Once a paper is retracted, do copies on mirror sites and shadow libraries keep it circulating long after the notice?
How do mirror sites and shadow libraries perpetuate retracted papers?
This explores what happens to a paper after it's been retracted or disowned, specifically how copies kept on mirror sites and shadow libraries keep it in circulation. The collection doesn't cover that directly, but it does cover the problem one step over: discredited research stays in circulation, and retrieval systems pick it up.
This explores how retracted papers keep circulating after they've been officially pulled, through mirrors, shadow libraries and other copies that never get the retraction notice. The collection has nothing specifically on mirror sites or shadow libraries, so it can't tell you how those archives handle retractions. What it does have is useful nearby material: why pulling a paper rarely undoes its influence, and what happens when AI systems retrieve from collections that contain bad documents.
The clearest case is MIT's public statement that it had no confidence in an AI-and-science preprint, followed by a request that arXiv withdraw it Can unreviewed preprints shape scientific debate before peer review?. The lesson goes beyond any one archive. By the time the institution acted, the paper had already shaped months of discussion, and that discussion had been cited, quoted and built on. A withdrawal changes the paper's status. It doesn't reach back into every conversation the paper started. Mirror sites and shadow libraries would make this worse, since they keep the original file, but the deeper problem is that the ideas spread before the paper was pulled.
The part you might not expect: the newest place for retracted work to keep living may be the document collections that AI systems search. When a RAG system pulls from a corpus, it treats every document as equally available evidence, and nothing in the basic setup asks whether a source has been retracted. The research on corpus poisoning is about deliberately planted documents, not retractions, but the defenses carry over. One approach splits the corpus into partitions so that no single document can dominate an answer. Another flags documents whose relevance collapses in suspicious ways when parts of the text are masked Can we defend RAG systems from corpus poisoning without retraining?. A related idea is grounded refusal: a system built over noisy historical newspapers was designed to say "I don't know" rather than answer from weak evidence Can RAG systems refuse to answer without reliable evidence?. Neither approach checks retraction status, though, and the broader point is that retrieval measures how related a document is to the question, not whether it's reliable Where do retrieval systems fail and why?.
There's also an odd mirror image of the retraction problem. Papers routinely lose the experiments that failed, because the story-driven format of a paper leaves out dead ends. One proposal is to publish research as packages that keep those failed branches as part of the record Can research papers preserve the experiments that failed?. So the scientific record has two problems at once. Work that should have died keeps circulating, and knowledge worth keeping gets erased. In both cases, what survives depends on how the work is packaged and copied, not on whether it's right.
The flood of AI-generated papers raises the stakes. Hundreds of plausible papers with made-up justifications can now be produced automatically Can AI generate hundreds of fake academic papers automatically?, and ICLR found that fabricated references were the easiest thing to catch and desk-reject How can conferences detect and handle LLM misuse in peer review?. Every rejected or withdrawn paper still leaves copies behind. If you want the actual mechanics of mirrors and shadow libraries, you'll need sources outside this collection. The question this collection opens up instead is whether AI search tools will learn to recognize a retraction, or will keep surfacing withdrawn work as if nothing happened.
Sources 7 notes
MIT's case demonstrates that an arXiv preprint shaped AI and science discussions extensively despite never undergoing peer review. When the institution later raised reliability concerns, the damage to discourse had already occurred.
RAGPart and RAGMask provide lightweight, retraining-free defenses that operate at the retrieval layer. RAGPart bounds poisoned-document influence via partitioned retriever learning; RAGMask flags suspicious documents through abnormal similarity collapse under token masking.
A multilingual RAG system for noisy historical newspapers succeeds by aggressively expanding retrieval while constraining generation to only grounded answers. The grounded-refusal prompt prevents hallucination when OCR errors and language drift degrade source quality, trading coverage for integrity.
RAG systems fail at three structural levels: adaptive triggering (fixed intervals waste context), semantic-task mismatch (embeddings measure association, not relevance), and mathematical limits (embedding dimension constrains representable document sets). These require fundamentally different retrieval approaches, not tuning.
Publishing imposes a Storytelling Tax (erasing process, failed branches, tacit reasoning) and Engineering Tax (omitting implementation specs). Agent-Native Research Artifacts address both by packaging logic, executable code, exploration graphs of failures, and evidence grounding—treating rejected branches as publishable deliverables rather than editorial casualties.
Show all 7 sources
A demonstration showed LLMs generating 288 complete finance papers from 96 statistically significant signals, each with invented theoretical justifications and fabricated citations, proving academic HARKing can be automated at scale.
Program chairs used imperfect detectors as one input for area chairs rather than automated filters, but desk-rejected papers with confirmed fabricated references as a tractable enforcement point. Multiple human review steps mitigated false positives.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Stop Automating Peer Review Without Rigorous Evaluation
- CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning
- Searching for Best Practices in Retrieval-Augmented Generation
- Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs
- Chain-of-Retrieval Augmented Generation
- AI for Auto-Research: Roadmap & User Guide
- RAG Does Not Work for Enterprises
- You Don't Need Pre-built Graphs for RAG: Retrieval Augmented Generation with Adaptive Reasoning Structures