Glossary
The vocabulary of AI text provenance, in the order you are likely to meet it.
- Paraphrase pass
- One full rewriting sweep over a text that changes the wording while keeping the meaning. The simplest known way to remove a statistical watermark, and the standard attack used in detector research.
- Watermark (text)
- A hidden statistical pattern a model embeds in its word choices during generation, detectable later by a party holding the key. Not visible, not a hidden character, and not metadata.
- Statistical watermark
- A watermark that works by biasing which tokens the model picks, so that a detector counting token patterns can tell marked text from unmarked text with a stated confidence.
- Green list
- In the Kirchenbauer scheme, the keyed subset of vocabulary the model is nudged toward at each step. The detector counts how often green tokens appear.
- Tournament sampling
- The SynthID-Text method: candidate tokens compete in rounds judged by keyed random functions, producing a detectable pattern with minimal effect on quality.
- SynthID
- Google DeepMind's family of watermarks for images, audio, video and text. The text version has run in Gemini since May 2024 and is open-sourced.
- textGrain
- OpenAI's text watermark, announced October 5, 2026. EU-only by default in ChatGPT and Codex, opt-in in the API, detector gated.
- Detector (classifier)
- A model trained to guess whether text was machine-written from its style and predictability. Needs no key, works on any model's output, and is wrong at a rate that varies by domain and writer.
- Zero-shot detector
- A detector that uses pretrained language models directly, without training on labelled examples. Binoculars is the common example.
- Perplexity
- A measure of how surprising a text is to a language model. Low perplexity means predictable text, which classifiers read as machine-like. Non-native writers often produce low-perplexity text.
- Burstiness
- Variation in sentence length and structure across a text. Early detectors treated low burstiness as a machine signal; it is easy to fake and easy to find in human writing.
- False positive
- Human text flagged as AI. The error that harms people. The 2023 Stanford study measured a 61.3 percent average rate on non-native essays.
- False negative
- Machine text that a detector misses. Paraphrasing, short length and constrained domains all raise this rate.
- Humanizer
- A paraphrasing tool marketed on passing AI detectors. Also called a bypasser.
- Bypasser detection
- Turnitin's name for a signal trained on humanizer output, launched August 27, 2025, that flags rewritten machine text.
- Provenance
- A record of where content came from and what was done to it. Delivered either as signed metadata (C2PA) or as an invisible mark.
- C2PA
- The Coalition for Content Provenance and Authenticity standard for signed manifests attached to media files. Also marketed as Content Credentials.
- Secured metadata
- The EU Code of Practice's term for provenance data that is cryptographically signed so tampering is detectable. The first of its two marking layers.
- Two-layer marking
- The EU Code of Practice approach: secured metadata plus an imperceptible watermark, so that a signal survives when either layer is stripped.
- Non-distortionary
- A watermark property meaning the average distribution of outputs is unchanged, so marked text is not measurably worse.
- Retrieval defence
- Detecting machine text by matching it against a provider's record of past outputs, which survives paraphrasing because meaning is preserved. Proposed by Krishna et al., 2023.
- Process evidence
- Proof of how a text was produced: version history, drafts, typing replay, notes. Unaffected by rewriting, which is why it outranks any score.
- Article 50
- The transparency article of the EU AI Act, applying from August 2, 2026. Providers mark output; deployers disclose deepfakes and public-interest AI text.
- Code of Practice (transparency)
- The voluntary EU code, final June 10, 2026, that gives signatories a presumption of conformity with Article 50 in exchange for two-layer marking and detector access.
- AI Transparency Act (California)
- SB 942 as amended by AB 853. Provider duties from August 2, 2026; platform duties, including not stripping provenance, from January 1, 2027.
- RAID
- A shared benchmark of more than six million generations with eleven adversarial attacks and a public leaderboard, used to compare detectors fairly.
- Token
- The unit a language model reads and writes, roughly a word or part of a word. Watermark strength is usually quoted per number of tokens.