Reference

AI Detection Glossary

The terms you’ll run into when reading about AI detectors and proof of authorship, defined in plain English.

AI detector: Software that estimates how likely it is that content was produced by an AI system. Its result is a probability, not proof.
AI-likeness score: The percentage or rating a detector gives. Higher means the text has more of the patterns the detector associates with AI writing.
Base rate: How common something is in the group being tested. When most writing checked is human, even a small false positive rate produces many wrongly flagged people.
Binoculars: A zero-shot detection method from 2024 research that compares how two related language models score the same text.
Burstiness: How much sentence length, structure, and predictability vary across a text. Human writing is usually burstier than AI writing.
C2PA / content credentials: An industry standard for attaching signed records to images, video, and audio that describe how a file was created and edited.
Classifier: A machine-learning model trained to sort inputs into categories, such as human-written or AI-generated.
Content credentials: Signed records attached to an image, video, or audio file describing which tools created or edited it. See C2PA.
False negative: AI-generated content that a detector labels as human-written.
False positive: Human-written content that a detector labels as AI-generated.
Hash: A short digital fingerprint of a piece of data. Changing even one character produces a completely different hash.
Hash chain: A sequence of records where each includes the hash of the one before it, so any later change to an earlier record is detectable.
Humanizer: A tool that rewrites AI text to make it look more human to detectors, usually by adding variation and less common words.
Large language model (LLM): An AI system trained on large amounts of text to predict and generate language. ChatGPT, Claude, and Gemini are built on LLMs.
Paraphrasing tool: Software that rewords text. Running AI text through one often lowers detector scores.
Paste: Text inserted all at once from the clipboard. Writing-process tools label pastes so reviewers can tell quotations from large blocks of outside text.
Perplexity: A measure of how surprised a language model is by a text. Low perplexity means predictable text, which detectors associate with AI.
Process evidence: Evidence of how a work was created, such as drafts, version history, or a writing record, rather than an analysis of the finished text.
Proof of authorship: Evidence showing who created a work and how, such as drafts, version history, or a sealed writing record.
Provenance: The documented history of where something came from and how it was made.
Sentence-level scoring: Detector results broken down by sentence, showing which passages drove the overall score.
Version history: Saved earlier versions of a document, kept automatically by editors such as Google Docs and Microsoft Word.
Watermark (AI): A hidden statistical signal deliberately added to AI output so a matching detector can recognize it later.
Zero-shot detection: Detection that uses existing language models directly, without training on labeled human and AI examples.