Two audits of the published literature find invalid or non-existent citations in 1.07 percent of 56,381 conference papers (1.61 percent in 2025, 80.9 percent above the 2020 to 2024 average) and an excess of unmatched references of 0.21 to 1.91 percent over the pre-LLM baseline in four preprint and journal corpora by August 2025. Both measure existence in papers written by researchers and neither establishes which tool, if any, produced a given reference.

The conference audit contributes a hand-confirmed count and the year-over-year change; the four-corpus audit contributes the excess over a pre-LLM baseline at far larger scale by automated matching. Together they show a downstream trace in 2025 in two samples: a rise of the per-paper share over 2020 to 2024 in one, an excess of unmatched references over a pre-LLM baseline as of August 2025 in the other; the units differ (papers, references) and the samples may overlap through arXiv. The step keeps the authors' own disclaimers: no causal attribution to assistants.

The join fails while both parents stand if the two audits are not independent readings of one trend: if the conference papers are largely the same documents as the arXiv corpus of the second audit, if both rest on the same matching databases so that one indexing gap produces both signals, or if the per-paper share and the per-reference excess turn out to move apart when computed on one corpus. Query: overlap of the 56,381 conference papers with the arXiv corpus, and both metrics computed on that overlap.

Hallucinated citations are flooding science. The measured level is about one paper in a hundred with at least one invalid citation, rising, with no attribution to a tool.