What is authorship attribution?

Definition

Authorship attribution asks whether a person, model, organization, or combination of contributors produced material. Evidence can include drafts, edit history, provenance records, writing patterns, watermarks, account activity, testimony, and the circumstances in which the work appeared.

A watermark detection result is not a complete authorship test because it may show only that a compatible model influenced part of the text. Fair attribution should account for assistance, editing, mixed authorship, detector uncertainty, and plausible alternative explanations before reaching a consequential conclusion.

ELI5

Authorship attribution is the process of working out who or what created a piece of work. It can consider drafts, writing patterns, saved history, tool records, and other evidence rather than relying on one clue.

For example, a detected watermark may show that an AI model helped with a document, but it cannot by itself show whether the model edited two sentences or wrote everything. A fair decision needs the wider history of how the work was made.

Frequently asked questions

What evidence can support authorship attribution?

Drafts, revision history, provenance, source files, account records, writing patterns, watermarks, testimony, and contextual evidence can contribute.

Why is mixed human and AI authorship difficult to attribute?

A final work can combine human writing, model generation, model editing, and repeated revisions without preserving clear boundaries between them.

Videos explaining authorship attribution

  1. A flat document fingerprint beside the words AI Text Leaves a Trace
  2. Dario Amodei, Nick Saraev and Jack Roberts beside the headline Pace the Frontier? on a black background.