How Generative AI Is Worsening the Meta-Analysis Glut

JM Rosen, SK Gupta SK, R Kellermayer. Pediatr Gastroenterol Nutr. 2026; 1-4. doi:10.1002/jpn3.70634. Four decades in the making: generative artificial intelligence and the secondary-literature glut in biomedical publishing.

Key points:

  • “Narrative reviews, systematic reviews, and meta-analyses (articles that synthesize existing evidence rather than generate new observations) have been accumulating at a pace that outstrips the growth of primary research for more than three decades. Generative artificial intelligence (AI) … has removed the last practical constraint on it and made the process much faster.”
  • Systematic reviews by 2019 had seen “a greater than 20-fold rise across two decades, compared with only a 2.6-fold rise in PubMed-indexed publications generally….Ioannidis argued that the production of systematic reviews and meta-analyses had “reached epidemic proportions,” estimating that a large share were redundant, misleading, or conflicted, with some clinical questions accumulating more than a dozen overlapping meta-analyses of the identical evidence base, and with production increasingly concentrated among a small number of prolific sources (geographic, topical, and/or contracted) rather than spread evenly across the field.1“
  • “The tasks that constitute a systematic review, framing a question, searching, screening, extracting data, synthesizing findings, and drafting prose, are precisely the tasks at which large language models now perform with superficial competence…the technology is now fluent enough to produce a document with the surface features of a competent review, but not trustworthy enough to be left unsupervised”
  • “Fabricated references in PubMed Central had increased more than 12-fold since 2023, to 4046 instances; these were concentrated in exactly the kind of citation-dense writing on which reviews and meta-analyses depend for their evidentiary claims.9“
  • “Nearly all current policy operates at the level of disclosure: authors must state whether and how they used AI.”

Proposed Changes:

  • “First, a mandatory novelty statement… that distinguishes their submission from the existing literature and from what a generative AI system, prompted on the same topic, could already produce unassisted.”
  • “Second, a stated human contribution, an explicit account of the clinical judgment, methodological innovation, or original synthesis that the authors, not a language model, provided.”
Four decades of growth in PubMed‐indexed systematic‐review publication, predating generative AI by roughly three decades.Blue circles show raw PubMed publication‐type tag counts for “systematic review” reported by Ioannidis. Orange squares show independent,model‐based stratified‐sample estimates reported by Hoffmann et al. Green triangles show the authors’ own PubMed query for the raw“systematic review” publication‐type count; the open triangle for 2026 (41,603) is an anticipated full‐year total.

My take: AI has exponentially worsen the problem of flooding journals with unnecessary and untrustworthy systematic reviews.

Related blog posts: