'Made with AI' is too blunt for readers
A sticker saying "made with AI" can punish accurate work and still miss fluent nonsense. Provenance Density asks a better question: how much of this text is backed by verified claims? Transparency moves from authorship theater to evidence readers can inspect.
What the source reports
Zhang, Huang, Lee, Starner, and Rekimoto describe the "Fluency Trap." People trust polished falsehoods. They also discount accurate text once it is disclosed as AI-made. Their proposed Provenance Density interface shows how many claims in a text are verified. In a user study with 81 participants, the interface produced a +4.15 point gap in telling truth from fabrication. The effect size was d = 1.82. Participants with no signal showed no detectable discrimination. A 200-sample technical audit found retrieval density alone was not enough. On dynamic queries, the Consistency Veto carried much of the measured signal.
Original source
- Title
- Beyond "Made with AI": Visualizing Provenance Density to Mitigate the Transparency Penalty
- Author
- Zhang, Huang, Lee, Starner, Rekimoto
- Publication
- arXiv