A field guide to AI text: who marks it, who detects it, who rewrites it.
The model companies now sign their words with invisible statistical marks. Detectors guess. Rewriters erase. The rules changed this year, and the tools change monthly. This is the dated, independent map, written for people and readable by machines.
The results suggest that further research is needed to confirm these findings.
The results suggest that further research is needed to confirm these findings. These results hint that more work will be needed before anyone trusts them.
Highlighted words carry a statistical mark. One rewrite keeps the meaning, changes the words, and the mark is gone. Every entry on this site tracks that gap.
The state of play
Six facts that explain the field as of October 31, 2026. Each links to its entry and its primary source.
Guides
Answer first, then the evidence, then the sources.
- How AI text watermarks work
A plain explanation of statistical text watermarks (SynthID, textGrain, the Claude watermark): how the mark is embedded in word choice, how detection works, why length matters, and what a positive result does and does not prove.
- Which AI models watermark their text?
A dated table of which AI models and products apply a text watermark, where it applies, whether users can opt out, and who can detect it: Gemini, Claude, ChatGPT, Codex, the OpenAI API, and open-weight models.
- Does paraphrasing remove AI watermarks?
Yes, a full rewrite removes a statistical text watermark, and light edits weaken it in proportion to how many words change. The numbers from Anthropic, OpenAI and the research literature, and what survives a rewrite.
- How accurate are AI detectors?
What detector accuracy claims mean, why vendor numbers are not comparable, the false positive problem for non-native writers, and where to find a fair comparison.
- How to check whether text is AI-generated
An honest method for checking text for AI generation: provider watermarks (gated), classifier detectors (fallible), provenance metadata (rare for text), and process evidence (the one that holds up). What each can and cannot tell you.
- What to do when a detector flags your writing
A practical guide for students and writers flagged by an AI detector: the evidence that proves your process, what the vendors themselves say about their scores, the research to cite, and how to respond.
- Turnitin AI detection, explained
How Turnitin's AI writing indicator works, what the percentage means, what bypasser detection added in August 2025, the language and length limits, and why some universities switched it off.
- Humanizers versus detectors: the arms race
How AI humanizer tools and AI detectors have chased each other since 2023, what changed when Turnitin and Pangram began training on humanizer output, and why no result from last year still holds.
- EU AI Act Article 50, explained
What Article 50 of the EU AI Act requires for AI-generated text, the provider and deployer duties, the August 2, 2026 start, the Code of Practice's two-layer marking approach, and what it does not require.
- California's AI Transparency Act, explained
What California's AI Transparency Act requires after AB 853: provider disclosures and detection tools from August 2, 2026, platform provenance duties and the ban on stripping provenance from January 1, 2027, and capture-device rules from 2028.
Marks and provenance
How the big model providers now tag their own output, what the tags can and cannot tell you, and the open standards around them.
-
SynthID for text Google DeepMind Built into Gemini. Detector portal for images, audio and video; text detection through the open-source library.
A statistical watermark that tilts word choice while the model writes. Applied to text in the Gemini app since May 2024. The method was published in Nature in October 2024 and released as an open-source library with Hugging Face, so anyone can apply and detect a mark with their own key.
Detection needs the key that made the mark, so a Gemini mark can only be checked by Google or by a party Google authorises.
deepmind.google, checked 2026-10-31 -
Reference implementation of the text watermark for Hugging Face Transformers. Useful for anyone who wants to watermark their own model's output, or to study how the signal behaves after editing.
github.com, checked 2026-10-31 -
Developer guide to applying SynthID to Gemma and Gemini outputs, with the configuration options and the trade-offs between mark strength and output quality.
ai.google.dev, checked 2026-10-31 -
Claude text watermark Anthropic Applied automatically. Detection API in private preview for eligible organisations.
Announced August 14, 2026 and updated September 1, 2026. A version of SynthID-Text applied globally to future Claude models, with models released before August 2, 2026 to follow over the coming months. Detection uses Anthropic's key and answers only whether Claude likely wrote part of a passage. Access is open by application to regulators, law enforcement, media, fact-checkers, researchers, educators, EU civil society groups and enterprises with EU compliance duties.
Anthropic's own words: light editing probably will not remove the mark completely; a complete rewrite where every word is replaced will.
anthropic.com, checked 2026-10-31
Detectors
Classifiers that guess whether text came from a model. None of them is proof, every vendor publishes its own accuracy numbers, and the independent evidence is thinner than the marketing.
-
Turnitin AI writing detection Turnitin Institutional licence, sold as an add-on to Turnitin Originality and iThenticate.
The detector most students meet. It reports a percentage of a submission that it judges likely AI-written. AI bypasser detection, announced August 27, 2025, flags text that humanizer tools may have rewritten and counts it toward the AI-generated total. Bypasser detection is trained and tested for English only.
Several universities have switched the AI indicator off over false-positive concerns. Turnitin's own guide says a score should start a conversation, not decide an outcome.
turnitin.com, checked 2026-10-31 -
Turnitin's own description of how the model is trained, what the percentage means, which languages are covered and where it says the score should not be used alone.
guides.turnitin.com, checked 2026-10-31 -
A trained classifier that explicitly targets humanized text as well as raw model output. Pangram publishes its own comparisons and reports detection above 90 percent on high-quality humanized content.
Those figures are vendor-reported. No independent study of the same claim had been published when this entry was checked.
pangram.com, checked 2026-10-31 -
Education-focused detector with per-sentence highlighting, document reports and a writing-replay feature for Google Docs that shows how a draft was typed.
gptzero.me, checked 2026-10-31
Paraphrasers and humanizers
Two kinds of rewriting tools. Paraphrasers change wording for clarity, tone or length. Humanizers are sold on one promise: that the rewritten text will pass a detector. The second group is why marks and detectors keep changing.
-
The most used paraphraser. Modes for fluency, formality, simplification, expansion and shortening, plus a grammar checker, summarizer and citation generator.
quillbot.com, checked 2026-10-31 -
Sentence-level rewrites with casual and formal options, as a web app and browser extension.
wordtune.com, checked 2026-10-31 -
Grammar and clarity first, with a paraphrase feature. The right choice when you want fewer changes rather than a full rewrite.
grammarly.com, checked 2026-10-31 -
Style and tone rewriting from the translation company. Strong on English and German, with alternatives offered per sentence.
deepl.com, checked 2026-10-31
Research and benchmarks
The papers that defined the field, and the benchmark that keeps score. Read these before trusting any accuracy claim.
-
The green-list scheme that most later watermark work builds on: the model is nudged toward a secret subset of tokens, and a detector counts how often they appear.
arxiv.org, checked 2026-10-31 -
Tests marks against paraphrasing, copy-paste mixing and human rewriting. Finds that marks weaken but often remain detectable once enough text is available.
arxiv.org, checked 2026-10-31 -
Argues that recursive paraphrasing defeats both classifiers and watermarks, and that detection gets harder as models get closer to human text.
arxiv.org, checked 2026-10-31 -
Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense Krishna et al., 2023 Open access.
Introduces DIPPER, an 11-billion-parameter paraphraser, shows it breaks the detectors of the day, and proposes retrieval of previously generated outputs as a defence that paraphrasing cannot beat.
arxiv.org, checked 2026-10-31
Policy and law
The rules that now require marks, the platform policies that restrict removing them, and the dates that matter.
-
Transparency obligations. Providers of generative systems must mark output in a machine-readable way; deployers must disclose deepfakes and AI-written text published on matters of public interest. Applies from August 2, 2026, with a transition period for systems already on the market.
artificialintelligenceact.eu, checked 2026-10-31 -
The voluntary route to Article 50 compliance. Final text published June 10, 2026; the Commission found it adequate on July 8, 2026. Signatories commit to a two-layer approach, secured metadata plus watermarks, and to keeping marks non-removable to the best of their ability.
digital-strategy.ec.europa.eu, checked 2026-10-31 -
The Official Journal version of the AI Act, for the exact wording of Article 50 and the definitions it relies on.
eur-lex.europa.eu, checked 2026-10-31 -
AB 853 was signed October 13, 2025. Covered providers owe disclosures and detection tools from August 2, 2026. Large online platforms and generative AI hosting platforms are covered from January 1, 2027, including a ban on knowingly stripping provenance data. Capture device makers follow on January 1, 2028.
leginfo.legislature.ca.gov, checked 2026-10-31
Where the news comes from
The feeds we read to keep this site current. Vendor blogs announce changes first; the papers and the regulators decide whether they matter.
-
Watermark, detector-access and policy announcements for Claude.
anthropic.com, checked 2026-10-31 -
textGrain updates, EU rollout dates and provenance announcements.
openai.com, checked 2026-10-31 -
SynthID releases and research.
deepmind.google, checked 2026-10-31 -
Detector comparisons and humanizer analysis. Vendor-written, but unusually specific.
pangram.com, checked 2026-10-31
Paraphrase Pass opens
First edition of the directory: 7 provenance entries, 11 detectors, 14 rewriting tools, 10 papers and benchmarks, 8 policy texts, 12 news sources. Every link was checked on this date.
OpenAI announces textGrain
OpenAI's text watermark goes opt-in for API customers worldwide, off by default, and begins rolling out to ChatGPT and Codex output in the EU. Detector access is by application only. OpenAI's own figures: about 95 percent detection on 400-token passages at a 1 percent false positive rate, falling to 17 percent after a quarter of the words are swapped for synonyms.
Anthropic announces a global Claude text watermark
A SynthID-Text variant applied to future Claude models worldwide, with earlier models to follow. A detection API enters private preview for regulators, media, researchers, educators and enterprises with EU compliance needs. Anthropic says light editing probably leaves the mark; a full rewrite removes it.
EU Article 50 and California provider duties apply
The AI Act's transparency obligations take effect for providers and deployers, with a transition period for systems already on the market. On the same date, covered providers under California's AI Transparency Act owe disclosures and detection tools.