Techniques for marking content as machine-generated so it can be recognised later. The Coalition for Content Provenance and Authenticity (‘C2PA’) defines a standard for attaching a signed record of how a file was made, and some models embed statistical watermarks in what they produce. Both help; neither is a solution, because a determined party can strip provenance data and the absence of a mark proves nothing.
Technical view
C2PA content credentials are cryptographically signed manifests carried in file metadata, robust to inspection but easily removed by re-encoding or screenshotting unless durable variants are used. Statistical watermarking for text biases token selection within a secret partition so a detector can identify machine generation at a stated false positive rate; it survives light editing and degrades under paraphrase. Post-hoc detectors that classify text as AI-generated without a watermark are unreliable and produce false accusations, particularly against people writing in a second language, and should not be used for consequential decisions.
Related
Part of AI Terminology Index, section: Governance, Law and Regulation.