The vocabulary of AI detection

Photo Response Non-Uniformity — the unique sensor noise fingerprint every camera leaves on its photos.

Fast Fourier Transform — converts an image to the frequency domain where AI models leave detectable signatures.

Exchangeable Image File Format — metadata embedded in photos that records camera, lens, exposure, GPS and more.

Google's invisible watermark embedded in images and text generated by their AI models.

Coalition for Content Provenance and Authenticity — a metadata standard that records how an image was created and edited.

A pattern unique to a specific generative model that detectors use to identify the source generator.

The subtle, repeatable patterns that diffusion models leave in their output — a primary target for AI image detectors.

A forensic technique that resaves an image at known quality to expose regions edited or generated differently from the rest.

Adding controlled, camera-like noise to an image to replace the sensor noise AI generators omit.

The telltale compression history of a real photo — and a signal detectors use against single-pass AI exports.

A measure of how predictable text is to a language model — low perplexity is a strong AI signal.

The variance in sentence length and structure — human writing is bursty, AI writing is uniform.

A tool that rewrites AI-generated text so its statistical signature reads as human-written.

A tool that restates text in different words — related to, but weaker than, a true AI humanizer for detection.

The recurring words, phrases and rhythms that give away machine-written text.

The long dash (—) whose heavy, uniform use has become a recognisable marker of AI-written text.

The statistical analysis of writing style — the foundation most AI text detectors are built on.

When an AI detector flags genuinely human-written text as machine-generated.

The percentage or probability an AI text detector assigns — an estimate, not a measurement.

Synthetic or manipulated video and audio depicting real people — the highest-scrutiny category for detectors and platforms.

An imperceptible signal embedded in AI-generated content to identify it as synthetic.

The process of modifying AI-generated content so it passes detectors as human-created.

The consumer-facing name for C2PA provenance metadata — the "nutrition label" platforms read to apply AI tags.

The "AI Info" / "made with AI" tag platforms attach to flagged content — often with a reach penalty.

The compressed internal representation a generative model works in — where an image exists before it becomes pixels.