When AI Writes Your Reading List, It Sees Roughly Half the Library

When a researcher, journalist, or policy team asks a large language model to scan the scholarly literature, the reading list it returns has already been filtered by a force that has nothing to do with relevance or quality: whether each paper is free to read online. A new exploratory study in Accountability in Research puts a number on that hidden filter and gives the pattern a name: "AI-induced evidence skew."

The study's authors prompted nine widely used LLMs to generate literature-supported content, then checked the accessibility of the references the models actually returned. Among 78 verifiable references, a mean of 95.38% had a freely available full text, and 87.43% were open access. Open access, in plain English, means the paper is free to read without a subscription, usually under a license that also permits reuse. The baseline the authors compared against was the open-access share of raw search results from three scholarly databases, a mean of 57.83%. The gap, roughly 30 to 38 percentage points depending on which slice you take, is the size of an entire publishing model.

It is also, the authors stress, illustrative rather than definitive. Their sample is small, the prompts are researcher-driven, and the paper itself is tagged "exploratory and non-generalizable." The point of the number is not to claim that 95% of AI-served research is free. The point is to show that the filter exists at all, and that it is large enough to reshape which findings end up cited in a literature review.

That is the structural problem the paper is actually trying to surface. The concern is not that AI tools favor weak papers; open access and low quality are not the same thing, and the authors say so explicitly. The concern is that the accessibility of a paper has become an unintended determinant of whether it enters an AI-mediated evidence base at all.

How the skew gets produced is the more useful part of the paper. The authors break it into four convergent pathways, and the order matters.

First, representation in training data. LLMs are trained on snapshots of the open web. The wider the crawl, the more the model's prior over which papers "exist" tilts toward freely indexed work. Second, retrieval-access constraints. When a model searches the web or queries an index to ground its answer, it can only return papers it can actually fetch. A paywalled paper behind a login wall is, to many retrieval pipelines, effectively a non-paper.

Third, human verification practices. Researchers tend to spot-check the references AI tools give them by clicking the link. A link that opens a full PDF passes the check; a link that lands on a paywall fails it, and the failure is more memorable than the success. Fourth, system- or user-imposed access restrictions, including the deliberate instruction to "use only open-access sources" or the integration of retrieval tools that themselves sit behind institutional credentials.

These four pathways compound rather than compete. A paper that is under-represented in training data is also the one most likely to be misdescribed or missed by retrieval, and therefore the one most likely to be cut during verification. By the time a paper appears in an AI-generated reading list, or does not, its absence is the product of a stack, not a single decision.

The mechanism is not new in name. A 2023 Nature commentary titled "Audit AI search tools now" warned that the rapid rollout of AI search products in research could distort the literature they touch. A 2024 paper presented at KDD (the ACM's annual data-mining conference) on bias and unfairness in information-retrieval systems in the LLM era described, at a technical level, how retrieval pipelines can re-weight sources by features that have nothing to do with relevance, the same dynamic the new paper now applies to scholarly work. An industry write-up from IntuitionLabs frames paywalled full text as the practical barrier it has become for AI research tools, and a 2025 Frontiers article on AI literature retractions situates AI-related integrity concerns inside the publishing system more broadly. What the new study adds is the name, the four-pathway framing, and a concrete, if small, empirical illustration of the size of the gap.

The constructive path the authors propose is closer to disclosure than to defiance. They call for transparent reporting of AI use, explicit acknowledgment of which retrieval sources a model drew on, and verification of AI-generated references against a database rather than a link click. For a researcher or newsroom that already leans on AI assistance, the smallest change with the largest effect is to ask the tool, in plain language, which papers it could not access and why. The answer is often the most informative part of the response.

The next move, the authors argue, is empirical. The 78-reference illustration is the seed. The larger study, across more disciplines, more platforms, more languages, and more than nine tools, has not been done. For now, the safe working assumption for any AI-assisted literature review is that the reading list is shorter than it looks, and the papers you cannot see are the ones the system was least able to fetch.