[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"kb-article-why-llms-invent-academic-citations-and-how-to-stop-ghost-references-en":3,"ArticleBody_k7U1Igw9TCQnNYEdvijh3CTUwH14FuIAPPizqQXiJaw":107},{"article":4,"relatedArticles":76,"locale":66},{"id":5,"title":6,"slug":7,"content":8,"htmlContent":9,"excerpt":10,"category":11,"tags":12,"metaDescription":10,"wordCount":13,"readingTime":14,"publishedAt":15,"sources":16,"sourceCoverage":57,"transparency":58,"seo":63,"language":66,"featuredImage":67,"featuredImageCredit":68,"isFreeGeneration":72,"niche":73,"geoTakeaways":57,"geoFaq":57,"entities":57},"69a1a50e83962bbe60b24fbe","Why LLMs Invent Academic Citations—and How to Stop Ghost References","why-llms-invent-academic-citations-and-how-to-stop-ghost-references","## Introduction  \n\nLarge language models now assist with theses, grant proposals, and journal articles, often drafting sections and reference lists.  \n\nAcross 13 state-of-the-art models, hallucinated citation rates range from 14.23% to 94.93%—far too high to trust AI-generated bibliographies [11]. Fake references are now appearing in AI\u002FML and security venues, ACL conferences, and NeurIPS 2025 papers [5][11][12].  \n\nThis article explains why LLMs fabricate citations, how ghost references enter citation indices and metrics, why retrieval alone is insufficient, and what a realistic architecture for citation-safe AI use requires.  \n\n---\n\n## 1. The Ghost Citation Crisis: Scope and Evidence  \n\nGhost references—citations to non-existent papers—have become a measurable systemic risk.  \n\nKey findings:  \n\n- A benchmark of 13 leading LLMs across 40 domains found hallucinated citation rates between 14.23% and 94.93% when asked to generate references [11].  \n- Independent audits of popular tools report 18–69% fabricated citations in typical use; some systems produce fake references more than half the time they attempt to cite scholarly work [2][11].  \n\n📊 **Corpus-level contamination**  \n\nUsing the CiteVerifier framework, one study analyzed 2.2 million citations from 56,381 AI\u002FML and Security papers (2020–2025):  \n\n- 1.07% of papers contained invalid or fabricated citations.  \n- Such papers increased by 80.9% in 2025, coinciding with widespread LLM-assisted writing [11].  \n\nTop venues show similar issues:  \n\n- In >4,000 NeurIPS 2025 papers, at least 53 accepted papers contained hallucinated citations, including fabricated titles, authors, venues, and dead URLs [5].  \n- ACL conference analyses found incorrect arXiv IDs, broken links, and references to non-existent pages in otherwise solid work [12].  \n\n⚠️ **Persistence and spread**  \n\nExamples include:  \n\n- A non-existent paper accumulating over 40 citations in Google Scholar as authors repeatedly reused the same ghost reference [3][4].  \n- Editors receiving submissions citing articles “authored” by the editors themselves—papers that do not exist—indicating wholesale pasting of AI-generated references [3].  \n\n**Mini-conclusion:** Ghost citations are now a structural threat to scholarly reliability, already visible in leading venues and citation databases [11][12].  \n\n---\n\n## 2. Why LLMs Invent Citations: Mechanics of Fabrication  \n\nLLMs are not bibliographic search engines; they are pattern-completion systems.  \n\nWhen prompted to “add references,” a model predicts the next most probable token based on training data saturated with citation-like strings—author-year patterns, venues, DOIs—without an indexed catalog of real papers [2][11]. It learns how references *look*, not whether any specific one exists [2].  \n\n💡 **Pattern mastery without world anchoring**  \n\nConsequences:  \n\n- Models generate plausible author combinations, titles, and venues.  \n- There is no native link from a generated reference to CrossRef, arXiv, PubMed, or publisher databases [2][11].  \n\nTwo main failure modes follow:  \n\n- **Ghost references:** entirely fabricated but plausible works.  \n- **Citation unfaithfulness:** real papers cited for claims they do not support, often chosen for textual similarity rather than semantic relevance [1].  \n\nAsking models how “confident” they are in a citation does not help:  \n\n- They generate numbers that “sound right” given training patterns, biased toward high confidence [8][9].  \n- Re-running the same scoring prompt yields different values due to non-deterministic sampling, so there is no stable internal certainty [8].  \n\n⚠️ **Fluent scholarly style without factual binding**  \n\nInstruction tuning and RLHF optimize for helpful, fluent, decisive output. Models “write like scholars, cite like experts” [2], but:  \n\n- Nothing in the architecture enforces that references correspond to real works.  \n- Nothing guarantees that cited works actually support the associated claims [1][2][11].  \n\n**Mini-conclusion:** Citation fabrication is an emergent property of pattern-completion on scholarly text without binding to authoritative bibliographic databases [1][2].  \n\n---\n\n## 3. How Fake Citations Cross Into the Scholarly Record  \n\nGhost references would be manageable if they stayed in chat logs. The real damage occurs when they enter formal scholarly infrastructure.  \n\nObserved patterns:  \n\n- Instructors and editors find bogus citations in student papers that already appear in dozens of published articles, often with minor variations [3][4].  \n- Once a fake paper is cited in a publication, it gains *derived legitimacy*: others copy its references or trust Google Scholar’s representation without checking [1][3].  \n\n📊 **Feedback loops in citation infrastructure**  \n\nCitation indices like Google Scholar:  \n\n- Create citation records for references they cannot match to real documents, inferring stubs from bibliographies [1].  \n- Allow those stubs to accumulate further citations from human authors and LLM-assisted workflows, reinforcing the illusion of reality [1][4].  \n\nExamples include:  \n\n- A fabricated article on education governance listed as a highly cited paper in Google Scholar, despite never being written by the named authors [3].  \n- Student essays whose ghost references resolve to published articles that themselves cite the same non-existent studies [4].  \n\nNeurIPS 2025 analyses found:  \n\n- Hybrid citations blending elements from multiple real papers—authors from one, title motif from another, venue from a third—into a single unresolved reference [5].  \n- Invented authors, swapped venues, and dead URLs that still passed multi-reviewer peer review [5].  \n\n⚠️ **Distorted metrics and claims**  \n\nBecause citation counts feed into promotion, funding, and journal rankings:  \n\n- Ghost and zombie citations distort impact indicators [3][11].  \n- Imaginary papers with many citations inflate the perceived rigor of work that leans on them.  \n- Polluted citations become training data for future AI systems.  \n\n**Mini-conclusion:** Once ghost citations enter journals, conferences, and indices, they fully participate in academic metrics, making them hard to detect and unwind [1][3][11].  \n\n---\n\n## 4. Why RAG and Tools Help—but Do Not Fully Solve It  \n\nRetrieval-Augmented Generation (RAG) is often proposed as a fix: ground model outputs in retrieved documents so models can only cite what they see [1].  \n\nEmpirical results:  \n\n- RAG reduces the number of claims a model marks as “cannot assess,” because retrieval supplies more evidence [6][7].  \n- In machine-generated news fact-checking, adding search results led to fewer abstentions and more decisive judgments [6][7].  \n\n💼 **Grounded on what?**  \n\nThe same work found a key downside:  \n\n- Irrelevant or low-quality retrieval increased incorrect assessments, as models overfit to whatever evidence they were given [6][7].  \n- If retrieval includes polluted bibliographies or secondary mentions of ghost references, RAG can *amplify* fabrication and misattribution [1][6].  \n\nCommon RAG citation failure modes:  \n\n- Letting models “fill in” missing fields—guessing DOIs, venues, or years from partial snippets.  \n- Blending metadata from multiple documents into a synthetic reference [5][12].  \n- Citing real papers from retrieval results that do not actually support the claim (citation unfaithfulness) [1][11].  \n\n⚠️ **Evidence layer, not guarantee**  \n\nEven with RAG:  \n\n- Ghost references may be reduced but not eliminated.  \n- There is no automatic check that each citation both exists and semantically supports the argument [1][11].  \n\n**Mini-conclusion:** RAG is valuable infrastructure but must be paired with explicit verification and quality-controlled retrieval. Grounding alone does not prevent ghost or unfaithful citations [1][6][7].  \n\n---\n\n## 5. Human and Institutional Failure Modes: The Verification Gap  \n\nLLM “hallucinations” explain only part of the problem. Ghost citations spread because human and institutional practices leave a large verification gap.  \n\nSurvey data from 94 researchers [11]:  \n\n- 41.5% copy-paste BibTeX entries without checking.  \n- 44.4% take no action when encountering suspicious references.  \n- 76.7% of reviewers do not thoroughly check references.  \n- 80.0% never suspect fake citations might be present.  \n\nIn such conditions, LLM-generated fabrications pass through by default.  \n\n💡 **Overload and misaligned incentives**  \n\nThe ACL community notes that exponential submission growth and reviewer overload push peer review toward shallow checks [12]. NeurIPS papers with hallucinated citations each passed three or more reviewers, showing that even elite venues do not systematically scrutinize bibliographies [5].  \n\nA parallel from journalism:  \n\n- Ars Technica retracted an article after AI-generated quotes were presented as real statements by a named individual, violating core trust in quotation [10].  \n- Policies against unlabeled AI-generated content existed; enforcement at publication time failed [10].  \n\nSimilar patterns in scholarship:  \n\n- Students submit AI-written assignments with overlapping sets of fabricated references.  \n- Editors receive papers citing fake work attributed to themselves, showing authors did not read or sanity-check their own lists [3][4].  \n- Publication incentives reward volume and citation counts over meticulous evidentiary integrity [3][11].  \n\n⚠️ **Governance and workflow, not just models**  \n\nVague standards like “citations must be accurate” and weak enforcement allow ghost citations to cross the publication boundary [10][11][12].  \n\n**Mini-conclusion:** LLMs accelerate a long-standing tendency to skip citation verification. Addressing it requires changes to incentives, workflows, and governance, not only better models [10][11].  \n\n---\n\n## 6. A Preventive Architecture for Citation-Safe LLM Use  \n\nCitation integrity must be treated as a systems-design problem. Effective defenses span tooling, workflows, and policy.  \n\n### 6.1 Separate generation from verification  \n\nUse LLMs for drafting, not for authorizing references. Route every citation through an automated verifier (e.g., CiteVerifier) that [11]:  \n\n- Resolves DOIs, arXiv IDs, and URLs.  \n- Confirms titles, authors, venues, and years against canonical records.  \n- Flags non-existent or mismatched references for human review.  \n\n📊 **Design principle:** the model proposes; independent systems verify.  \n\n### 6.2 Evidence-based confidence, not model introspection  \n\nAvoid asking models how sure they are. Instead, compute trust from:  \n\n- Exact or near-exact matches between citation metadata and source documents.  \n- Successful URL or DOI resolution.  \n- Cross-document consistency in how a paper is described [8][9].  \n\nThis avoids relying on numerically expressed “confidence” that LLMs cannot reliably calibrate [8][9].  \n\n### 6.3 Constrained RAG and adversarial checking  \n\nFor RAG-based literature tools:  \n\n- Restrict citations to items that actually appear in the retrieved corpus; forbid invented DOIs or venues [1][6].  \n- Log retrieval traces for each citation so reviewers can inspect supporting documents.  \n- Use a second-pass “adversarial” model with retrieval access to challenge citations from the first pass, flagging suspect items for humans [6][7].  \n\n💼 **Use LLMs as both writers and critics**  \n\nModels are often better at spotting inconsistencies given evidence than at generating ground-truth citations. Allocate capacity to critique as well as draft.  \n\n### 6.4 Governance, review, and education  \n\nInstitutions should adopt explicit rules that:  \n\n- Treat unverified AI-generated references like fabricated quotes in journalism—prohibited unless clearly labeled and independently validated before publication [10].  \n- Update reviewer forms and editorial checklists to ask whether a sample of references was independently verified for existence and relevance [5][12].  \n- Train authors and students that AI tools are unreliable citation engines; ghost references are a known failure mode, and humans remain responsible [2][4].  \n\n⚠️ **Default stance:** AI-supplied citations are untrusted until proven otherwise.  \n\n**Mini-conclusion:** Robust citation safety combines constrained generation, independent verification, adversarial checking, and clear governance. No single layer suffices; together, they can prevent LLMs from poisoning citation ecosystems [10][11].  \n\n---\n\n## Conclusion: Design for Citation Safety Up Front  \n\nLLMs fabricate citations because they are optimized for plausible text, not verified bibliographies. Even elite models hallucinate 14.23–94.93% of citations when asked to generate references, and ghost citations already appear in NeurIPS, ACL, and other AI\u002FML venues [5][11][12].  \n\nRAG, browsing, and self-reported confidence cannot, by themselves, solve a problem rooted in both model behavior and human verification gaps [1][6][8][9]. Citation indices and overloaded peer review pipelines allow fabricated references to enter the scholarly record and accumulate illegitimate influence [1][11].  \n\nA resilient response treats every AI-generated citation as untrusted until proven otherwise. That means coupling retrieval with automated verifiers, using LLMs as adversarial fact-checkers, instrumenting workflows with audits, and updating editorial and educational policies so ghost citations are seen as a preventable systems failure, not an inevitable side effect of automation [2][4][10][11].  \n\nAny deployment of LLMs in scholarly or high-stakes settings should be designed for citation safety from the outset—before ghost references become embedded in the bodies of work institutions rely on.","\u003Ch2>Introduction\u003C\u002Fh2>\n\u003Cp>Large language models now assist with theses, grant proposals, and journal articles, often drafting sections and reference lists.\u003C\u002Fp>\n\u003Cp>Across 13 state-of-the-art models, hallucinated citation rates range from 14.23% to 94.93%—far too high to trust AI-generated bibliographies \u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>. Fake references are now appearing in AI\u002FML and security venues, ACL conferences, and NeurIPS 2025 papers \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>\u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>This article explains why LLMs fabricate citations, how ghost references enter citation indices and metrics, why retrieval alone is insufficient, and what a realistic architecture for citation-safe AI use requires.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>1. The Ghost Citation Crisis: Scope and Evidence\u003C\u002Fh2>\n\u003Cp>Ghost references—citations to non-existent papers—have become a measurable systemic risk.\u003C\u002Fp>\n\u003Cp>Key findings:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>A benchmark of 13 leading LLMs across 40 domains found hallucinated citation rates between 14.23% and 94.93% when asked to generate references \u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Independent audits of popular tools report 18–69% fabricated citations in typical use; some systems produce fake references more than half the time they attempt to cite scholarly work \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>📊 \u003Cstrong>Corpus-level contamination\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Using the CiteVerifier framework, one study analyzed 2.2 million citations from 56,381 AI\u002FML and Security papers (2020–2025):\u003C\u002Fp>\n\u003Cul>\n\u003Cli>1.07% of papers contained invalid or fabricated citations.\u003C\u002Fli>\n\u003Cli>Such papers increased by 80.9% in 2025, coinciding with widespread LLM-assisted writing \u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Top venues show similar issues:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>In >4,000 NeurIPS 2025 papers, at least 53 accepted papers contained hallucinated citations, including fabricated titles, authors, venues, and dead URLs \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>ACL conference analyses found incorrect arXiv IDs, broken links, and references to non-existent pages in otherwise solid work \u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>⚠️ \u003Cstrong>Persistence and spread\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Examples include:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>A non-existent paper accumulating over 40 citations in Google Scholar as authors repeatedly reused the same ghost reference \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Editors receiving submissions citing articles “authored” by the editors themselves—papers that do not exist—indicating wholesale pasting of AI-generated references \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>Mini-conclusion:\u003C\u002Fstrong> Ghost citations are now a structural threat to scholarly reliability, already visible in leading venues and citation databases \u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>\u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>2. Why LLMs Invent Citations: Mechanics of Fabrication\u003C\u002Fh2>\n\u003Cp>LLMs are not bibliographic search engines; they are pattern-completion systems.\u003C\u002Fp>\n\u003Cp>When prompted to “add references,” a model predicts the next most probable token based on training data saturated with citation-like strings—author-year patterns, venues, DOIs—without an indexed catalog of real papers \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>. It learns how references \u003Cem>look\u003C\u002Fem>, not whether any specific one exists \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>💡 \u003Cstrong>Pattern mastery without world anchoring\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Consequences:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Models generate plausible author combinations, titles, and venues.\u003C\u002Fli>\n\u003Cli>There is no native link from a generated reference to CrossRef, arXiv, PubMed, or publisher databases \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Two main failure modes follow:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Ghost references:\u003C\u002Fstrong> entirely fabricated but plausible works.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Citation unfaithfulness:\u003C\u002Fstrong> real papers cited for claims they do not support, often chosen for textual similarity rather than semantic relevance \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Asking models how “confident” they are in a citation does not help:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>They generate numbers that “sound right” given training patterns, biased toward high confidence \u003Ca href=\"#source-8\" class=\"citation-link\" title=\"View source [8]\">[8]\u003C\u002Fa>\u003Ca href=\"#source-9\" class=\"citation-link\" title=\"View source [9]\">[9]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Re-running the same scoring prompt yields different values due to non-deterministic sampling, so there is no stable internal certainty \u003Ca href=\"#source-8\" class=\"citation-link\" title=\"View source [8]\">[8]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>⚠️ \u003Cstrong>Fluent scholarly style without factual binding\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Instruction tuning and RLHF optimize for helpful, fluent, decisive output. Models “write like scholars, cite like experts” \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>, but:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Nothing in the architecture enforces that references correspond to real works.\u003C\u002Fli>\n\u003Cli>Nothing guarantees that cited works actually support the associated claims \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>Mini-conclusion:\u003C\u002Fstrong> Citation fabrication is an emergent property of pattern-completion on scholarly text without binding to authoritative bibliographic databases \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>3. How Fake Citations Cross Into the Scholarly Record\u003C\u002Fh2>\n\u003Cp>Ghost references would be manageable if they stayed in chat logs. The real damage occurs when they enter formal scholarly infrastructure.\u003C\u002Fp>\n\u003Cp>Observed patterns:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Instructors and editors find bogus citations in student papers that already appear in dozens of published articles, often with minor variations \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Once a fake paper is cited in a publication, it gains \u003Cem>derived legitimacy\u003C\u002Fem>: others copy its references or trust Google Scholar’s representation without checking \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>📊 \u003Cstrong>Feedback loops in citation infrastructure\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Citation indices like Google Scholar:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Create citation records for references they cannot match to real documents, inferring stubs from bibliographies \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Allow those stubs to accumulate further citations from human authors and LLM-assisted workflows, reinforcing the illusion of reality \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Examples include:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>A fabricated article on education governance listed as a highly cited paper in Google Scholar, despite never being written by the named authors \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Student essays whose ghost references resolve to published articles that themselves cite the same non-existent studies \u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>NeurIPS 2025 analyses found:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Hybrid citations blending elements from multiple real papers—authors from one, title motif from another, venue from a third—into a single unresolved reference \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Invented authors, swapped venues, and dead URLs that still passed multi-reviewer peer review \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>⚠️ \u003Cstrong>Distorted metrics and claims\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Because citation counts feed into promotion, funding, and journal rankings:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Ghost and zombie citations distort impact indicators \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Imaginary papers with many citations inflate the perceived rigor of work that leans on them.\u003C\u002Fli>\n\u003Cli>Polluted citations become training data for future AI systems.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>Mini-conclusion:\u003C\u002Fstrong> Once ghost citations enter journals, conferences, and indices, they fully participate in academic metrics, making them hard to detect and unwind \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>4. Why RAG and Tools Help—but Do Not Fully Solve It\u003C\u002Fh2>\n\u003Cp>Retrieval-Augmented Generation (RAG) is often proposed as a fix: ground model outputs in retrieved documents so models can only cite what they see \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>Empirical results:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>RAG reduces the number of claims a model marks as “cannot assess,” because retrieval supplies more evidence \u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>In machine-generated news fact-checking, adding search results led to fewer abstentions and more decisive judgments \u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>💼 \u003Cstrong>Grounded on what?\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>The same work found a key downside:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Irrelevant or low-quality retrieval increased incorrect assessments, as models overfit to whatever evidence they were given \u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>If retrieval includes polluted bibliographies or secondary mentions of ghost references, RAG can \u003Cem>amplify\u003C\u002Fem> fabrication and misattribution \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Common RAG citation failure modes:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Letting models “fill in” missing fields—guessing DOIs, venues, or years from partial snippets.\u003C\u002Fli>\n\u003Cli>Blending metadata from multiple documents into a synthetic reference \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Citing real papers from retrieval results that do not actually support the claim (citation unfaithfulness) \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>⚠️ \u003Cstrong>Evidence layer, not guarantee\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Even with RAG:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Ghost references may be reduced but not eliminated.\u003C\u002Fli>\n\u003Cli>There is no automatic check that each citation both exists and semantically supports the argument \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>Mini-conclusion:\u003C\u002Fstrong> RAG is valuable infrastructure but must be paired with explicit verification and quality-controlled retrieval. Grounding alone does not prevent ghost or unfaithful citations \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>5. Human and Institutional Failure Modes: The Verification Gap\u003C\u002Fh2>\n\u003Cp>LLM “hallucinations” explain only part of the problem. Ghost citations spread because human and institutional practices leave a large verification gap.\u003C\u002Fp>\n\u003Cp>Survey data from 94 researchers \u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>41.5% copy-paste BibTeX entries without checking.\u003C\u002Fli>\n\u003Cli>44.4% take no action when encountering suspicious references.\u003C\u002Fli>\n\u003Cli>76.7% of reviewers do not thoroughly check references.\u003C\u002Fli>\n\u003Cli>80.0% never suspect fake citations might be present.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>In such conditions, LLM-generated fabrications pass through by default.\u003C\u002Fp>\n\u003Cp>💡 \u003Cstrong>Overload and misaligned incentives\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>The ACL community notes that exponential submission growth and reviewer overload push peer review toward shallow checks \u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>. NeurIPS papers with hallucinated citations each passed three or more reviewers, showing that even elite venues do not systematically scrutinize bibliographies \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>A parallel from journalism:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Ars Technica retracted an article after AI-generated quotes were presented as real statements by a named individual, violating core trust in quotation \u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Policies against unlabeled AI-generated content existed; enforcement at publication time failed \u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Similar patterns in scholarship:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Students submit AI-written assignments with overlapping sets of fabricated references.\u003C\u002Fli>\n\u003Cli>Editors receive papers citing fake work attributed to themselves, showing authors did not read or sanity-check their own lists \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Publication incentives reward volume and citation counts over meticulous evidentiary integrity \u003Ca href=\"#source-3\" class=\"citation-link\" title=\"View source [3]\">[3]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>⚠️ \u003Cstrong>Governance and workflow, not just models\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Vague standards like “citations must be accurate” and weak enforcement allow ghost citations to cross the publication boundary \u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>\u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>Mini-conclusion:\u003C\u002Fstrong> LLMs accelerate a long-standing tendency to skip citation verification. Addressing it requires changes to incentives, workflows, and governance, not only better models \u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>6. A Preventive Architecture for Citation-Safe LLM Use\u003C\u002Fh2>\n\u003Cp>Citation integrity must be treated as a systems-design problem. Effective defenses span tooling, workflows, and policy.\u003C\u002Fp>\n\u003Ch3>6.1 Separate generation from verification\u003C\u002Fh3>\n\u003Cp>Use LLMs for drafting, not for authorizing references. Route every citation through an automated verifier (e.g., CiteVerifier) that \u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Resolves DOIs, arXiv IDs, and URLs.\u003C\u002Fli>\n\u003Cli>Confirms titles, authors, venues, and years against canonical records.\u003C\u002Fli>\n\u003Cli>Flags non-existent or mismatched references for human review.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>📊 \u003Cstrong>Design principle:\u003C\u002Fstrong> the model proposes; independent systems verify.\u003C\u002Fp>\n\u003Ch3>6.2 Evidence-based confidence, not model introspection\u003C\u002Fh3>\n\u003Cp>Avoid asking models how sure they are. Instead, compute trust from:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Exact or near-exact matches between citation metadata and source documents.\u003C\u002Fli>\n\u003Cli>Successful URL or DOI resolution.\u003C\u002Fli>\n\u003Cli>Cross-document consistency in how a paper is described \u003Ca href=\"#source-8\" class=\"citation-link\" title=\"View source [8]\">[8]\u003C\u002Fa>\u003Ca href=\"#source-9\" class=\"citation-link\" title=\"View source [9]\">[9]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>This avoids relying on numerically expressed “confidence” that LLMs cannot reliably calibrate \u003Ca href=\"#source-8\" class=\"citation-link\" title=\"View source [8]\">[8]\u003C\u002Fa>\u003Ca href=\"#source-9\" class=\"citation-link\" title=\"View source [9]\">[9]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Ch3>6.3 Constrained RAG and adversarial checking\u003C\u002Fh3>\n\u003Cp>For RAG-based literature tools:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Restrict citations to items that actually appear in the retrieved corpus; forbid invented DOIs or venues \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Log retrieval traces for each citation so reviewers can inspect supporting documents.\u003C\u002Fli>\n\u003Cli>Use a second-pass “adversarial” model with retrieval access to challenge citations from the first pass, flagging suspect items for humans \u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-7\" class=\"citation-link\" title=\"View source [7]\">[7]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>💼 \u003Cstrong>Use LLMs as both writers and critics\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cp>Models are often better at spotting inconsistencies given evidence than at generating ground-truth citations. Allocate capacity to critique as well as draft.\u003C\u002Fp>\n\u003Ch3>6.4 Governance, review, and education\u003C\u002Fh3>\n\u003Cp>Institutions should adopt explicit rules that:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Treat unverified AI-generated references like fabricated quotes in journalism—prohibited unless clearly labeled and independently validated before publication \u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Update reviewer forms and editorial checklists to ask whether a sample of references was independently verified for existence and relevance \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fli>\n\u003Cli>Train authors and students that AI tools are unreliable citation engines; ghost references are a known failure mode, and humans remain responsible \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>⚠️ \u003Cstrong>Default stance:\u003C\u002Fstrong> AI-supplied citations are untrusted until proven otherwise.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>Mini-conclusion:\u003C\u002Fstrong> Robust citation safety combines constrained generation, independent verification, adversarial checking, and clear governance. No single layer suffices; together, they can prevent LLMs from poisoning citation ecosystems \u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Chr>\n\u003Ch2>Conclusion: Design for Citation Safety Up Front\u003C\u002Fh2>\n\u003Cp>LLMs fabricate citations because they are optimized for plausible text, not verified bibliographies. Even elite models hallucinate 14.23–94.93% of citations when asked to generate references, and ghost citations already appear in NeurIPS, ACL, and other AI\u002FML venues \u003Ca href=\"#source-5\" class=\"citation-link\" title=\"View source [5]\">[5]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>\u003Ca href=\"#source-12\" class=\"citation-link\" title=\"View source [12]\">[12]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>RAG, browsing, and self-reported confidence cannot, by themselves, solve a problem rooted in both model behavior and human verification gaps \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-6\" class=\"citation-link\" title=\"View source [6]\">[6]\u003C\u002Fa>\u003Ca href=\"#source-8\" class=\"citation-link\" title=\"View source [8]\">[8]\u003C\u002Fa>\u003Ca href=\"#source-9\" class=\"citation-link\" title=\"View source [9]\">[9]\u003C\u002Fa>. Citation indices and overloaded peer review pipelines allow fabricated references to enter the scholarly record and accumulate illegitimate influence \u003Ca href=\"#source-1\" class=\"citation-link\" title=\"View source [1]\">[1]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>A resilient response treats every AI-generated citation as untrusted until proven otherwise. That means coupling retrieval with automated verifiers, using LLMs as adversarial fact-checkers, instrumenting workflows with audits, and updating editorial and educational policies so ghost citations are seen as a preventable systems failure, not an inevitable side effect of automation \u003Ca href=\"#source-2\" class=\"citation-link\" title=\"View source [2]\">[2]\u003C\u002Fa>\u003Ca href=\"#source-4\" class=\"citation-link\" title=\"View source [4]\">[4]\u003C\u002Fa>\u003Ca href=\"#source-10\" class=\"citation-link\" title=\"View source [10]\">[10]\u003C\u002Fa>\u003Ca href=\"#source-11\" class=\"citation-link\" title=\"View source [11]\">[11]\u003C\u002Fa>.\u003C\u002Fp>\n\u003Cp>Any deployment of LLMs in scholarly or high-stakes settings should be designed for citation safety from the outset—before ghost references become embedded in the bodies of work institutions rely on.\u003C\u002Fp>\n","Introduction  \n\nLarge language models now assist with theses, grant proposals, and journal articles, often drafting sections and reference lists.  \n\nAcross 13 state-of-the-art models, hallucinated cit...","hallucinations",[],1879,9,"2026-02-27T14:12:42.043Z",[17,22,26,30,34,38,42,45,49,53],{"title":18,"url":19,"summary":20,"type":21},"Why Ghost References Still Haunt Us in 2025—And Why It's Not Just About LLMs","https:\u002F\u002Faarontay.substack.com\u002Fp\u002Fwhy-ghost-references-still-haunt","As early as late 2022, I understood that Retrieval Augmented Generation (RAG) would be the future. By grounding LLM responses in retrieved content, RAG should reduce or even eliminate certain types of...","kb",{"title":23,"url":24,"summary":25,"type":21},"The Fabrication Problem: How AI Models Generate Fake Citations, URLs, and References","https:\u002F\u002Fmedium.com\u002F@nomannayeem\u002Fthe-fabrication-problem-how-ai-models-generate-fake-citations-urls-and-references-55c052299936","The Fabrication Problem: How AI Models Generate Fake Citations, URLs, and References\n\n[Image 1]\n\nNayeem Islam\n\n31 min read\n\nJun 12, 2025\n\nYou have already faced it ; Now let’s understand why AI tools ...",{"title":27,"url":28,"summary":29,"type":21},"LLMs Generate Fake Citations in Academic Papers","https:\u002F\u002Fwww.linkedin.com\u002Fposts\u002Fpetercorke_the-problem-of-bullshit-i-think-this-is-activity-7408260816644534272-MEMw","The problem of bullshit (I think this is a more accurate term than hallucination, an hallucination is a perception problem, LLMs don’t perceive, they have a problem with incorrect outputs, but I digre...",{"title":31,"url":32,"summary":33,"type":21},"AI Chatbots Are Poisoning Research Archives With Fake Citations","https:\u002F\u002Fwww.rollingstone.com\u002Fculture\u002Fculture-features\u002Fai-chatbot-journal-research-fake-citations-1235485484\u002F","By Miles Klee\nDecember 17, 2025\n\nAcademic papers increasingly cite articles and publications invented by large language models. Gettу Images\n\nAs the fall semester came to a close, Andrew Heiss, an ass...",{"title":35,"url":36,"summary":37,"type":21},"NeurIPS research papers contained 100+ AI-hallucinated citations, new report claims","https:\u002F\u002Ffortune.com\u002F2026\u002F01\u002F21\u002Fneurips-ai-conferences-research-papers-hallucinations\u002F","NeurIPS, one of the world’s top academic AI conferences, accepted research papers with 100+ AI-hallucinated citations, new report claims\n\nSharon Goldman\n\nAI Reporter\n\nJanuary 21, 2026, 9:00 AM ET\n\nNeu...",{"title":39,"url":40,"summary":41,"type":21},"Fact-checking AI-generated news reports: Can LLMs catch their own lies?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.18293","Authors: Jiayi Yao, Haibo Sun, Nianwen Xue\n\nSubmitted on 24 Mar 2025\n\nAbstract:\nIn this paper, we evaluate the ability of Large Language Models (LLMs) to assess the veracity of claims in \"news reports...",{"title":39,"url":43,"summary":44,"type":21},"https:\u002F\u002Farxiv.org\u002Fhtml\u002F2503.18293v1","Abstract\nIn this paper, we evaluate the ability of Large Language Models (LLMs) to assess the veracity of claims in “news reports” generated by themselves or other LLMs. Our goal is to determine wheth...",{"title":46,"url":47,"summary":48,"type":21},"Why Confidence Scoring With LLMs Is Dangerous","https:\u002F\u002Fwww.epiqglobal.com\u002Fen-us\u002Fresource-center\u002Farticles\u002Fwhy-confidence-scoring-with-llms-is-dangerous","What to know before relying on confidence scoring of LLMs in a document review setting.\n\n When it comes to confidence assessments from LLMs, scoring predictions is essential. The most important thing ...",{"title":50,"url":51,"summary":52,"type":21},"Rethinking Confidence in LLM Data Extraction","https:\u002F\u002Fwww.linkedin.com\u002Fpulse\u002Frethinking-confidence-llm-data-extraction-bogdan-raduta-hpuwf","Rethinking Confidence in LLM Data Extraction\n\nExtracting structured data from invoices using Large Language Models (LLMs) is alluring – just prompt a model like GPT to read an OCR’d invoice and output...",{"title":54,"url":55,"summary":56,"type":21},"When Fabricated Quotes Cross the Publication Boundary","https:\u002F\u002Fwww.linkedin.com\u002Fpulse\u002Fwhen-fabricated-quotes-cross-publication-boundary-paul-mitchell-qt1hc","Yesterday, a Sunday, Ars Technica retracted an article after discovering that it contained fabricated quotations generated by an AI tool and attributed to a real individual who did not say them. The e...",null,{"generationDuration":59,"kbQueriesCount":60,"confidenceScore":61,"sourcesCount":62},210048,12,100,10,{"metaTitle":64,"metaDescription":65},"LLMs invent citations: 7 drivers, 6 fixes, 2025–2026","AI models fabricate 14–95% of citations in some tests. Learn why LLMs invent academic references, how this is polluting research, and what teams can do to prevent it.","en","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1752697589018-64fc6b19cc9a?w=1200&h=630&fit=crop&crop=entropy&q=60&auto=format,compress",{"photographerName":69,"photographerUrl":70,"unsplashUrl":71},"Krists Luhaers","https:\u002F\u002Funsplash.com\u002F@kristsll?utm_source=coreprose&utm_medium=referral","https:\u002F\u002Funsplash.com\u002Fphotos\u002Fbooks-related-to-law-and-human-rights-are-visible-5m4-lY2FXto?utm_source=coreprose&utm_medium=referral",false,{"key":74,"name":75,"nameEn":75},"ai-engineering","AI Engineering & LLM Ops",[77,84,92,99],{"id":78,"title":79,"slug":80,"excerpt":81,"category":11,"featuredImage":82,"publishedAt":83},"69df1f93461a4d3bb713a692","AI Financial Agents Hallucinating With Real Money: How to Build Brokerage-Grade Guardrails","ai-financial-agents-hallucinating-with-real-money-how-to-build-brokerage-grade-guardrails","Autonomous LLM agents now talk to market data APIs, draft orders, and interact with client accounts. The risk has shifted from “bad chatbot answers” to agents that can move cash and positions. When an...","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1621761484370-21191286ff96?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHxmaW5hbmNpYWwlMjBhZ2VudHMlMjBoYWxsdWNpbmF0aW5nJTIwcmVhbHxlbnwxfDB8fHwxNzc2MjMwNzM5fDA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-04-15T05:25:38.954Z",{"id":85,"title":86,"slug":87,"excerpt":88,"category":89,"featuredImage":90,"publishedAt":91},"69de1167b1ad61d9624819d5","When Claude Mythos Meets Production: Sandboxes, Zero‑Days, and How to Not Burn the Data Center Down","when-claude-mythos-meets-production-sandboxes-zero-days-and-how-to-not-burn-the-data-center-down","Anthropic did something unusual with Claude Mythos: it built a frontier model, then refused broad release because it is “so good at uncovering cybersecurity vulnerabilities” that it could supercharge...","security","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1508361727343-ca787442dcd7?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxfHxtb2Rlcm4lMjB0ZWNobm9sb2d5fGVufDF8MHx8fDE3NzYxNjE2Njh8MA&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-04-14T10:14:27.151Z",{"id":93,"title":94,"slug":95,"excerpt":96,"category":89,"featuredImage":97,"publishedAt":98},"69ddbd0e0e05c665fc3c620d","Inside the Anthropic Claude Fraud Attack on 16M Startup Conversations","inside-the-anthropic-claude-fraud-attack-on-16m-startup-conversations","A fraud campaign siphoning 16 million Claude conversations from Chinese startups is not science fiction; it is a plausible next step on a risk curve we are already on. [1][9] This article treats that...","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1487017159836-4e23ece2e4cf?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxNnx8YnVzaW5lc3MlMjBvZmZpY2V8ZW58MXwwfHx8MTc3NjEzOTczM3ww&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-04-14T04:08:51.872Z",{"id":100,"title":101,"slug":102,"excerpt":103,"category":104,"featuredImage":105,"publishedAt":106},"69dd95fa0e05c665fc3c5fde","Designing Acutis AI: A Catholic Morality-Shaped Search Platform for Safer LLM Answers","designing-acutis-ai-a-catholic-morality-shaped-search-platform-for-safer-llm-answers","Most search copilots optimize for clicks, not conscience. For Catholics asking about sin, sacraments, or vocation, answers must be doctrinally sound, pastorally careful, and privacy-safe.  \n\nAcutis AI...","safety","https:\u002F\u002Fimages.unsplash.com\u002Fphoto-1675557009285-b55f562641b9?ixid=M3w4OTczNDl8MHwxfHNlYXJjaHwxNnx8YXJ0aWZpY2lhbCUyMGludGVsbGlnZW5jZSUyMHRlY2hub2xvZ3l8ZW58MXwwfHx8MTc3NjEyOTgwMHww&ixlib=rb-4.1.0&w=1200&h=630&fit=crop&crop=entropy&auto=format,compress&q=60","2026-04-14T01:23:19.348Z",["Island",108],{"key":109,"params":110,"result":112},"ArticleBody_k7U1Igw9TCQnNYEdvijh3CTUwH14FuIAPPizqQXiJaw",{"props":111},"{\"articleId\":\"69a1a50e83962bbe60b24fbe\",\"linkColor\":\"red\"}",{"head":113},{}]