A recent large-scale study warns that AI-generated fake citations are rapidly flooding scientific literature, undermining the integrity of research across major repositories. The surge is directly linked to the widespread adoption of large language models in academic writing, where automated systems frequently invent references that do not exist. Researchers auditing over 2.5 million scientific papers discovered an estimated 146,900 hallucinated citations in 2025 alone. The audit covered four primary repositories: arXiv, bioRxiv, SSRN, and PubMed Central. The findings indicate that this is not an isolated issue caused by a few individuals but a systemic problem where many papers contain small numbers of non-existent references. This pattern suggests that researchers are increasingly relying on AI tools to generate content without sufficiently verifying the output. The study, posted on the arXiv preprint server, analyzed 111 million references using a combination of automated and manual checks. Over 95 percent of citations could be matched to real publications. For the remaining entries, researchers corrected minor typos or used search engines to find obscure matches. To isolate the impact of AI, the team compared unmatched citation rates before and after the rise of consumer-grade large language models in 2023. The data revealed a sharp increase in fake references starting in mid-2024. The investigation highlighted several concerning trends regarding who is responsible for these errors. Early-career scientists and small research teams were found to be most likely to include hallucinated citations. In some cases, these researchers reported a threefold increase in productivity following the adoption of AI. Furthermore, the hallucinated references disproportionately credited prominent and male scholars, potentially reinforcing existing inequalities in scientific recognition. Generative AI models are designed to predict the next word based on patterns in their training data rather than to access or verify factual information. Consequently, they can produce output that sounds plausible and is internally consistent while being entirely fabricated. This capability poses a significant risk because scientific progress depends on building upon a solid foundation of existing, verified knowledge. The study also exposed critical gaps in the current peer-review and publication infrastructure. While moderation systems on platforms like arXiv are in place, they failed to catch the vast majority of these errors. Approximately 78.8 percent of non-existent citations still appeared on the platform, passing through preprint moderation, journal editing, and peer review processes. Scientists warn that without intervention, the infiltration of hallucinated content threatens the reliability and equity of global scientific knowledge production. The consequences could extend far beyond academic publishing, potentially influencing public policy and the general public's understanding of scientific discoveries. The findings underscore the urgent need for better validation tools and revised editorial standards to preserve the trustworthiness of the scientific record.