Scientific papers all come packed as one fixed PDF bundle — what if AI let each piece move separately?

This explores how the format social scientists publish in, mostly the peer-reviewed PDF, shapes what evidence they can show and how readers weigh it. The corpus has one note that takes this on directly, and several others that come at it from the side. The direct one is Munger's argument that the PDF is a bundle Can AI help social science move beyond the peer-reviewed PDF?. A single paper has to act as an archive, a literature review, a theory statement, a methods record and a results report all at once. Evidence only enters the record if it fits into that fixed sequence. Munger's surprising claim is that the main obstacle to change isn't AI capability. It's the document form itself and the career incentives built around it. The tools to split these functions into separate, recombinable pieces already exist. The PDF is what keeps them stuck together.

You can see why this matters when you look at the new kinds of evidence AI is producing that don't fit a results section. LLM-driven simulations guided by causal models can test social hypotheses in settings like negotiation, bail and job interviews. They reliably show which direction an effect goes, but they can't say how big it is Can structural causal models automate social science with language models?. A paper format built around effect sizes and significance tests has no natural place for evidence that is directional only. A related idea is that LLM outputs shouldn't be treated as observations at all. They are draws from the model's learned assumptions, a 'subjective prior' shaped partly by how the prompt was written. They should influence conclusions only through an explicitly stated trust weight Should we treat LLM outputs as real empirical data?. That kind of evidence comes with a dial attached, and a static PDF has nowhere to put the dial.

Form also shapes credibility separately from content. An arXiv preprint shaped public debate about AI and science well before anyone reviewed it. MIT's later withdrawal request couldn't undo that Can unreviewed preprints shape scientific debate before peer review?. Looking like a paper was enough to do the work of evidence. At the reader's end, people prefer answers with more citations, and irrelevant citations raise trust almost as much as relevant ones Do users trust citations more when there are simply more of them?. The visible apparatus of scholarship works as a trust signal on its own. Research on persuasion points the same way: claims slipped in as assumed background persuade more than claims stated outright Why are presuppositions more persuasive than direct assertions?. That's worth keeping in mind for a genre where the literature review sets out what readers are expected to take as settled.

Some notes suggest what an unbundled form could look like. One paper-generation system requires authors to specify their evidence before seeing results, and keeps model judgment separate from checks a machine can verify Can separating judgment from verification improve research paper reliability?. That builds something like pre-registration into how the document is made. On the reading side, retrieval works better when it first builds a map of what role each part of a document plays Can building a document map first improve retrieval over long texts?. This hints that a document's internal roles, like 'this is the method' and 'this is the finding', could become handles that machines and readers use directly. There's also a caution. Horning argues that answer-style AI output presents frozen facts while hiding the social processes that produced them Do LLMs obscure the historical processes behind their answers?. Replacing the PDF with AI-generated summaries could hide even more of how the evidence was made, not less.

In short, the corpus has one sharp argument (Munger's) and a set of adjacent findings. Together they suggest that the form of a document decides which evidence can be shown, how much it gets trusted, and how easily it can be checked. The corpus doesn't yet include empirical studies comparing what social scientists actually report under different publication formats, so that part of the question stays open.