Flash) scoring the changed text against the original.
02 / 05
Length
Length matters. Cut to the first 200 tokens, 67% are still detected; cut to 100 tokens, 29%.2
-
Share of watermarked pages still detected
Our study · 28 Sep 20260%50%100%AI judge, 0 to 10Control
Untouched 88%Copy and paste 88%Length
Cut to 200 tokens 67%Cut to 100 tokens 29%AI light edits
Asked 5%6% changed 65% 9.9Asked 10%10% changed 45% 9.9Asked 15%13% changed 36% 9.9Asked 20%17% changed 17% 9.8Asked 30%26% changed 3.3% 9.7Asked 40%31% changed 3.3% 9.5Mixing
Half mixed with other text 38%AI rewrite
AI rewrite, Mistral Nemo 5.3% 9.5AI rewrite, GPT-5.5 (67 pages) 0% 9.7Translation
Japanese and back 2.7% 9.6Formatting
5 kinds, from added spacing to writing in capitals: see “Removal without rewording”
150 watermarked pages (GPT-5.5 rewrite: 67), Gemma 4 26B, our key, 28 Sep 2026. Whiskers: 95% intervals. Quality: one AI judge (DeepSeek V4 Flash) scoring the changed text against the original.
03 / 05
AI light edits
AI light edits wear the mark down step by step. Edits that changed 6% of the words leave 65% detected, edits that changed 17% leave 17%, and edits that changed 26% leave 3%.2
-
Share of watermarked pages still detected
Our study · 28 Sep 20260%50%100%AI judge, 0 to 10Control
Untouched 88%Copy and paste 88%Length
Cut to 200 tokens 67%Cut to 100 tokens 29%AI light edits
Asked 5%6% changed 65% 9.9Asked 10%10% changed 45% 9.9Asked 15%13% changed 36% 9.9Asked 20%17% changed 17% 9.8Asked 30%26% changed 3.3% 9.7Asked 40%31% changed 3.3% 9.5Mixing
Half mixed with other text 38%AI rewrite
AI rewrite, Mistral Nemo 5.3% 9.5AI rewrite, GPT-5.5 (67 pages) 0% 9.7Translation
Japanese and back 2.7% 9.6Formatting
5 kinds, from added spacing to writing in capitals: see “Removal without rewording”
150 watermarked pages (GPT-5.5 rewrite: 67), Gemma 4 26B, our key, 28 Sep 2026. Whiskers: 95% intervals. Quality: one AI judge (DeepSeek V4 Flash) scoring the changed text against the original.
-
Share of watermarked pages still detected
Our study · 28 Sep 20260%50%100%AI judge, 0 to 10Control
Untouched 88%Copy and paste 88%Length
Cut to 200 tokens 67%Cut to 100 tokens 29%AI light edits
Asked 5%6% changed 65% 9.9Asked 10%10% changed 45% 9.9Asked 15%13% changed 36% 9.9Asked 20%17% changed 17% 9.8Asked 30%26% changed 3.3% 9.7Asked 40%31% changed 3.3% 9.5Mixing
Half mixed with other text 38%AI rewrite
AI rewrite, Mistral Nemo 5.3% 9.5AI rewrite, GPT-5.5 (67 pages) 0% 9.7Translation
Japanese and back 2.7% 9.6Formatting
5 kinds, from added spacing to writing in capitals: see “Removal without rewording”
150 watermarked pages (GPT-5.5 rewrite: 67), Gemma 4 26B, our key, 28 Sep 2026. Whiskers: 95% intervals. Quality: one AI judge (DeepSeek V4 Flash) scoring the changed text against the original.
05 / 05
Rewrites and translation
A full AI rewrite removes it: 5.3% detected after Mistral Nemo and 0 of 67 after GPT-5.5. A round trip through Japanese leaves 2.7%. An AI judge still scores those rewrites 9.5 and 9.7 out of 10.2
Words changed, not words asked for
How much rewording it takes
-
Detected, by share of words actually changed
Our study · 28 Sep 2026All 150 pages, every edit, rewrite and translation Most room (over 0.592 bits)Middle thirdLeast room (under 0.533 bits) Pilot, 12 articles, Gemma 4 E2B, 26 Sep 2026Main study, Gemma 4 26B, our key, 28 Sep 2026. Words changed is measured, not the share the editor was asked for. Room is the model's own entropy while writing, split into thirds.
01 / 05
Placed by words changed
Here every edit, rewrite and translation is placed by how many words actually changed, not by what the editor was asked to do. With under 5% of the words changed, 85% of pages are detected.2
-
Detected, by share of words actually changed
Our study · 28 Sep 2026All 150 pages, every edit, rewrite and translation Most room (over 0.592 bits)Middle thirdLeast room (under 0.533 bits) Pilot, 12 articles, Gemma 4 E2B, 26 Sep 2026Main study, Gemma 4 26B, our key, 28 Sep 2026. Words changed is measured, not the share the editor was asked for. Room is the model's own entropy while writing, split into thirds.
02 / 05
One word in ten
Detection falls fast. Between 5% and 10% changed, 54% are detected; between 10% and 15%, 32%. Half the pages lose the mark at about one word in ten.2
-
Detected, by share of words actually changed
Our study · 28 Sep 2026All 150 pages, every edit, rewrite and translation Most room (over 0.592 bits)Middle thirdLeast room (under 0.533 bits) Pilot, 12 articles, Gemma 4 E2B, 26 Sep 2026Main study, Gemma 4 26B, our key, 28 Sep 2026. Words changed is measured, not the share the editor was asked for. Room is the model's own entropy while writing, split into thirds.
-
Detected, by share of words actually changed
Our study · 28 Sep 2026All 150 pages, every edit, rewrite and translation Most room (over 0.592 bits)Middle thirdLeast room (under 0.533 bits) Pilot, 12 articles, Gemma 4 E2B, 26 Sep 2026Main study, Gemma 4 26B, our key, 28 Sep 2026. Words changed is measured, not the share the editor was asked for. Room is the model's own entropy while writing, split into thirds.
04 / 05
Room while writing
Pages written with the most room start stronger: 96% detected under 5% change, against 70% for the least room. By 15% to 20% changed, all three groups sit between 9% and 17%.2
-
Detected, by share of words actually changed
Our study · 28 Sep 2026All 150 pages, every edit, rewrite and translation Most room (over 0.592 bits)Middle thirdLeast room (under 0.533 bits) Pilot, 12 articles, Gemma 4 E2B, 26 Sep 2026Main study, Gemma 4 26B, our key, 28 Sep 2026. Words changed is measured, not the share the editor was asked for. Room is the model's own entropy while writing, split into thirds.
Tap a version to compare
One article, nine versions
-
One pilot article, token by token
Our pilot · 26 Sep 2026Original z 6.8 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 11% changed z 4.9 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the correct choice is vitally critical for preventing injury and guaranteeing an enjoyable occasion. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 19% changed z 2.7 · detected
Setting out on the journey to your first marathon is a thrilling goal, but before you lace up those first race shoes, making the right choice is absolutely vital for avoiding injury and ensuring an enjoyable experience. Choosing the perfect running shoe is less about finding the trendiest new model and more about finding a dependable companion that aligns with your foot mechanics and the demands of long-distance running. Begin by grasping your stride; are you a heel striker, a midfoot runner, or a forefoot strider? Although not a formal analysis, paying attention to howCut to 100 tokens z 1.9 · not detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider?Half mixed z 2.8 · detected
Tired of shouting into the void when trying to connect with your customers? If you run a small business, building a direct line of communication is one of the most powerful, yet often underutilized, tools at your disposal, and that’s where email newsletters come in. Think of an email newsletter not just as a promotional blast, but as a personal touchpoint, a way to nurture the relationships you’ve built with your audience and keep your brand top-of-mind long after they’ve left your physical store or website. StartingSpanish and back z 2.8 · detected
Starting the journey toward your first marathon is an exciting goal, but before lacing up for the race, choosing the right shoes is absolutely crucial to prevent injuries and ensure a pleasant experience. Selecting the perfect running shoes isn't about finding the flashiest model, but rather about finding a reliable companion that matches your foot's biomechanics and the demands of long-distance running. Begin by understanding your stride: do you land on your heel, midfoot, or forefoot? While not a formal analysis, paying attention to how your foot lands andJapanese and back z 0.50 · not detected
The goal of tackling your first marathon is exciting, but choosing the right running shoes before lacing up is crucial to prevent injury and ensure an enjoyable experience. Selecting the perfect running shoe isn't about finding the latest flashy model—it's about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait pattern. Do you land on your heel, midfoot, or forefoot? Even without a professional analysis, observing how your foot strikes the ground and how your weight shifts can provide valuableRewrite, Mistral Nemo z 0.78 · not detected
Preparing for Your Maiden Marathon: The Perfect Running Shoe Selection
Embarking on your first marathon is an exciting challenge, but choosing the right running shoe is vital for injury prevention and a pleasant experience. The ideal shoe isn't about aesthetics; it's about finding a trustworthy companion that aligns with your foot mechanics and marathon demands.
Begin by understanding your running style. Observe how your foot lands and rolls - are you a heel striker, midfoot runner, or forefoot strider? This informal analysis will hint at the support youRewrite, GPT-5.5 z 1.5 · not detected
Training for your first marathon is an exciting challenge, but before you put in serious miles, choosing the right running shoes is essential. The right pair can help reduce your risk of injury and make the entire experience far more comfortable. Picking a marathon shoe is not about grabbing the trendiest or most eye-catching model; it is about finding a dependable option that suits the way your feet move and can handle the demands of long-distance running. A good place to begin is by noticing your running gait. Do you usually land on your heel, midfootWritten without the key z 0.55 · not detected
Embarking on your first marathon is an exhilarating achievement, but the journey doesn't end at the starting line; the right running shoes are absolutely crucial for comfort, injury prevention, and overall performance. Choosing the perfect pair is a deeply personal process that requires paying close attention to your running style, foot mechanics, and training progression. Before you dive into the endless catalog of options, the most important first step is visiting a specialized running store where experts can analyze your gait. They will observe how you run, whether you overpronate, supinatescore above 0.5: evidence for our mark faded: below 0.5Pilot article blog-02, Gemma 4 E2B, 26 Sep 2026. One article; the averages are in the two scenes before this one. The z and the verdict are for the whole text; the first 110 tokens are shown.
01 / 04
Untouched
This is one pilot article, each word tinted by how much it supports our mark. Untouched, it scores z 6.8.1
-
One pilot article, token by token
Our pilot · 26 Sep 2026Original z 6.8 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 11% changed z 4.9 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the correct choice is vitally critical for preventing injury and guaranteeing an enjoyable occasion. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 19% changed z 2.7 · detected
Setting out on the journey to your first marathon is a thrilling goal, but before you lace up those first race shoes, making the right choice is absolutely vital for avoiding injury and ensuring an enjoyable experience. Choosing the perfect running shoe is less about finding the trendiest new model and more about finding a dependable companion that aligns with your foot mechanics and the demands of long-distance running. Begin by grasping your stride; are you a heel striker, a midfoot runner, or a forefoot strider? Although not a formal analysis, paying attention to howCut to 100 tokens z 1.9 · not detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider?Half mixed z 2.8 · detected
Tired of shouting into the void when trying to connect with your customers? If you run a small business, building a direct line of communication is one of the most powerful, yet often underutilized, tools at your disposal, and that’s where email newsletters come in. Think of an email newsletter not just as a promotional blast, but as a personal touchpoint, a way to nurture the relationships you’ve built with your audience and keep your brand top-of-mind long after they’ve left your physical store or website. StartingSpanish and back z 2.8 · detected
Starting the journey toward your first marathon is an exciting goal, but before lacing up for the race, choosing the right shoes is absolutely crucial to prevent injuries and ensure a pleasant experience. Selecting the perfect running shoes isn't about finding the flashiest model, but rather about finding a reliable companion that matches your foot's biomechanics and the demands of long-distance running. Begin by understanding your stride: do you land on your heel, midfoot, or forefoot? While not a formal analysis, paying attention to how your foot lands andJapanese and back z 0.50 · not detected
The goal of tackling your first marathon is exciting, but choosing the right running shoes before lacing up is crucial to prevent injury and ensure an enjoyable experience. Selecting the perfect running shoe isn't about finding the latest flashy model—it's about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait pattern. Do you land on your heel, midfoot, or forefoot? Even without a professional analysis, observing how your foot strikes the ground and how your weight shifts can provide valuableRewrite, Mistral Nemo z 0.78 · not detected
Preparing for Your Maiden Marathon: The Perfect Running Shoe Selection
Embarking on your first marathon is an exciting challenge, but choosing the right running shoe is vital for injury prevention and a pleasant experience. The ideal shoe isn't about aesthetics; it's about finding a trustworthy companion that aligns with your foot mechanics and marathon demands.
Begin by understanding your running style. Observe how your foot lands and rolls - are you a heel striker, midfoot runner, or forefoot strider? This informal analysis will hint at the support youRewrite, GPT-5.5 z 1.5 · not detected
Training for your first marathon is an exciting challenge, but before you put in serious miles, choosing the right running shoes is essential. The right pair can help reduce your risk of injury and make the entire experience far more comfortable. Picking a marathon shoe is not about grabbing the trendiest or most eye-catching model; it is about finding a dependable option that suits the way your feet move and can handle the demands of long-distance running. A good place to begin is by noticing your running gait. Do you usually land on your heel, midfootWritten without the key z 0.55 · not detected
Embarking on your first marathon is an exhilarating achievement, but the journey doesn't end at the starting line; the right running shoes are absolutely crucial for comfort, injury prevention, and overall performance. Choosing the perfect pair is a deeply personal process that requires paying close attention to your running style, foot mechanics, and training progression. Before you dive into the endless catalog of options, the most important first step is visiting a specialized running store where experts can analyze your gait. They will observe how you run, whether you overpronate, supinatescore above 0.5: evidence for our mark faded: below 0.5Pilot article blog-02, Gemma 4 E2B, 26 Sep 2026. One article; the averages are in the two scenes before this one. The z and the verdict are for the whole text; the first 110 tokens are shown.
02 / 04
AI edits
An AI edit that changes 11% of the words leaves z 4.9. One that changes 19% drops it to 2.7, just above the line.1
-
One pilot article, token by token
Our pilot · 26 Sep 2026Original z 6.8 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 11% changed z 4.9 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the correct choice is vitally critical for preventing injury and guaranteeing an enjoyable occasion. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 19% changed z 2.7 · detected
Setting out on the journey to your first marathon is a thrilling goal, but before you lace up those first race shoes, making the right choice is absolutely vital for avoiding injury and ensuring an enjoyable experience. Choosing the perfect running shoe is less about finding the trendiest new model and more about finding a dependable companion that aligns with your foot mechanics and the demands of long-distance running. Begin by grasping your stride; are you a heel striker, a midfoot runner, or a forefoot strider? Although not a formal analysis, paying attention to howCut to 100 tokens z 1.9 · not detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider?Half mixed z 2.8 · detected
Tired of shouting into the void when trying to connect with your customers? If you run a small business, building a direct line of communication is one of the most powerful, yet often underutilized, tools at your disposal, and that’s where email newsletters come in. Think of an email newsletter not just as a promotional blast, but as a personal touchpoint, a way to nurture the relationships you’ve built with your audience and keep your brand top-of-mind long after they’ve left your physical store or website. StartingSpanish and back z 2.8 · detected
Starting the journey toward your first marathon is an exciting goal, but before lacing up for the race, choosing the right shoes is absolutely crucial to prevent injuries and ensure a pleasant experience. Selecting the perfect running shoes isn't about finding the flashiest model, but rather about finding a reliable companion that matches your foot's biomechanics and the demands of long-distance running. Begin by understanding your stride: do you land on your heel, midfoot, or forefoot? While not a formal analysis, paying attention to how your foot lands andJapanese and back z 0.50 · not detected
The goal of tackling your first marathon is exciting, but choosing the right running shoes before lacing up is crucial to prevent injury and ensure an enjoyable experience. Selecting the perfect running shoe isn't about finding the latest flashy model—it's about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait pattern. Do you land on your heel, midfoot, or forefoot? Even without a professional analysis, observing how your foot strikes the ground and how your weight shifts can provide valuableRewrite, Mistral Nemo z 0.78 · not detected
Preparing for Your Maiden Marathon: The Perfect Running Shoe Selection
Embarking on your first marathon is an exciting challenge, but choosing the right running shoe is vital for injury prevention and a pleasant experience. The ideal shoe isn't about aesthetics; it's about finding a trustworthy companion that aligns with your foot mechanics and marathon demands.
Begin by understanding your running style. Observe how your foot lands and rolls - are you a heel striker, midfoot runner, or forefoot strider? This informal analysis will hint at the support youRewrite, GPT-5.5 z 1.5 · not detected
Training for your first marathon is an exciting challenge, but before you put in serious miles, choosing the right running shoes is essential. The right pair can help reduce your risk of injury and make the entire experience far more comfortable. Picking a marathon shoe is not about grabbing the trendiest or most eye-catching model; it is about finding a dependable option that suits the way your feet move and can handle the demands of long-distance running. A good place to begin is by noticing your running gait. Do you usually land on your heel, midfootWritten without the key z 0.55 · not detected
Embarking on your first marathon is an exhilarating achievement, but the journey doesn't end at the starting line; the right running shoes are absolutely crucial for comfort, injury prevention, and overall performance. Choosing the perfect pair is a deeply personal process that requires paying close attention to your running style, foot mechanics, and training progression. Before you dive into the endless catalog of options, the most important first step is visiting a specialized running store where experts can analyze your gait. They will observe how you run, whether you overpronate, supinatescore above 0.5: evidence for our mark faded: below 0.5Pilot article blog-02, Gemma 4 E2B, 26 Sep 2026. One article; the averages are in the two scenes before this one. The z and the verdict are for the whole text; the first 110 tokens are shown.
-
One pilot article, token by token
Our pilot · 26 Sep 2026Original z 6.8 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 11% changed z 4.9 · detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the correct choice is vitally critical for preventing injury and guaranteeing an enjoyable occasion. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider? While not a formal analysis, paying attention to howAI edit · 19% changed z 2.7 · detected
Setting out on the journey to your first marathon is a thrilling goal, but before you lace up those first race shoes, making the right choice is absolutely vital for avoiding injury and ensuring an enjoyable experience. Choosing the perfect running shoe is less about finding the trendiest new model and more about finding a dependable companion that aligns with your foot mechanics and the demands of long-distance running. Begin by grasping your stride; are you a heel striker, a midfoot runner, or a forefoot strider? Although not a formal analysis, paying attention to howCut to 100 tokens z 1.9 · not detected
Embarking on the journey to your first marathon is an exhilarating goal, but before you lace up those first race shoes, making the right choice is absolutely crucial for preventing injury and ensuring an enjoyable experience. Selecting the perfect running shoe is less about finding the flashiest new model and more about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait; are you a heel striker, a midfoot runner, or a forefoot strider?Half mixed z 2.8 · detected
Tired of shouting into the void when trying to connect with your customers? If you run a small business, building a direct line of communication is one of the most powerful, yet often underutilized, tools at your disposal, and that’s where email newsletters come in. Think of an email newsletter not just as a promotional blast, but as a personal touchpoint, a way to nurture the relationships you’ve built with your audience and keep your brand top-of-mind long after they’ve left your physical store or website. StartingSpanish and back z 2.8 · detected
Starting the journey toward your first marathon is an exciting goal, but before lacing up for the race, choosing the right shoes is absolutely crucial to prevent injuries and ensure a pleasant experience. Selecting the perfect running shoes isn't about finding the flashiest model, but rather about finding a reliable companion that matches your foot's biomechanics and the demands of long-distance running. Begin by understanding your stride: do you land on your heel, midfoot, or forefoot? While not a formal analysis, paying attention to how your foot lands andJapanese and back z 0.50 · not detected
The goal of tackling your first marathon is exciting, but choosing the right running shoes before lacing up is crucial to prevent injury and ensure an enjoyable experience. Selecting the perfect running shoe isn't about finding the latest flashy model—it's about finding a reliable partner that matches your foot mechanics and the demands of long-distance running. Start by understanding your gait pattern. Do you land on your heel, midfoot, or forefoot? Even without a professional analysis, observing how your foot strikes the ground and how your weight shifts can provide valuableRewrite, Mistral Nemo z 0.78 · not detected
Preparing for Your Maiden Marathon: The Perfect Running Shoe Selection
Embarking on your first marathon is an exciting challenge, but choosing the right running shoe is vital for injury prevention and a pleasant experience. The ideal shoe isn't about aesthetics; it's about finding a trustworthy companion that aligns with your foot mechanics and marathon demands.
Begin by understanding your running style. Observe how your foot lands and rolls - are you a heel striker, midfoot runner, or forefoot strider? This informal analysis will hint at the support youRewrite, GPT-5.5 z 1.5 · not detected
Training for your first marathon is an exciting challenge, but before you put in serious miles, choosing the right running shoes is essential. The right pair can help reduce your risk of injury and make the entire experience far more comfortable. Picking a marathon shoe is not about grabbing the trendiest or most eye-catching model; it is about finding a dependable option that suits the way your feet move and can handle the demands of long-distance running. A good place to begin is by noticing your running gait. Do you usually land on your heel, midfootWritten without the key z 0.55 · not detected
Embarking on your first marathon is an exhilarating achievement, but the journey doesn't end at the starting line; the right running shoes are absolutely crucial for comfort, injury prevention, and overall performance. Choosing the perfect pair is a deeply personal process that requires paying close attention to your running style, foot mechanics, and training progression. Before you dive into the endless catalog of options, the most important first step is visiting a specialized running store where experts can analyze your gait. They will observe how you run, whether you overpronate, supinatescore above 0.5: evidence for our mark faded: below 0.5Pilot article blog-02, Gemma 4 E2B, 26 Sep 2026. One article; the averages are in the two scenes before this one. The z and the verdict are for the whole text; the first 110 tokens are shown.
04 / 04
Rewrites
AI rewrites leave nothing readable: z 0.8 after Mistral Nemo, 1.5 after GPT-5.5. The article written without the key scores 0.55.1
Formatting, not words
Removal without rewording
-
Detected after formatting changes
Our study · 28 Sep 2026Changed after writing
Formatted while writing, converted back
Why: each word is scored with the tokens just before it Illustration
Writtenthe◆right◆shoe“shoe” scored next to ◆
Symbols removedfortherightshoesame word, new context: a fresh, unrelated score
150 pages per row, Gemma 4 26B, our key, 28 Sep 2026. The formatted-while-writing rows are fresh watermarked pages written under a formatting instruction, then converted back. Cleaning up whitespace before scoring does not raise their rates.
01 / 04
Spacing added after writing
Some changes alter no words at all. Scattering tabs and line breaks through a finished page drops detection from 88% to 2% for a detector that reads the text as is.2
-
Detected after formatting changes
Our study · 28 Sep 2026Changed after writing
Formatted while writing, converted back
Why: each word is scored with the tokens just before it Illustration
Writtenthe◆right◆shoe“shoe” scored next to ◆
Symbols removedfortherightshoesame word, new context: a fresh, unrelated score
150 pages per row, Gemma 4 26B, our key, 28 Sep 2026. The formatted-while-writing rows are fresh watermarked pages written under a formatting instruction, then converted back. Cleaning up whitespace before scoring does not raise their rates.
02 / 04
Reversible
That change is reversible, and so is the damage. A detector that collapses whitespace before scoring recovers 84%.2
-
Detected after formatting changes
Our study · 28 Sep 2026Changed after writing
Formatted while writing, converted back
Why: each word is scored with the tokens just before it Illustration
Writtenthe◆right◆shoe“shoe” scored next to ◆
Symbols removedfortherightshoesame word, new context: a fresh, unrelated score
150 pages per row, Gemma 4 26B, our key, 28 Sep 2026. The formatted-while-writing rows are fresh watermarked pages written under a formatting instruction, then converted back. Cleaning up whitespace before scoring does not raise their rates.
03 / 04
Formatted while writing
Formatting while writing is different. When the model writes with a symbol after every word, one word per line or in capitals, and the text is converted back afterwards, 0.7% to 2.7% of pages are detected, even after cleanup.2
-
Detected after formatting changes
Our study · 28 Sep 2026Changed after writing
Formatted while writing, converted back
Why: each word is scored with the tokens just before it Illustration
Writtenthe◆right◆shoe“shoe” scored next to ◆
Symbols removedfortherightshoesame word, new context: a fresh, unrelated score
150 pages per row, Gemma 4 26B, our key, 28 Sep 2026. The formatted-while-writing rows are fresh watermarked pages written under a formatting instruction, then converted back. Cleaning up whitespace before scoring does not raise their rates.
04 / 04
Why: context
The reason is context. Each word's score depends on the tokens just before it, so words written next to a symbol sit in a different context once it is gone. Kirchenbauer et al. described this “emoji attack” in 2023; a provider can defend by ignoring such symbols when it applies the mark, and whether Claude or Gemini does is not public.2,9
Why some pages start weaker
AI pages leave less room
-
Room per token: the real pages and oursRoom while writing and the strength of the mark
Our study · 28 Sep 2026Blog post 0.58 / 1.20 97% detectedCase study 0.55 / 1.24 93% detectedGuide 0.61 / 1.35 93% detectedPromo page 0.55 / 1.21 83% detectedService page 0.51 / 1.20 73% detected0.00.51.01.5human-written page on the real site our model's page, same brief our model's page, detected hollow: missed blog post service pageHuman room: Gemma 4 E2B reading the real page. AI room: Gemma 4 26B measuring its own odds while writing. Two models, so read the gap as indicative, not exact. 30 pages per type, 28 Sep 2026.
01 / 03
About half the room
The real site's human-written pages carry 1.20 to 1.35 bits per token by page type. The pages our model wrote from the same briefs carry 0.51 to 0.61, about half. A model writes what it finds most predictable.2,3
Human room was scored by Gemma 4 E2B reading the page; AI room was measured by Gemma 4 26B while writing. Read the gap as indicative, not exact.
-
Room per token: the real pages and oursRoom while writing and the strength of the mark
Our study · 28 Sep 2026Blog post 0.58 / 1.20 97% detectedCase study 0.55 / 1.24 93% detectedGuide 0.61 / 1.35 93% detectedPromo page 0.55 / 1.21 83% detectedService page 0.51 / 1.20 73% detected0.00.51.01.5human-written page on the real site our model's page, same brief our model's page, detected hollow: missed blog post service pageHuman room: Gemma 4 E2B reading the real page. AI room: Gemma 4 26B measuring its own odds while writing. Two models, so read the gap as indicative, not exact. 30 pages per type, 28 Sep 2026.
02 / 03
Less room, weaker mark
Less room means a weaker mark. Across the 150 untouched pages, more room goes with a higher score, a rank correlation of 0.40.2
-
Room per token: the real pages and oursRoom while writing and the strength of the mark
Our study · 28 Sep 2026Blog post 0.58 / 1.20 97% detectedCase study 0.55 / 1.24 93% detectedGuide 0.61 / 1.35 93% detectedPromo page 0.55 / 1.21 83% detectedService page 0.51 / 1.20 73% detected0.00.51.01.5human-written page on the real site our model's page, same brief our model's page, detected hollow: missed blog post service pageHuman room: Gemma 4 E2B reading the real page. AI room: Gemma 4 26B measuring its own odds while writing. Two models, so read the gap as indicative, not exact. 30 pages per type, 28 Sep 2026.
03 / 03
Service pages and blog posts
Service pages, the tightest, are detected 73% of the time untouched; blog posts, 97%. With 30 pages per type, those two figures carry wide margins.2
In plain language
What the experiments show, and what they do not
Method in one paragraph
We used Google's open SynthID-Text code in Hugging Face Transformers with a 30-round key we made up and a 5-token context, sampling at temperature 0.8 with top-k 40.1 A text counts as detected when its weighted score is 2.33 standard deviations above chance, a 1% false-alarm rate by design. Light edits came from an AI editor (Llama 3.3 70B) told to change a set share of words; rewrites came from Mistral Nemo and GPT-5.5; translation from Qwen 3 235B; quality scores from one AI judge (DeepSeek V4 Flash), out of 10, against the original. We measured how many words each change actually altered.
What it shows
- Copy and paste, reformatting that can be undone, and very light edits leave the mark readable.
- Any AI rewrite or translation removes it almost completely, and the rewritten text still scores well on quality.
- Editing removes it gradually: half the pages lose it at about 10% of words changed, and almost none keep it past 20%. Two models, two days apart, trace the same curve.
- False alarms are rare and match the design: 0.4% of 2,250 unwatermarked texts.
- Formatting instructions given at writing time can defeat a mark that is keyed on raw tokens.
What it does not show
- Anything direct about Claude's or Gemini's deployed marks. Their keys, models, settings, detectors and defences differ from ours and are not public.
- How human editors affect the mark. Our edits were made by an AI editor; the 30 human edits in the original plan were not run.
- Any effect on search rankings. The planned ranking field test has not run.
- Quality as readers judge it. One AI judge scored each change; no human panel.
- Precise per-type rates. Each page type has 30 pages, and the GPT-5.5 rewrite ran on 67 of 150 pages because of rate limits.
How it fits published work Reported
Published tests of open SynthID-Text code report the same pattern. One 2026 study finds paraphrase removes 98.3% of marks;13 a 2024 study finds translation drops detection to near chance;14 and a 2023 study finds a watermark can survive human paraphrase given about 800 tokens.15 None of them tested the deployed Claude or Gemini keys.
Chapter 04 · Access and deadlines
Who can check it
Set against the measurements: who can actually read a real mark, what a hit means, and the dates that change the picture.
Reported by the providers and regulators
Keys, access and deadlines
-
Who marks text, who can check it, and when
Reported · as of 26 Sep 2026- 10 Jun 2026EU final Code of Practice on marking AI-generated content
- 2 Aug 2026EU AI Act Article 50 applies
- 14 Aug 2026Anthropic explains Claude's text watermark, a version of SynthID-Text
- ~1 Sep 2026Claude text detection opens as a private preview
- 2 Dec 2026Claude models released before Aug 2 carry text marks; Mistral's stated date for marking and detection
- 2 Feb 2027EU Code: providers' detection must interoperate
A hit means the text
may have been generated or processed by Claude
.Claude Help Center, updated 22 Sep 2026. A miss proves nothing.
ProviderMarks textMethod and detectionAnthropic Yes A version of SynthID-Text. Detection: private preview for named groupsGoogle Yes, in the Gemini app and web SynthID-Text. Detector portal for early testers; Gemini API text marking not confirmedOpenAI No Built one in 2024 and held it back; marks images with C2PA and SynthIDMeta No deployment found TextSeal research, open-sourced May 2026Mistral Planned Says marking and detection by 2 Dec 2026Cohere Not stated Signed the EU Code; method unstatedMicrosoft, xAI, Amazon None found Image, audio or video marks only in reviewed materialAs of 26 Sep 2026, from the providers' own pages where possible; future dots (dashed) are commitments or legal deadlines, not facts yet. The private-preview date is approximate (a press report).
01 / 05
The EU rule
EU AI Act Article 50 has applied since 2 Aug 2026. It requires AI-generated text to be marked in a machine-readable form that can be detected.18
-
Who marks text, who can check it, and when
Reported · as of 26 Sep 2026- 10 Jun 2026EU final Code of Practice on marking AI-generated content
- 2 Aug 2026EU AI Act Article 50 applies
- 14 Aug 2026Anthropic explains Claude's text watermark, a version of SynthID-Text
- ~1 Sep 2026Claude text detection opens as a private preview
- 2 Dec 2026Claude models released before Aug 2 carry text marks; Mistral's stated date for marking and detection
- 2 Feb 2027EU Code: providers' detection must interoperate
A hit means the text
may have been generated or processed by Claude
.Claude Help Center, updated 22 Sep 2026. A miss proves nothing.
ProviderMarks textMethod and detectionAnthropic Yes A version of SynthID-Text. Detection: private preview for named groupsGoogle Yes, in the Gemini app and web SynthID-Text. Detector portal for early testers; Gemini API text marking not confirmedOpenAI No Built one in 2024 and held it back; marks images with C2PA and SynthIDMeta No deployment found TextSeal research, open-sourced May 2026Mistral Planned Says marking and detection by 2 Dec 2026Cohere Not stated Signed the EU Code; method unstatedMicrosoft, xAI, Amazon None found Image, audio or video marks only in reviewed materialAs of 26 Sep 2026, from the providers' own pages where possible; future dots (dashed) are commitments or legal deadlines, not facts yet. The private-preview date is approximate (a press report).
-
Who marks text, who can check it, and when
Reported · as of 26 Sep 2026- 10 Jun 2026EU final Code of Practice on marking AI-generated content
- 2 Aug 2026EU AI Act Article 50 applies
- 14 Aug 2026Anthropic explains Claude's text watermark, a version of SynthID-Text
- ~1 Sep 2026Claude text detection opens as a private preview
- 2 Dec 2026Claude models released before Aug 2 carry text marks; Mistral's stated date for marking and detection
- 2 Feb 2027EU Code: providers' detection must interoperate
A hit means the text
may have been generated or processed by Claude
.Claude Help Center, updated 22 Sep 2026. A miss proves nothing.
ProviderMarks textMethod and detectionAnthropic Yes A version of SynthID-Text. Detection: private preview for named groupsGoogle Yes, in the Gemini app and web SynthID-Text. Detector portal for early testers; Gemini API text marking not confirmedOpenAI No Built one in 2024 and held it back; marks images with C2PA and SynthIDMeta No deployment found TextSeal research, open-sourced May 2026Mistral Planned Says marking and detection by 2 Dec 2026Cohere Not stated Signed the EU Code; method unstatedMicrosoft, xAI, Amazon None found Image, audio or video marks only in reviewed materialAs of 26 Sep 2026, from the providers' own pages where possible; future dots (dashed) are commitments or legal deadlines, not facts yet. The private-preview date is approximate (a press report).
-
Who marks text, who can check it, and when
Reported · as of 26 Sep 2026- 10 Jun 2026EU final Code of Practice on marking AI-generated content
- 2 Aug 2026EU AI Act Article 50 applies
- 14 Aug 2026Anthropic explains Claude's text watermark, a version of SynthID-Text
- ~1 Sep 2026Claude text detection opens as a private preview
- 2 Dec 2026Claude models released before Aug 2 carry text marks; Mistral's stated date for marking and detection
- 2 Feb 2027EU Code: providers' detection must interoperate
A hit means the text
may have been generated or processed by Claude
.Claude Help Center, updated 22 Sep 2026. A miss proves nothing.
ProviderMarks textMethod and detectionAnthropic Yes A version of SynthID-Text. Detection: private preview for named groupsGoogle Yes, in the Gemini app and web SynthID-Text. Detector portal for early testers; Gemini API text marking not confirmedOpenAI No Built one in 2024 and held it back; marks images with C2PA and SynthIDMeta No deployment found TextSeal research, open-sourced May 2026Mistral Planned Says marking and detection by 2 Dec 2026Cohere Not stated Signed the EU Code; method unstatedMicrosoft, xAI, Amazon None found Image, audio or video marks only in reviewed materialAs of 26 Sep 2026, from the providers' own pages where possible; future dots (dashed) are commitments or legal deadlines, not facts yet. The private-preview date is approximate (a press report).
04 / 05
What a hit means
A hit means the text “may have been generated or processed by Claude”, which includes proofreading and translation. A miss proves nothing.6
-
Who marks text, who can check it, and when
Reported · as of 26 Sep 2026- 10 Jun 2026EU final Code of Practice on marking AI-generated content
- 2 Aug 2026EU AI Act Article 50 applies
- 14 Aug 2026Anthropic explains Claude's text watermark, a version of SynthID-Text
- ~1 Sep 2026Claude text detection opens as a private preview
- 2 Dec 2026Claude models released before Aug 2 carry text marks; Mistral's stated date for marking and detection
- 2 Feb 2027EU Code: providers' detection must interoperate
A hit means the text
may have been generated or processed by Claude
.Claude Help Center, updated 22 Sep 2026. A miss proves nothing.
ProviderMarks textMethod and detectionAnthropic Yes A version of SynthID-Text. Detection: private preview for named groupsGoogle Yes, in the Gemini app and web SynthID-Text. Detector portal for early testers; Gemini API text marking not confirmedOpenAI No Built one in 2024 and held it back; marks images with C2PA and SynthIDMeta No deployment found TextSeal research, open-sourced May 2026Mistral Planned Says marking and detection by 2 Dec 2026Cohere Not stated Signed the EU Code; method unstatedMicrosoft, xAI, Amazon None found Image, audio or video marks only in reviewed materialAs of 26 Sep 2026, from the providers' own pages where possible; future dots (dashed) are commitments or legal deadlines, not facts yet. The private-preview date is approximate (a press report).
Chapter 05 · The marketer's question
What it means for search
The question most marketers ask, with what Google has said kept apart from what people infer.
On record, and our read
What Google has and has not said
-
Watermarks and Google Search: on record, and our read
Reported · as of 26 Sep 2026- 21 Aug 2026 Google called the August 2026 spam update (Aug 18 to 21) a normal spam update and published no new policy
- 24 Sep 2026 The September 2026 spam update began Sep 24; Google said it may take up to two weeksCompletion not confirmed in a search on 9 Oct 2026.
- accessed 26 Sep 2026 Google's policy targets content made at scale to manipulate rankings, however it is produced
- 27 Aug 2026 SE Ranking: 16.71% of prior top-10 URLs fell out of the top 100 in the August window, against 9.2% in July
- 26 Sep 2026 No primary Google or Anthropic statement links text watermarks to Search ranking
Google demotes watermarked text
One correlational study (First Page Sage) that does not say how it detected watermarks or control for quality. Google would need Anthropic's key.
Unsupported
Google runs its own AI-text detector on each page
AI-heavy pages do worse in several studies (Ahrefs, Lily Ray), but AI text and low-quality text overlap.
Possible
Google looks for patterns at scale: templates, networks, coordination
Fits the August case studies and Google's published S-CTS and SAFE research, which is about video channel clusters.
Most plausible
The long September rollout means a new system
Only the unusual length. The SAFE link is a journalist's guess.
Speculation
Google statements from its Search Status Dashboard and spam policies; findings from the studies named in the sources. The meters are our judgement of the evidence, not a measurement.
-
Watermarks and Google Search: on record, and our read
Reported · as of 26 Sep 2026- 21 Aug 2026 Google called the August 2026 spam update (Aug 18 to 21) a normal spam update and published no new policy
- 24 Sep 2026 The September 2026 spam update began Sep 24; Google said it may take up to two weeksCompletion not confirmed in a search on 9 Oct 2026.
- accessed 26 Sep 2026 Google's policy targets content made at scale to manipulate rankings, however it is produced
- 27 Aug 2026 SE Ranking: 16.71% of prior top-10 URLs fell out of the top 100 in the August window, against 9.2% in July
- 26 Sep 2026 No primary Google or Anthropic statement links text watermarks to Search ranking
Google demotes watermarked text
One correlational study (First Page Sage) that does not say how it detected watermarks or control for quality. Google would need Anthropic's key.
Unsupported
Google runs its own AI-text detector on each page
AI-heavy pages do worse in several studies (Ahrefs, Lily Ray), but AI text and low-quality text overlap.
Possible
Google looks for patterns at scale: templates, networks, coordination
Fits the August case studies and Google's published S-CTS and SAFE research, which is about video channel clusters.
Most plausible
The long September rollout means a new system
Only the unusual length. The SAFE link is a journalist's guess.
Speculation
Google statements from its Search Status Dashboard and spam policies; findings from the studies named in the sources. The meters are our judgement of the evidence, not a measurement.
-
Watermarks and Google Search: on record, and our read
Reported · as of 26 Sep 2026- 21 Aug 2026 Google called the August 2026 spam update (Aug 18 to 21) a normal spam update and published no new policy
- 24 Sep 2026 The September 2026 spam update began Sep 24; Google said it may take up to two weeksCompletion not confirmed in a search on 9 Oct 2026.
- accessed 26 Sep 2026 Google's policy targets content made at scale to manipulate rankings, however it is produced
- 27 Aug 2026 SE Ranking: 16.71% of prior top-10 URLs fell out of the top 100 in the August window, against 9.2% in July
- 26 Sep 2026 No primary Google or Anthropic statement links text watermarks to Search ranking
Google demotes watermarked text
One correlational study (First Page Sage) that does not say how it detected watermarks or control for quality. Google would need Anthropic's key.
Unsupported
Google runs its own AI-text detector on each page
AI-heavy pages do worse in several studies (Ahrefs, Lily Ray), but AI text and low-quality text overlap.
Possible
Google looks for patterns at scale: templates, networks, coordination
Fits the August case studies and Google's published S-CTS and SAFE research, which is about video channel clusters.
Most plausible
The long September rollout means a new system
Only the unusual length. The SAFE link is a journalist's guess.
Speculation
Google statements from its Search Status Dashboard and spam policies; findings from the studies named in the sources. The meters are our judgement of the evidence, not a measurement.
-
Watermarks and Google Search: on record, and our read
Reported · as of 26 Sep 2026- 21 Aug 2026 Google called the August 2026 spam update (Aug 18 to 21) a normal spam update and published no new policy
- 24 Sep 2026 The September 2026 spam update began Sep 24; Google said it may take up to two weeksCompletion not confirmed in a search on 9 Oct 2026.
- accessed 26 Sep 2026 Google's policy targets content made at scale to manipulate rankings, however it is produced
- 27 Aug 2026 SE Ranking: 16.71% of prior top-10 URLs fell out of the top 100 in the August window, against 9.2% in July
- 26 Sep 2026 No primary Google or Anthropic statement links text watermarks to Search ranking
Google demotes watermarked text
One correlational study (First Page Sage) that does not say how it detected watermarks or control for quality. Google would need Anthropic's key.
Unsupported
Google runs its own AI-text detector on each page
AI-heavy pages do worse in several studies (Ahrefs, Lily Ray), but AI text and low-quality text overlap.
Possible
Google looks for patterns at scale: templates, networks, coordination
Fits the August case studies and Google's published S-CTS and SAFE research, which is about video channel clusters.
Most plausible
The long September rollout means a new system
Only the unusual length. The SAFE link is a journalist's guess.
Speculation
Google statements from its Search Status Dashboard and spam policies; findings from the studies named in the sources. The meters are our judgement of the evidence, not a measurement.
Chapter 06 · For executives
Where we land
- A watermark shows a model was involved. It says nothing about quality or authorship, and a missing mark proves nothing.
- It is a compliance tool, not a policing tool. Removing it is cheap, as our study shows, and checking for it is restricted to a few groups.
- For search, nothing on record links it to rankings as of 26 Sep 2026. What gets hit is content published at scale without adding anything.
- Stripping marks for SEO has no evidence behind it. Editing AI drafts is normal, and human editorial review is also what exempts AI-assisted public-interest text from the EU's disclosure rule.18
- Keep drafts, sources and editor sign-offs. They settle an authorship question better than a watermark can.
Text watermark
Lives in the word choice
- Survives
- Copy and paste, light edits
- Removed by
- AI rewrites, translation, heavy edits
- Who can check
- Only the key-holder
- False alarms
- Set by design; we measured 0.4%2
File credentials (C2PA)
Signed metadata on images and files
- Removed by
- A screenshot or a re-save
- Who can check
- Anyone
- Claude
- Uses them for PNG, JPG and SVG6
“AI detectors”
A guess from style
- Key
- None
- Accuracy
- One study found seven detectors flagged 61% of non-native English essays as AI17
Sources and method
As of 26 Sep 2026. Outside facts on this page come from our verified research brief of 26 Sep 2026 and the sources below, and were not re-checked for this page unless dated otherwise. The September spam update's completion was searched again on 9 Oct 2026; no notice was found.
Our runs used Google's open SynthID-Text code in Hugging Face Transformers with a key we made up. The pilot ran on 26 Sep 2026 (Gemma 4 E2B, a laptop), the site map on 27 Sep 2026, the main study on 28 Sep 2026 (Gemma 4 26B, a rented A100) and the next-word probabilities on 29 Sep 2026. The real website behind the page types is not named and appears only as aggregates. Every chart's data, the code and the raw scores are kept with the project. “Accessed” is the date our research opened the page.
An earlier interactive version appeared on 29 Sep 2026; every number on this page was recomputed from our stored results on 9 Oct 2026. Evidence marks: filled for our own runs, half for a figure others report, dashed for an illustration.
Our measurements
- 1 M1 Pilot: Gemma 4 E2B on a laptop, 12 articles, 17 kinds of handling, our key · run 26 Sep 2026 Chart data: pilot-vs-main, detection-trace, pilot-heat, dose-response (pilot).
- 2 M2 Main study: Gemma 4 26B on a rented A100, 150 pages, 18 kinds of handling, our key · run 28 Sep 2026 Chart data: study-b-tactics, dose-response, format-tactics, room-and-strength.
- 3 M3 Site map: room per page type on one real website (not named), scored by Gemma 4 E2B · run 27 Sep 2026 Chart data: room-by-page-type (aggregates only).
- 4 M4 Next-word probabilities for one sentence, Gemma 4 E2B · run 29 Sep 2026 Chart data: next-word-path.
Outside sources
- 5 R1 Anthropic: How Claude's text watermark works · 14 Aug 2026 · accessed 26 Sep 2026
- 6 R2 Claude Help Center: How Claude marks AI-generated content · updated 22 Sep 2026 · accessed 26 Sep 2026
- 7 R3 Dathathri et al., Scalable watermarking for identifying large language model outputs (Nature 634) · 23 Oct 2024 · accessed 25 Sep 2026
- 8 R4 Google DeepMind: Watermarking AI-generated text and video with SynthID · 14 May 2024 · accessed 26 Sep 2026
- 9 R5 Kirchenbauer et al., A Watermark for Large Language Models (ICML 2023; includes the emoji attack) · 24 Jan 2023 · accessed 25 Sep 2026
- 10 R6 Scott Aaronson, My AI Safety Lecture for UT Effective Altruism · Nov 2022 · accessed 25 Sep 2026
- 11 R7 Hou et al., SemStamp (NAACL 2024) · Oct 2023 · accessed 25 Sep 2026
- 12 R8 Liang et al., Watermark under Fire (WaterPark), Findings of EMNLP 2025 · Nov 2024 · accessed 26 Sep 2026
- 13 R9 Tamim et al., AI Watermark Evidence Fails Forensic Readiness · Jul 2026 · accessed 26 Sep 2026
- 14 R10 He et al., Can Watermarks Survive Translation? (ACL 2024) · Feb 2024 · accessed 25 Sep 2026
- 15 R11 Kirchenbauer et al., On the Reliability of Watermarks for LLMs (ICLR 2024) · Jun 2023 · accessed 25 Sep 2026
- 16 R12 Jovanovic et al., Watermark Stealing (ICML 2024) · Feb 2024 · accessed 25 Sep 2026
- 17 R13 Liang et al., GPT detectors are biased against non-native English writers (Patterns) · Apr 2023 · accessed 25 Sep 2026
- 18 R14 EU AI Act, Article 50 · this duty applies from 2 Aug 2026 · accessed 25 Sep 2026
- 19 R15 European Commission: Code of Practice on AI-generated content · final 10 Jun 2026 · accessed 26 Sep 2026
- 20 R16 Freshfields: EU AI Act unpacked #33, the final Code of Practice on transparency · 25 Jun 2026 · accessed 26 Sep 2026
- 21 R17 The Decoder: Anthropic opens Claude AI text detection to regulators, media, fact-checkers and others · about 1 Sep 2026 · accessed 25 Sep 2026
- 22 R18 Google Search Status Dashboard: August 2026 spam update · 18 to 21 Aug 2026 · accessed 25 Sep 2026
- 23 R19 Google Search Status Dashboard: September 2026 spam update · from 24 Sep 2026 · accessed 25 Sep 2026; completion searched 9 Oct 2026
- 24 R20 Google Search Central: Spam policies · current · accessed 26 Sep 2026
- 25 R21 Glenn Gabe: August 2026 Google spam update case studies · 31 Aug 2026 · accessed 25 Sep 2026
- 26 R22 Search Engine Land: SE Ranking impact data, August 2026 spam update · 27 Aug 2026 · accessed 25 Sep 2026
- 27 R23 First Page Sage: Impact of AI watermarking on SEO and GEO results · 2026 · accessed 29 Sep 2026
- 28 R24 Ahrefs: AI content in top-10 results · 2026 · accessed 29 Sep 2026
- 29 R25 Lily Ray: It works until it doesn't · 2026 · accessed 29 Sep 2026
- 30 R26 Google Research: SAFE (Scaled Abuse Forensics Examiner) · 2026 · accessed 26 Sep 2026
- 31 R27 Search Engine Journal: SAFE · Sep 2026 · accessed 26 Sep 2026
- 32 R28 Mistral Legal Center: AI governance, Vibe Code · current · accessed 26 Sep 2026
- 33 R29 OpenAI: Supporting the EU trustworthy AI ecosystem · 11 Jun 2026 · accessed 26 Sep 2026
- 34 R30 Hugging Face: Introducing SynthID Text · 23 Oct 2024 · accessed 25 Sep 2026
- 35 R31 SynthID-Text reference code · Oct 2024 · accessed 25 Sep 2026
- 36 R32 Claude Content Checker · about 2 Sep 2026 · accessed 25 Sep 2026
- 37 R33 Our verified research brief (unpublished) · 26 Sep 2026
- 38 R34 The Word-Choice Watermark, earlier interactive version (unpublished) · 29 Sep 2026