Plagiarism Ex Machina: When Originality Becomes a Statistical Effect

Large language models have destabilized the classical definition of plagiarism. My paper Plagiarism Ex Machina: Structural Appropriation in Large Language Models argues that the central problem is no longer only copying. It is the transformation of human-authored corpora into generative capacity under conditions where source lineage becomes difficult or impossible to recover.

Traditional plagiarism assumes an identifiable relation between source, copy, and author. Generative systems complicate that model. They absorb distributed textual labor, recombine learned structures, and produce outputs that may appear new even when their rhetorical, conceptual, stylistic, and argumentative competence depends on prior human work.

Structural appropriation

I call this condition structural appropriation. The issue is not necessarily verbatim reproduction. A model can generate enough surface variation to evade similarity-based notions of copying while still depending on patterns extracted from a vast archive of intellectual production. The resulting text can look original precisely because the source relation has become opaque.

This produces what the paper describes as synthetic originality. Originality becomes an effect of recombination depth and source opacity rather than clear evidence of autonomous intellectual creation. The system can reproduce the forms of expertise, argumentation, academic prose, legal reasoning, or technical explanation without exposing the authors and communities whose accumulated practices made those forms possible.

The problem of generative provenance

This is why provenance matters. If generative systems are increasingly used in scholarship, law, journalism, education, and professional writing, the inability to reconstruct intellectual dependency creates a new problem for attribution, authorship, academic integrity, and the political economy of knowledge.

The deeper question is not whether an AI copied a sentence. It is whether institutions are prepared to treat source-neutral fluency as original production when that fluency is built from distributed human labor that the system cannot transparently acknowledge at the point of generation.

Full paper: https://zenodo.org/records/20070859