A glossary of AI detection and image forensics

In shortThe terms that appear in a ProofVerity report and in conversations about AI detection, defined in one or two sentences each.

Last reviewed 28 September 2026 · ProofVerity

Terms

AI detector: Software that estimates how likely a text, image or other content was produced by a generative model. Good detectors report probabilities with error rates; bad ones report verdicts.
Perplexity: A measure of how surprising each word is given the words before it, according to a language model. Machine-written prose tends to have low perplexity; human writing is less predictable.
Burstiness: How much sentence length and structure vary across a passage. Human writing is bursty; model output tends to be even.
Marker phrases: Constructions that language models over-produce relative to people, such as "it is important to note" or "delve into". A signal, never a proof.
Style break: A passage whose statistics differ sharply from the rest of the same author’s document. Evidence that a section was written or rewritten by someone, or something, else.
Calibrated probability: A score whose stated confidence matches its measured accuracy on that kind of document. When a calibrated detector says 94%, about 94 in 100 such spans were machine-written in evaluation.
False positive rate: The share of human-written spans a detector wrongly marks at a given threshold. ProofVerity publishes it per document type, next to each probability.
Verdict: A yes-or-no judgement about authorship. ProofVerity does not issue them; it reports evidence for a person to weigh.
Hallucinated citation: A reference invented by a language model: plausible authors, a real venue, an article that does not exist. Reported by ProofVerity as "not found", with the nearest real match.
Misquoted source: A real source that does not make the claim attributed to it. Reported with the discrepancy.
Misattribution: A claim that exists but belongs to a different work, author or year than the one cited.
DOI: Digital Object Identifier: a persistent identifier for scholarly works. Resolving a DOI is the first step in verifying a citation.
Content credentials (C2PA): Cryptographically signed provenance metadata attached to some images and videos by cameras and editing software. Present credentials prove history; absent credentials prove nothing.
EXIF: Metadata written into image files by cameras and software: device, time, settings. Easily stripped or edited, so a weak signal on its own.
Error level analysis (ELA): A technique that recompresses an image and compares error levels across regions. Areas edited and saved separately stand out.
Clone stamping: Copying a patch of an image onto another part of the same image to hide or duplicate something. Detected by finding repeated patches.
Splicing: Combining parts of two or more images into one. Detected at the boundaries, where noise and colour statistics change.
Inpainting: Regenerating a region of a real photo with a model, for example to remove or add an object. Reported as an altered region with its share of the frame.
Chain of custody: The record of who submitted an asset, when, what was checked, what was found and who decided. Exported with every ProofVerity report.
Evidence appendix: The section of an exported report listing, for every flag, the signals that produced it and the sources checked.
Credit: ProofVerity’s billing unit: one document of up to 100 pages, or four images.