Can AI help social science move beyond the peer-reviewed PDF?

Kevin Munger, a political science professor writing on his Substack "Never Met a Science," argues that "the successful transition to the age of AI must entail greater diversity in the forms that quantitative social science takes." He frames this against a failed precedent: "quantitative social science did not have a successful transition to the internet age" because the field simply "took old paper journals and put them online with no innovation in form." He names "the tyranny of the peer-reviewed PDF" as the current bottleneck — "the only one for which we are officially recognized and rewarded" — even though many epistemically valuable actions "are really very poorly fit for being stored and communicated in this modality."

His reasoning is that the PDF bundles several distinct functions — "an archive of what actions the scientist has taken," plus "the literature review, the theoretical arguments, the methods section... the statistical results, the conclusions, the references" — that only made sense to combine when "there were many humans who were going to read these PDFs." He proposes "three dimensions on which new forms can transcend the PDF": a more explicit "meta-ontology" that tracks each bundled component "separately, as well as how they have been combined," knowledge kept current over time instead of frozen in a static document, and "aggressive empirical exploration." The excerpt develops only the first at length: the PDF's narrative form requires authors to "play a kind of trick," convincing readers that a specific RCT, convenience-sample survey, or game-theoretic model really connects to the real-world claim attached to it — a linkage "contemporary methodology pays very little attention to" even though "whether they hang together is more of an art form."

Munger's diagnosis puts the limiting factor in academic incentives and document form, not in AI capability — a different target from What stops AI from discovering science without human help?, which locates the bottleneck in model-side design gaps. For Munger, the reward structure built around the PDF is itself what has to change before capability becomes the binding constraint. His push for more diverse forms also sits awkwardly against Does AI help individual scientists while narrowing scientific focus?, which measures AI-augmented natural-science work moving toward narrower, data-rich topics rather than toward varied forms of output — diversity of form is not evidence of diversity of content. And his aside that peer review "is going to have to change as AI becomes more widespread" points at the same institution that Can one AI system complete a full research cycle end-to-end? shows being automated from the opposite direction: that system writes and reviews its own manuscript inside the existing PDF form, while Munger wants the form itself replaced.

The excerpt is a programmatic essay, not a study: it offers no instance of a journal, funder, or field actually adopting an unbundled or meta-ontological alternative, and it breaks off before developing the "time" and "rapid empirical iteration" dimensions it names, leaving the framework two-thirds unillustrated. Munger writes as an interested party — a journal editor proposing to redesign the system he edits within — and his confidence that "social science will muster the energy to reject the entrenched interests preserving the status quo" is a stated hope, not a measured trend. At the strength a single author's position essay supports, the implication is that institutional reward structures, not AI capability, may be the open design question for AI-and-science work — worth tracking against future evidence of whether unbundled forms get built and rewarded, but not yet shown by this excerpt.