Persuasion Index scores text across 15 rhetorical dimensions

A new arXiv preprint proposes Persuasion Index, or PI, as a taxonomy of 15 dimensions for analyzing rhetorical cues in text, grounded in persuasion theories from psychology and communication. The authors ship it with one transparent implementation using 55 sub-features built from lexicons and rule-based detectors, and describe the taxonomy as modular so individual detectors can be replaced without breaking the theoretical structure.

The framing worth paying attention to is who the authors say the tool is for. The abstract puts detecting information manipulation, improving AI safety, and advancing public health communication in the same opening sentence. Three audiences that currently reach for different, often opaque classifiers to answer versions of the same question: what in this text is doing the persuading? PI's answer is a shared feature space you can inspect.

Evaluation is on four public datasets varying in domain, style, and outcome measures. On those, the paper reports that linear models over PI features "carry meaningful predictive signal while remaining computationally lightweight," and that dimension-level analyses reveal "recurring associations between PI dimensions and persuasion outcomes across datasets" alongside topic- and stance-specific variation.

The honest caveat is that a rules-and-lexicons implementation is a different tool from an end-to-end LLM classifier, and the abstract does not name the four datasets, spell out the 15 dimensions, or benchmark PI against LLM-based baselines. So the write-up is a scaffold rather than a settled result. What it does give safety and communications teams is a principled, auditable vocabulary for why a piece of content reads as persuasive, released as an open-source package and a web interface.