The latest arXiv preprint challenges the framing of AI-text detection. Instead of asking whether a machine wrote a text, the authors argue that as AI becomes an integral and accepted part of writing, the more relevant question is how much of the text reflects human contribution. This shifts the goal from a binary label to a continuous measurement.

The paper's title—"Likelihood Is Not Authorship"—signals its central concern: statistical likelihood, often used by detectors, may not be a valid proxy for authorship. The abstract asks whether we can reliably measure human contribution in AI-assisted writing at all, suggesting that current methods may conflate probability with provenance.

Because the source is only an abstract, the specifics of the proposed alternative are not detailed. But the significance is clear: if likelihood cannot distinguish human effort from machine output, then the growing reliance on AI detectors for evaluation, plagiarism, or authenticity may be misplaced. The paper calls for a more nuanced approach to understanding authorship in an AI-assisted world.